Rank Dough builds high intent landing pages in 8 steps: find the decision behind the search, collect evidence for it, turn each decision into a page component, then use AI to build the page and check it. For a UK student accommodation marketplace we mapped 26,716 keywords and ideas before AI wrote a line, and turned the London searches into the buying decisions the page now answers.

This guide walks through each step with the prompts we used, 16 figures from the real build, and the mistakes we made on the way.

What problem does a high intent landing page need to solve?

A high intent landing page has one job: help someone who is ready to choose make the choice on that page. Most pages I review do something else. They repeat the keyword, add a hero photo and a few lines about quality, then send the visitor off to five other tabs to work out the price, the terms and whether the thing fits.

Take “student accommodation London”. The keyword says London. The person searching wants to know which rooms they can afford, how far each one is from their lectures, how many weeks they pay for, and whether the person paying will agree. A page that answers the keyword and none of those questions ranks for a while and converts badly.

The fix is not more copy. It is working out the decision behind the search, finding evidence for it in your data, and building the page around that decision. AI is very good at the building part. It is bad at the deciding part, and it fills every gap you leave with confident guesses.

Research outside our own build points the same way. Hidden or unclear costs are a top reason people walk away at checkout, most attention goes to the first screen, and plain copy converts better than clever copy.

Sources: Baymard Institute, cart abandonment rate (updated September 2025); Nielsen Norman Group, Scrolling and Attention (2018); Unbounce Conversion Benchmark Report (2024); Save the Student, National Student Accommodation Survey 2025.

None of those studies measured a student accommodation page. I use them as reasons to test a layout, not as proof that a layout works.

What does the workflow look like from data to page?

It is eight steps, and each one hands a concrete output to the next. Steps 1 to 4 are data work: what the page is for, what evidence you have, which decisions the searches point to, and which components answer them. Steps 5 to 8 are build and proof: briefing AI, laying the page out, making it readable for search and AI answers, then checking and measuring it.

The example all the way through is a London city page we prototyped for a UK student accommodation marketplace in October and November 2026. The client is not named, and the brand is blurred in every screenshot.

Step 1: Which decision should the page own?

Write the page’s job in one sentence before you touch any tool. Ours was: “Help a student find London rooms that fit their budget and campus, see the full cost of the stay, and take the plan home.” Everything that doesn’t serve that sentence goes lower on the page or off it.

Then check that you aren’t about to build a second page for a job an existing URL already does. Two pages chasing the same searches split the clicks and confuse Google about which one to show.

1. In Search Console, open Performance, filter queries containing the topic (here “london”) and switch to the Pages tab.
2. Note every URL that already gets impressions for those queries, and its average position.
3. If an existing page sits in the top five, improve it rather than adding a new one.
4. If an older page targets the same intent and performs worse, plan a 301 redirect to the new one.
5. Write the chosen URL into the brief so nobody builds a duplicate later.

Step 2: What data do you need before you open an AI tool?

You need four kinds of evidence, and a keyword export is only one of them. Keyword data tells you what people type. It cannot tell you what you sell, what it costs, or why buyers hesitate.

Source: Rank Dough evidence pack template, used on the London prototype, October 2026.

Keep the source and date next to every row. When a figure turns up later in copy, a chart or an AI answer, you should be able to point to the row it came from in under a minute.

Source: Rank Dough UK keyword map: Ahrefs UK volumes pulled 30 October and 2 November 2026, Search Console 1 August to 30 October 2026.

Step 3: How do you turn keywords into buying decisions?

Read every keyword as a question the visitor is trying to answer. “Cheap student accommodation London” is a budget question. “Student accommodation near UCL” is a location question. “How many weeks do you pay for student accommodation” is a contract question. Group the rows by decision, not by topic.

Start wide, then go narrow. Across the UK map, transactional searches split into topics such as city and area searches, flats, studios and private accommodation.

Source: Rank Dough UK keyword map: Ahrefs UK volumes pulled 30 October and 2 November 2026, Search Console 1 August to 30 October 2026. Volumes are summed estimates for the measured keywords only.

Then zoom into the page you are building. For London, the broad term carries the volume, but the smaller searches name the decisions the page has to support.

Source: Ahrefs UK volumes via the Rank Dough keyword map, pulled 30 October to 2 November 2026. Volumes overlap, so the rows are not added together.

Method: we took the London rows with a measured volume from the map and tagged each with the one decision it most clearly points to. Rows that fit no decision stayed as general demand.

Volume sizes demand. It does not rank decisions. The 100-search campus query asks for a harder feature than the 4,700-search broad term, and the page fails that visitor if it is missing.

Step 4: How do decisions become page components?

Give every decision one main component and one supporting detail, and write down the data field each one needs. If you can’t name the field, you can’t build the component properly yet.

The contract decision shows why this matters. A weekly price filter tells you which room is cheaper per week. It doesn’t tell you which room costs less, because rooms run on different contract lengths.

Source: room examples recorded in the source listing editor on 23 October 2026 for tenancies from 12 September 2026. Historical, not current offers.

We put that comparison on the page itself as a small table under the cost planner, with a one-line summary a visitor can read without opening anything.

Source: London prototype room examples, recorded 23 October 2026. £225 × 42 = £9,450; £225 × 51 = £11,475; £230 × 44 = £10,120.

The main component for this decision is a full-stay cost planner. The visitor enters the exact weekly rent and contract weeks, adds their own living and travel estimates, and gets one number for the whole stay.

Source: London prototype planner with its default inputs, 4 November 2026. £10,120 rent + £3,520 living + £1,100 travel + £200 one-off = £14,940.

Step 5: How do you brief AI so it builds what the data says?

Split the job into three passes and read the output of each one before you start the next. One long prompt that says “build me a landing page from this data” lets a single weak assumption run through the copy, the filters and the maths at once.

Pass 1: the evidence table. Ask AI to list the decisions the page must support, each with the source row behind it, and to keep facts and ideas in separate columns.

Pass 2: the component spec. For each decision you approved, ask for the fields, labels, formula, empty state and a test.

Pass 3: the build. Only now does AI write code and copy, as a separate version so you can compare it with the last one.

Step 6: In what order should the page be laid out?

Lay the page out in the order people make the decision. Each block answers the next question a visitor has, and optional blocks sit below the tools, never above them.

The first tool asks where the visitor studies and what they can pay. London is too big for “near the centre” to mean anything, so the campus field comes first, and the page links to a route rather than printing a travel time it can’t stand behind.

The property cards then carry the same facts in the same place, so a visitor can compare without opening each one. Price always has its unit. Each card says what the listing claims and what still needs confirming.

The page ends with a plan the student can take home. For a student, the decision often isn’t theirs alone. The summary puts the room, the weeks and the total in one card they can download or copy.

Source: Rank Dough London city page prototype, captured 4 November 2026. In the prototype, 12 properties carry reference starting rents from £230 to £450 a week; 7 of them are at or under £270.

Step 7: How do you make the page readable for Google and AI answers?

Put the important facts in plain text on the page, not only inside a calculator or an image. Google says there are no extra requirements to appear in AI Overviews or AI Mode, and lists making important content available in text and keeping structured data in line with the visible page among its best practices.

Source: Google Search Central, AI features and your website, checked 4 November 2026.

1. Write each key fact as a sentence that stands on its own: “At £225 a week for 51 weeks, rent comes to £11,475.”
2. Keep units, dates and assumptions next to the number, so the line still makes sense when it is quoted on its own.
3. Use one clear H1 and a title that names the place and the offer.
4. Answer the question searches in a short FAQ near the bottom, with FAQ schema that matches the visible text word for word.
5. Link each card to the exact property page and each area to its own guide.
6. Check the rendered page, not the template: the facts must be in the HTML a crawler receives.

Don’t spin up a page for every filter combination because AI makes it cheap. A studio filter on the London page serves “student studio london” well. A separate studio page only earns its place if it has enough studios to show and a reason to exist on its own.

Step 8: What do you check before the page goes live?

Check the numbers, then the states, then the task. In my experience the first check gets done and the last one gets skipped.

Source: Rank Dough QA list for the London prototype. The 7 of 12 answer comes from the prototype’s reference rents.

What went wrong in our build, and what did we change?

I’d rather show the misses than pretend the first version was right. These are the ones that taught us something.

One problem is still open. On a laptop screen 698 pixels tall, the first screen of the prototype is mostly headline and photo, and the campus and rent tools start right at the bottom edge.

Nielsen Norman Group found people spend 57% of their viewing time above the fold. That is a reason to test a shorter opening with the tools visible, not proof that it will convert better. It is the next change on our list.

Does your landing page pass the decision check?

Tick what your current page already does. The check lists what is missing and links to the step that fixes it. The pass bands are Rank Dough’s judgement, not a measured score.

Plain version of the check: One sentence states the decision the page helps someone make; Every keyword group is tagged with a decision, not just a topic; Price shows its unit, and the full cost of the stay or order is visible; Each component on the page traces back to a source row; AI was briefed in separate passes, with a check after each; The selection tools are visible on the first screen; Key facts are written as plain sentences in the HTML; Someone new has completed the main task without help.

How do you know the new page works?

Pick one commercial outcome before launch and measure that, not a proxy. For the London page that is enquiries or bookings that start from the page. A saved room or a downloaded summary is a useful signal, but it isn’t a sale.

Source: Rank Dough measurement plan for the London page.

Compare the same weeks year on year where you can, because student housing is seasonal. Note the launch date, and any price, stock or campaign change in the same window. If several things changed at once, report the result as a pattern, not proof.

To be clear: the London page is a prototype. We have not measured rankings, bookings or AI citations from it yet, and this guide doesn’t claim any.

What else should you know before you build one?

Sources

• Rank Dough build record: UK keyword map (Ahrefs UK 30 October and 2 November 2026, Search Console 1 August to 30 October 2026) and the London city page prototype, October to November 2026
• Google Search Central, AI features and your website: developers.google.com
• Baymard Institute, cart abandonment rate statistics: baymard.com

The post How to build a high intent landing page with data and AI: an 8-step guide appeared first on Rank Dough Digital.

Rank Dough compared 8 AI SEO agencies working in London, ours included, and charted 8 sites in Ahrefs: 6 agency clients and 2 sites from my earlier roles. Only 1 of the 8 sites is up on a year ago, and only 2 of the 7 London agencies give dates you can check a claim against.

Not sure which client story is closest to your business? Take the quick check. Pick one and the matching section opens, with a question to put to any agency.

Which London AI SEO agencies did we compare?

We compared 8: Croud, Elixirr Digital, Reboot Online, Kaizen, Passion Digital, Position Digital, BE VISIBLE and Rank Dough. They run from two firms that each state 600+ staff to small teams, so the first table shows each agency’s client evidence and the second shows scale, price and awards. Rank Dough is ours and is not London based, so it is disclosed and checked on the same evidence as the rest.

An agency made the list only when we read three things on its own pages: a London base, a named client and a result with a number. Two of the seven, Elixirr Digital and BE VISIBLE, list London without a street address, and the table says so.

We used the same checks in our Dublin list and our UK travel AI SEO agencies list, so you can compare agencies across all three.

Source: agency websites and case study pages, read 2 November 2026; agency award pages and the Rank Dough home page, read 3 November 2026; Ahrefs Site Explorer, estimated organic visits a month, pulled 2 November 2026. Evidence label: Backed means Ahrefs shows the claimed change in the claimed window; Partly means the direction fits but dates, domain or metric cannot be matched; Not checkable means the claim is not a traffic figure; Not charted means no client domain was named.

What happened to their clients’ organic traffic?

We pulled estimated monthly organic visits for each client’s whole domain from Ahrefs Site Explorer on 2 November 2026. Every chart runs to October 2026, so the last 12 months compare October 2025 with October 2026.

Source: Ahrefs Site Explorer, estimated organic visits a month, subdomains included, pulled by Rank Dough on 2 November 2026.

1 of the 8 sites grew over the last 12 months: Thorne, by 7%. The other 7 fell: Decentriq by 7%, Hostelworld by 23%, Gozney by 33%, TUI by 34%, Aer Lingus by 41%, MS Trust by 46% and Shoosmiths by 63%.

Most of the case studies describe 2025. Most of the falls happened in 2026. That is a pattern in 8 sites, not a finding, and I have not tested why.

Change compares October 2026 with October 2025 for the same domain. Domains: gozney.com is named in Kaizen’s case study; thorne.com, mstrust.org.uk, tui.co.uk, shoosmiths.com and decentriq.com are matched to the client name. Estimated organic visits a month, Ahrefs Site Explorer, pulled 2 November 2026; October 2025 column checked again on 3 November 2026.

One method note. Ahrefs estimates traffic from rankings and search volumes for the whole domain, so it can disagree with a client’s own analytics, and it cannot see AI assistant sessions at all. Where the two differ, the agency’s figure may still be right, but it needs dates and a domain to be checked.

What do Croud’s clients show?

Thorne is a US supplement brand. Croud reports 30% more organic revenue year on year, 22% ahead of target, and a 66% AI mention rate on its own brand tracker. No dates are given.

Ahrefs shows Thorne’s 2025 average at 315,229 visits a month, 1.4 times its 2024 average of 220,796. It peaked at 482,641 in April 2026 and had 323,945 in October, 7% above a year earlier. It is the only site here that is up.

What do Elixirr Digital’s clients show?

MS Trust is a UK multiple sclerosis charity. Elixirr reports 4,061 AI sessions, up 907.7% year on year, and AI Overview coverage that grew from 1 tracked keyword in April to 42 in October 2025. Ahrefs cannot see AI sessions.

What it can see is the whole site: 46,885 visits in October 2024 and 84,209 in October 2025, 1.8 times. The site peaked at 107,194 in November 2025 and had 45,880 in October 2026, down 46%.

What do Reboot Online’s clients show?

TUI is a holiday company, and the case study covers its family holidays. Reboot reports TUI ranked first in AI answers for “best family holiday provider” with a 63% share of voice, up from 46% to 78% across its tracked prompts. Those are AI visibility measures, so a traffic chart cannot confirm or deny them.

The chart is context. tui.co.uk peaks each summer and each January. June 2026 had 6,954,313 visits against 13,239,146 in June 2025, and October 2026 is 34% below October 2025.

What do Kaizen’s clients show?

Gozney makes pizza ovens. Kaizen names the domain, Gozney.com, and reports 70% more UK non-brand traffic year on year, 116% more top 3 rankings, 1,900% more LLM traffic and 32% more global organic revenue. No start month is given.

Ahrefs cannot split UK non-brand traffic out of a whole domain. The whole domain went from 210,121 visits in June 2024 to 309,786 in June 2025, 1.5 times. It had 157,083 in October 2026, 33% below a year earlier.

What do Passion Digital’s clients show?

Shoosmiths is a UK law firm, and the case study covers its serious injury practice. Passion reports 33% more sitewide organic clicks year on year, 213 new page one rankings and 21 new AI Overview features, shown at the client’s conference in January 2025.

Ahrefs shows the domain from June 2023. July to December 2024 averaged 22,439 visits a month against 12,068 a year earlier, 1.9 times. The site peaked at 28,644 in May 2025 and had 5,041 in October 2026, down 63%.

What do Position Digital’s clients show?

Decentriq is a data clean room company. Position Digital gives the window, April to October 2025, and the numbers: organic traffic from 1,385 to 2,233, up 61%, and AI citations from 14 to 45.

Ahrefs shows 1,311 visits in April 2025 and 2,376 in October 2025, 1.8 times, so the claim holds in the stated months. The site peaked at 2,908 in March 2026 and had 2,220 in October, 7% below a year earlier.

What do BE VISIBLE’s clients show?

BE VISIBLE’s Caliber case study reports 28% more organic clicks in three months, to 4,780 on 1.24 million impressions, and a first enterprise lead from AI search after two months. It does not name Caliber’s domain or a start month, so there is nothing we can chart without guessing. Not charted stays not charted.

What do Rank Dough’s results show?

Rank Dough has no London client we can chart yet. The two sites below are my work before Rank Dough, so I do not count them as agency clients.

At Aer Lingus I worked through iProspect from January 2021 to March 2022. Traffic went from 88,779 to 440,720 visits a month, 5.0 times, while travel reopened, so I do not credit that to SEO alone. At Hostelworld I was in-house from May 2022 to January 2025, and whole-site traffic rose 19%. Both are down on a year ago: Aer Lingus by 41% and Hostelworld by 23%.

What do these charts tell a London business?

Read together, the 8 charts show five patterns a London business should expect from SEO and AI search work. Each comes from the charts above. They are my reading of the data, not tested rules.

Which of these client stories is closest to your business?

Pick the story closest to your business. The matching section opens below, with a question to put to any agency.

How should you choose an AI SEO agency in London?

Four checks follow from the data above. They are my judgement. I have not tested whether they predict results.

1. Ask for the client’s domain and the month the work started. Only Kaizen named a domain and only Position Digital and Elixirr Digital gave dates.
2. Check the claim in Ahrefs or ask for a Search Console export. Decentriq’s rise shows in the stated months; TUI’s share of voice cannot be checked from traffic at all.
3. Ask what happened after the case study ended. 7 of the 8 sites are down on a year ago, including Shoosmiths by 63% and MS Trust by 46%.
4. Match the client to your business, not the agency to your size. A pizza oven brand, a charity, a law firm and a data company need different work, and the closest client tells you more than the headcount.

Price and scale matter too. Croud and Elixirr Digital each state 600+ staff and publish no prices. Three show a number: Reboot Online from £2,500 a month, Position Digital from $2,000 a month and Passion Digital budget bands from under £3k to £10k+. Ask the others for a range before the first call.

Awards help less than a dated case study. 5 of the 7 London agencies show a win on their own site, but Croud’s and Passion Digital’s search awards are from 2023, Kaizen’s is from 2018 and Elixirr Digital’s is an employer award. Position Digital shows no award and has the one claim Ahrefs fully backs. Ask which client and which year an award covers, then check that client the same way.

What do people ask about AI SEO agencies in London?

Who is Rank Dough, and what are the limits of this list?

Rank Dough is a small team. Our site names 2 people, Roman Sadowski and David Mason. We trade as Fortunata and work from Loughrea, Co. Galway, with a UK address in Essex, not London. We put ourselves in this list because London businesses compare us with London agencies.

I have worked in search since 2004. The record charted here is Aer Lingus at iProspect and Hostelworld in-house. Neither was a Rank Dough engagement. The awards on our home page come from the same work: Best Use of Search and Travel and Tourism for Aer Lingus, and Best Blog for Workplace from Meta. This article is published by Rank Dough, which appears in the table, so check every claim against the named domain in Ahrefs.

The limits are clear. Traffic is estimated at domain level from one pull on 2 November 2026. We did not measure sales, leads, AI sessions or AI citations. There is no before-and-after test, 5 of the 6 other-agency domains are matched to the client name rather than linked, start dates are unknown for 4 of the 6, and each agency is judged on the one client it chose to publish. If you work in Ireland, our Dublin list uses the same method. We also checked Rank4AI (Colchester base), Digitaloft (Kendal and Manchester), Blue Array (Reading), Re:signal (remote), Rise at Seven (Manchester head office), Builtvisible, NetTrackers, Brainlabs, Semetrical, Impression and Exposure Ninja, and left them out because we could not read a London base, or a named client with an AI search result, on their pages.

Sources

• Ahrefs Site Explorer, estimated monthly organic traffic for 8 client domains, pulled by Rank Dough on 2 November 2026
• Position Digital, Decentriq case study, April to October 2025 figures
• Google Search Central, Do you need an SEO?

The post Best AI SEO Agencies in London (2026): Compared on Client Results appeared first on Rank Dough Digital.

You master it in five stages over about 24 weeks, and Rank Dough built this roadmap from 17 sources we opened and checked on 3 November 2026. It starts with how an agent reads a page, because the best GPT-4 agent in the WebArena benchmark finished only 14.41% of web tasks while humans finished 78.24%.

What is the agentic web?

The agentic web is the web used by an AI agent on a person’s behalf. You tell the agent “find me a studio near the university under a set budget and send an enquiry”, and the agent does the browsing, the comparing and the form.

That changes what success means for a website. A person reads the design and the copy. An agent reads the code: labels, buttons, structured data and the form itself.

SEO has always been about getting found. On the agentic web, getting found isn’t enough. The agent also has to finish the job, and I see that as the part most sites haven’t tested yet.

How is the roadmap laid out?

The roadmap runs in five stages, and each one needs the one before it. It assumes 3 to 5 hours a week and an SEO who already knows the basics, so it skips beginner material on purpose.

Source: Rank Dough agentic web roadmap, 3 November 2026.

We built it in a fixed order: Rank Dough’s own notes and tests first, then the live web for primary sources, then every source opened before it went in. Anything we couldn’t open and date was cut.

Where should you start? The skip test

Answer five questions without looking anything up. Every one you get right lets you skip ahead, so you don’t spend weeks on what you already know.

The starting point is Rank Dough’s judgement, not a scored test. Without JavaScript, the rule is:

1. Four or five answered: start at Stage 3.
2. Two or three answered: start at Stage 2.
3. One or none: start at Stage 1.

Stage 1: how does an agent read, decide and act?

You’ll be able to: explain to a client, from the agent’s side, why their booking flow fails an agent when a person gets through it. Time box: weeks 1 to 4.

World class techniques

1. Read the page as the agent does. Agents work from a mix of the page code, the accessibility tree and screenshots. OpenAI says its Atlas browser uses ARIA roles and labels to read page structure. It fails where the agent works from screenshots only. Established.
2. Map the task, not the page. Score a site on whole tasks (find, compare, enquire, pay), the way WebArena and WebMall do, not page by page. It fails on sites with one simple action. Established.
3. Separate the three traffic types. Training crawlers, search crawlers and agents acting for a user behave differently and need different rules. Established.

Sources: arXiv papers above; W3C WAI; OpenAI Publishers and Developers FAQ, checked 3 November 2026.

Unconventional approach: learn from accessibility first. Run a screen reader through a client’s enquiry form before any agent. Where the screen reader stalls, the agent usually stalls too. That’s a hypothesis, and the drill tests it.

Drill: pick three tasks on a live site you work on: find an item under a budget, compare two items, start an enquiry. Run each with an agent and with a screen reader, and log where each one fails.

Proof of mastery: a one page “why the agent failed” teardown per task, each failure tied to a cause: a missing label, a CAPTCHA, a choice lost from the web address, or a control that only works in JavaScript.

Stage 2: how do you see agents on a live site?

You’ll be able to: show a client how many agent visits they get, which tasks agents attempt and how many finish. Time box: weeks 5 to 9.

World class techniques

1. Verify, don’t guess from user agents. Signed agents sign each request with HTTP Message Signatures, and the site finds the key through a Signature-Agent header. It fails for agents driving the user’s own browser, which look human. Established.
2. Split the logs by purpose. Tag every AI hit as training, search, user action or undeclared, using Cloudflare’s four groups. Cloudflare measured crawl to refer ratios of about 887 to 1 for OpenAI and 50,000 to 1 for Anthropic in early August 2025. Established.
3. Measure task success, not visits. An agent visit that ends at a CAPTCHA is a lost lead. Build an agent funnel: landed, found the item, reached the form, submitted. Hypothesis, because there’s no public benchmark for live sites yet.

Sources: IETF datatracker; Cloudflare blog, checked 3 November 2026. The IETF draft has since been replaced by a newer version.

Unconventional approach: build the tool. A small script that reads raw logs, labels each AI hit by purpose and signature, and pushes a weekly table to a dashboard. Building it teaches you what can and can’t be seen.

Drill: run an agent test on a live site before and after a change to its bot settings. Compare verified agent hits, form reaches and submissions.

Proof of mastery: an agent traffic panel on one live site with four weeks of data, labelled by purpose, with the agent funnel next to it.

Stage 3: how do you make a site agent ready?

You’ll be able to: tell a client which standard fits their business in one line, and ship the first piece. Time box: weeks 10 to 15.

Sources: the specification pages linked in the table, all checked 3 November 2026. Chrome status from Daniel Vaughan’s October 2026 write up.

World class techniques

1. Schema first, standard second. NLWeb reads Schema.org, so clean structured data is the cheapest agent work and pays off in search too. Established for NLWeb, a hypothesis for other agents.
2. Fix the form before adding a standard. Labelled fields, ARIA roles, choices kept in the web address, and no tick box CAPTCHA for verified agents. In our Cloudflare agent test the agent failed on a reCAPTCHA tick box, not on a missing standard. Practitioner evidence, one test.
3. Pilot WebMCP on one form. Wrap a single enquiry form as a declarative WebMCP tool. It’s cheap, you can undo it, and few sites have tried it. It does nothing today in a browser without the preview switched on. Hypothesis.
4. Shops: pick UCP or ACP by platform. Shopify merchants get UCP through Shopify. Stripe merchants have the ACP route. Don’t build both. Practitioner evidence, vendor documents only.

Unconventional approach: break it on purpose. On a test page you own, ship a WebMCP form tool, then remove the labels, then add a CAPTCHA. Run the same agent each time and watch what it does.

Drill: write an agent readiness checklist (crawl access, a rule for verified agents, ARIA, schema, the form, the /.well-known files) and run it on two real sites.

Proof of mastery: one live WebMCP or NLWeb pilot on a site you own, and the checklist scored on two real sites.

Stage 4: where do the experts disagree?

You’ll be able to: hold a view on each open fight and say where it breaks. Time box: weeks 16 to 19.

1. Who owns it. John Mueller asked the question on LinkedIn and Joe Hall answered it: SEO owns discovery and understanding, developers own login, payments and APIs. Natzir Turrado’s AX idea, agent experience next to UX, pushes it towards a new role. Your view should name which half you sell.
2. ARIA for agents. OpenAI asks sites to add ARIA so Atlas can act. Adrian Roselli argues this invites ARIA keyword stuffing that hurts people who use screen readers. Read both.
3. Block or let in. Training crawlers give little back, with crawl to refer ratios in the hundreds to tens of thousands. Agents acting for a user bring buyers. One blanket rule is wrong for most businesses. That’s my reading of Cloudflare’s data.
4. Three checkout standards. UCP, ACP and AP2 overlap and nobody knows which wins. Treat any client spend here as a small bet you can undo.
5. WebMCP adoption. It’s still a draft in an early Chrome preview, and the API name has already changed once. Early pilots may need rewriting.

Unconventional approach: read against yourself. For each fight, write your view in three lines, then find the strongest source against it and rewrite.

Drill: one LinkedIn post per fight for business owners, one a fortnight. Track which ones draw objections and from whom.

Proof of mastery: a one page position paper, “What we do and don’t do on the agentic web”, with each claim labelled established, practitioner evidence or hypothesis.

Stage 5: how do you teach, sell and predict it?

You’ll be able to: sell an agent readiness check, run it with an agent, and make calls on the market that come true more often than not. Time box: weeks 20 to 24.

1. Turn the checklist into a service. An agent readiness check: an agent task test, a rule for verified agents, form fixes, schema and one standard picked. Hypothesis, because demand isn’t tested yet.
2. Publish proof, not opinion. Anonymised agent tests on real sites are what most agencies don’t have. Each one becomes a short case note.
3. Keep a prediction log. Ten dated calls, for example “WebMCP reaches stable Chrome by a set date” or “a top Irish retailer publishes /.well-known/ucp”. Score them every quarter.

Unconventional approach: ask the builder. Send one precise question to a WebMCP author or an A2A steering committee member, for example how agents should handle a step that needs a verified human. Publish the answer with permission.

Drill: run one agent readiness check end to end for a warm prospect, with an agent doing the task test.

Proof of mastery: one paid or signed off check, one published case note, and a prediction log with at least ten scored calls.

What should a prediction log look like?

Five columns are enough. Write the call before you check it, and never edit a row after the date passes.

The point is honesty with yourself. The reality is this: most predictions about the web sound right when you make them and nobody checks them later.

What doesn’t this roadmap prove?

It’s a learning plan, not a measured method. The order of stages and the time boxes are Rank Dough’s judgement, and no one has run all 24 weeks end to end yet.

The standards change fast. Every status in the Stage 3 table is dated 3 November 2026, and WebMCP in particular may look different in a few months. The screen reader link in Stage 1 and the agent funnel in Stage 2 are hypotheses until they’re measured on live sites.

Frequently asked questions about the agentic web roadmap

Sources

• Rank Dough agentic web roadmap research, 17 sources opened and checked on 3 November 2026
• Zhou et al., WebArena: A Realistic Web Environment for Building Autonomous Agents, arXiv 2307.13854v4: arxiv.org/abs/2307.13854v4
• W3C Web Machine Learning community group, WebMCP: github.com/webmachinelearning/webmcp

The post How do you master the agentic web? A 24 week roadmap appeared first on Rank Dough Digital.

It might be, and most owners wouldn’t know. Since 15 September 2026, Cloudflare blocks AI agents by default on pages that show ads for every new domain it takes on, and Rank Dough’s own test on 3 November 2026 found the bigger blocker sat further down the page: the agent got through Cloudflare on 3 of 3 prompts and through the enquiry form on 0 of 3.

What did Cloudflare change for AI agents?

On 1 July 2026 Cloudflare stopped treating AI traffic as one thing. It now sorts AI bots into three groups by what they do on your site: Search, Agent and Training.

Each group gets its own switch. You can block it on every page, block it only on pages that show ads, or leave it alone. Before this, Cloudflare offered one “Block AI bots” toggle that did everything at once.

The default is the part to watch. From 15 September 2026, every new domain added to Cloudflare blocks Agent and Training on pages that show ads, while Search stays allowed. If your site carries ads and moved to Cloudflare this autumn, agents may already be turned away on those pages.

Source: Cloudflare changelog, “New options to manage AI traffic”, 1 July 2026, and the Cloudflare blog post of the same date.

One thing Cloudflare doesn’t publish is how it decides a page “displays ads”. Its blog says an ad is a sign the page was built for a person to land on. It doesn’t say which ad scripts count, so I wouldn’t assume a page without banner ads is safe.

Why would a business want AI agents on its site?

Because the agent is the customer’s hands. Cloudflare’s own definition says there is often a human waiting on the other end. A student asks an AI to find a studio near their university and send an enquiry, and the agent goes off and does it.

If your site turns that agent away, the student doesn’t see an error page. The agent tries the next site on its list, and the enquiry lands with a competitor. Nothing shows in your analytics, because the visit never happened.

That’s why I treat this as a sales question, not a technical one. Blocking Training is a choice about your content. Blocking Agent is a choice about whether people who shop through AI can buy from you.

How do you check if your site is on Cloudflare?

Ask whoever runs your domain, or check it yourself in under a minute. Open your site in Chrome, press F12, go to the Network tab and reload. Click the first request and look at the response headers.

If you see “server: cloudflare” and a “cf-ray” header, your traffic goes through Cloudflare and this setting applies to you. If the server says Apache, nginx or something else, Cloudflare isn’t in front of your site and you can skip to the forms section.

When we checked on 3 November 2026, the student accommodation site in our test returned “server: cloudflare” with a cf-ray ID. Our own site, rankdough.com, returned Apache, so the setting doesn’t apply to us.

Where is the Agent setting in Cloudflare?

Log in to the Cloudflare dashboard and pick your domain. Go to Security, then Settings, filter by Bot traffic and open “AI Crawler and Agent access policies”.

You’ll see Search, Agent and Training listed with a choice for each:

1. Block: blocks that group on every page.
2. Block on pages with ads: blocks it only where Cloudflare detects ads.
3. Allow (do not block): lets it through.

Source: Cloudflare bot solutions docs, “AI Crawler and Agent access policies”, checked 3 November 2026.

Take a screenshot of what’s there before you change anything. Cloudflare’s changelog applies the new Agent default to new domains. On an older domain someone may still have switched on the old “Block AI bots” toggle and forgotten. The screenshot tells you where you started.

The Free plan gets these options too. Cloudflare’s changelog says all customers, including Free, can manage the three groups separately.

What happened when we sent an agent through a live site?

Rank Dough ran three booking prompts through a student accommodation site we work on, using an AI agent driving a browser. The prompts were a studio near King’s College London for September 2027, the cheaper of two properties near the University of Manchester, and the price of one named London property with an enquiry.

We logged each step the agent took and where it stopped, and sent no real enquiries. Every test used the name “TEST Rankdough” so the sales team could ignore it.

Source: Rank Dough agent test log, 3 November 2026.

What the agent saw, step by step

These are the pages the agent worked through on the first prompt, captured on 3 November 2026 with the site’s brand blurred. The red boxes mark where it moved on and where it stopped.

So Cloudflare let the agent in. The site’s own form kept it out. I expected the Cloudflare setting to be the story, and it wasn’t on this site. The tick box was.

If Cloudflare lets agents in, what else stops them?

The tick box CAPTCHA is the first one. It exists to stop robots, and an agent is a robot working for a real person. Every enquiry form on the test site used one: the form on the property page, the Enquire Now page and the Speak with us form.

The fix we gave the developer was to swap the tick box for an invisible check, such as reCAPTCHA v3 or Cloudflare Turnstile in managed mode. Real users stop seeing a puzzle, and the server still scores each submission for spam. We haven’t rerun the test since that change, so treat it as the plan, not a result.

Filters that don’t work in the web address are the second. The studio filter and the distance sort worked as URL parameters, so the agent could jump straight to them. The move in month only worked inside the Filters pop up, so the agent couldn’t filter to September 2027 directly.

Wrong data is the third, and it hurts people as much as agents. The “Premium Studio” Enquire Now page showed “Silver 2 Bed Apartment Plus” in its summary box. A person might spot that. An agent will send the enquiry for the wrong room.

The good news on that site: the forms are plain HTML that post with a security token, not built in JavaScript. That makes them easy for agents to read and easy to label later with WebMCP, a proposed web standard that tells agents what a form does.

Buttons and fields with no labels are the fourth. Agents read the page structure as well as the words, and OpenAI says its Atlas browser uses ARIA roles and labels to work out what a button or form field does. A button that is only an icon, or a drop down built from plain boxes, gives the agent nothing to go on.

The fix is ordinary accessibility work: a real label on every field and a clear name on every button. Don’t stuff those labels with keywords. Adrian Roselli, an accessibility consultant, warns that SEO tricks in ARIA make sites worse for people who use screen readers, and the same labels serve both readers.

Facts the agent can’t lift are the fifth. Price, availability and room type should sit in plain text and in structured data on the page, not only inside an image or a pop up. That’s also the cheapest agent work, because it already helps you in search.

Sources: OpenAI Publishers and Developers FAQ, checked 3 Nov 2026; Adrian Roselli, OpenAI, ARIA, and SEO, Oct 2025, updated Jun 2026.

Who was the agent, and did it reach checkout?

The agent was Claude, the AI assistant, driving the browser built into the Claude desktop app on my computer. It wasn’t a separate shopping bot. It searched, sorted, opened pages and read the forms the way an agent does for a person.

I’ll be plain about one thing. On this first run the agent read the page code and web addresses more than it clicked like a shopper. That makes it a kinder test than a browsing agent that clicks every button, which tends to get stuck earlier on pop ups and cookie banners.

It didn’t reach a checkout, because there isn’t one. A student accommodation site doesn’t take card payments online, it takes enquiries. There are two routes: Enquire Now, which stopped at the tick box, and Book Now, which opens a six step application with education, address, emergency contact and guarantor steps.

The application asks for a date of birth, a guarantor and an emergency contact before anyone replies. An agent acting for a student in London or Mumbai also has to notice the American defaults and change them. That’s a lot of steps to get right without a person checking.

What I took from it: the agent did the research part well. Price, distance and room type were found in seconds. It’s the hand over to the business, the form, where it broke, and that’s the part most SEO work never touches.

What happens on a big retail site like Currys?

We ran the same kind of test on Currys UK on 3 November 2026, because it sits on Cloudflare and has a real online checkout. The prompt was simple: find wireless earbuds under £100 and get them to checkout. We ran it twice, once from a cloud server and once from a normal desktop browser.

From the cloud server the agent never saw a product. The home page and the search page both returned a Cloudflare block, “Sorry, you have been blocked”, with a 403 status. Several agents, ChatGPT’s agent among them, browse from remote servers, so this is the first thing many of them meet.

From the desktop browser it got further. The home page passed a Cloudflare check after a few seconds. The search results page then showed a “Verify you are human” box, and the agent stopped there, because it doesn’t solve CAPTCHAs.

The way round was the search box itself. Typing “earbuds” showed product suggestions with prices, and the agent picked Sony WF-C510 earbuds at £39.99 straight from the list. It’s the same lesson as the student site: the more a page shows in plain text, the further an agent gets.

Adding them to the basket opened a pop up offering a three year Instant Replacement plan for £13. The agent had to read it and choose “Continue without Instant Replacement”, or it would buy something nobody asked for. A second panel then suggested a smartwatch and a cable.

The basket and checkout loaded with no more checks. Guest checkout asked for an email address, and the box to join Currys Perks for marketing was already ticked. We stopped there, because the next step needs real personal details.

An agent acting for a real person has to untick that box, or the customer lands on a mailing list they never chose. That’s nothing for a person and an easy miss for an agent.

Source: Rank Dough agent test log, Currys UK, 3 November 2026. No order placed and no personal details entered.

So on a big retailer, Cloudflare was the story after all. On the student accommodation site it let the agent in and the form stopped it. On Currys the same service decided which agents got in at all, depending on where they came from and which page they asked for.

How far did each agent get?

Rank Dough scores every agent test on four steps, because a visit is not the result. An agent that finds the page and can’t send the enquiry is a lost lead that never shows in your analytics.

Source: Rank Dough agent test logs, 3 November 2026. One run per site. The cloud agent on Currys was blocked at step 1 with a 403.

Which setting should you choose?

This is my view, not a tested rule. If you sell, take bookings or collect enquiries, allow Agent on the pages where people buy or enquire. Those are the pages an agent visits for a customer who has already decided to act.

Training is a separate call. Blocking it stops your content going into model training and doesn’t stop a customer’s agent from booking. Plenty of owners will block Training and allow Agent, and that’s a fair split.

Publishers who live on ad revenue have a real reason to block agents on ad pages. My reading is that an agent reads the page without looking at the ad, so the visit earns nothing. For them the default makes sense. For a business that sells something, it usually doesn’t.

There is a middle way. Some agents now sign every request so a site can check who they are. Cloudflare calls them signed agents, and the method is an open draft standard written by engineers at Cloudflare and Google. Letting verified agents through while you keep challenging unknown bots is a fairer rule than one switch for everything.

It has a gap. An agent driving a person’s own browser, like the one in our tests, looks like that person and carries no signature. So the Cloudflare setting is only half the check. The form test is the other half.

Sources: Cloudflare, The age of agents, 28 Aug 2025; IETF draft, HTTP Message Signatures for automated traffic, 2 Mar 2026.

How fast is AI agent traffic growing?

Fast, from a small base. HUMAN Security, which measures bot traffic across its customers’ sites, says traffic from AI agents grew 7,851% year over year in 2025. Its monthly reports for 2026 show the curve getting steeper, not flatter.

Chain the monthly changes together and agent traffic in September 2026 was more than four times its March level. Most of that came in one month: September rose 209.7% on August, with Meta’s new Muse agent going straight to number one.

The index is our calculation. We chained HUMAN’s reported month on month changes from March 2026, so it covers HUMAN’s customers only and moves if HUMAN revises a month. Ecommerce took about 60% of agent traffic in September and travel 17.1%, and booking a room sits close to both.

Sites are pushing back at the same time. The share of agent traffic HUMAN saw blocked went from 4.3% in March to 21.9% in August 2026. That’s about one agent visit in five turned away, before anyone checks whether the agent was working for a customer.

The money side is moving too. Adobe says AI-referred visits to US retail sites converted 38% worse than other traffic in March 2025. By March 2026 they converted 42% better, and by July 2026, 60% better.

Those are shoppers arriving from AI assistants, not agents buying on their own. It’s still the same shopper getting used to letting AI do the legwork, and the next step for that shopper is letting it send the enquiry too.

McKinsey’s forecast is the big number: up to $1 trillion in US retail revenue run through agents by 2030, and $3 trillion to $5 trillion worldwide. It’s a forecast, not a measurement, so I’d read it as a direction rather than a promise.

Sources: HUMAN Security, 2026 State of AI Traffic report (26 March 2026) and State of Agentic Traffic reports (May to October 2026); Adobe blog, 16 April 2026; Digital Commerce 360, 19 August 2026; McKinsey, The agentic commerce opportunity, 17 October 2025.

Who owns the agentic web on your team?

John Mueller at Google asked this on LinkedIn, and Joe Hall answered it in a longer piece. As I read his argument, SEOs shouldn’t own the agentic web. SEO owns whether an agent can find you and understand you. Developers own the parts where it acts: logins, payments, forms and APIs.

Our test fits that split almost exactly. Everything SEO touches worked: robots.txt, crawlable pages, prices in plain text, filters in the web address. Everything that belongs to the developer stopped the agent: the CAPTCHA, the room data behind the form and the six step application.

So I’d change the question. It isn’t who owns it, it’s who checks the hand over. Somebody has to run an agent from the first search to the submitted form, and right now that job sits between two teams and belongs to neither.

The developer side is where the new standards are landing. MCP lets an AI assistant connect to a business’s own tools. A2A lets agents talk to other agents. WebMCP, an early preview in Chrome and still a draft, lets a page tell an agent what a form does and which fields it needs, with a few attributes on the form rather than a rebuild.

In the comments under that debate, Natzir Turrado called this AX, agent experience, sitting next to UX. I like the name because it puts the form at the centre, not the ranking.

For a shop there are three more names to know. They all deal with the checkout, and none of them is a clear winner yet.

Sources: WebMCP repository; Google, Under the Hood: UCP, 11 Jan 2026; Stripe, Agentic Commerce Protocol; AP2 core concepts. All checked 3 Nov 2026. The views are mine, not tested.

The order matters more than the standard. Fix the CAPTCHA, the labels and the data first. A new protocol on a form an agent can’t submit changes nothing.

Where is the web heading?

This is my reading, not a forecast. The web is turning into one page with two readers. People still read, compare and decide who to trust. Agents read the same page for facts they can act on: price, dates, room type and the next step.

The finding part looks a lot like SEO today. An agent finds you the way AI search finds you, through pages it can crawl and facts it can lift. That’s why I don’t see agents as a threat to SEO. They’re one more reader we have to write for.

The new part is the last mile. Up to now a website’s job ended when a person landed and filled in the form. With agents the form is the shop window and the till at once, and a tick box that costs you nothing with people can cost you every enquiry an agent tries to send.

The simple truth is that most sites will block agents by accident long before they decide to welcome them. The ones that check the hand over now will get the enquiries the others never see in their analytics.

What I’d do this month

1. Check whether your site is on Cloudflare and screenshot the AI Crawler and Agent access policies page before changing anything.
2. Allow Agent on the pages where people book, buy or enquire. Decide Training separately.
3. Replace tick box CAPTCHAs on enquiry forms with an invisible check.
4. Make the main filters work in the web address, and check the data on each form matches the page it came from.
5. Run one agent from a real customer prompt to a submitted test form every quarter, and log where it stops.

Is Cloudflare blocking agents on your site?

Tick what’s true for your site. The answer tells you where to look first and links to the part of this post that explains it.

The verdicts are Rank Dough’s reading of Cloudflare’s published defaults and our one test, not a measured model. Without JavaScript, the same checks are:

1. Site on Cloudflare and domain added after 15 September 2026, with ads on key pages: Agent is likely blocked on those pages. Open the Agent setting.
2. Site on Cloudflare and nobody knows who set the bot settings: screenshot the AI Crawler and Agent access policies page before changing anything. See where it is.
3. Tick box CAPTCHA on the form: agents stop there even when Cloudflare lets them in. See what else stops agents.
4. Site not on Cloudflare: this setting doesn’t apply, but the form still might block agents. Check the headers.

What doesn’t this post prove?

It’s one site, one agent and one day, 3 November 2026. A different agent, or the same agent on another day, may get stopped earlier or later.

The agent was Claude reading pages more than clicking them on the student site, so a browsing agent that clicks every button may stop earlier. On Currys we tested one product on one day and haven’t seen its Cloudflare settings, so we can’t say which rule blocked which run. The growth figures come from HUMAN Security’s and Adobe’s own customer data, not the whole web, and McKinsey’s 2030 figure is a forecast.

We haven’t seen that site’s Cloudflare dashboard yet. The agent got through, but we can’t tell you whether the Agent setting is on Allow or whether the site’s pages simply don’t count as ad pages. That’s the next thing we’re checking with the site’s developer.

We also haven’t rerun the test after the form fix. When we do, we’ll add the result here with the date.

Frequently asked questions about Cloudflare and AI agents

Sources

• Rank Dough agent test log, student accommodation site on Cloudflare, three prompts, 3 November 2026
• Cloudflare changelog, New options to manage AI traffic, 1 July 2026: developers.cloudflare.com/changelog/post/2026-07-01-ai-traffic-options/
• HUMAN Security, State of Agentic Traffic, September 2026: humansecurity.com/learn/blog/september-2026-agentic-traffic-triples-meta-muse-debuts-at-no-1/

The post Is Cloudflare blocking AI agents from your site? appeared first on Rank Dough Digital.

Rank Dough compared 6 AI SEO agencies working out of Dublin, ours included, and charted 8 of their client sites in Ahrefs. Over the last 12 months 3 of the 8 sites grew and 5 fell, and only 3 of the 6 other-agency clients came with a dated start.

Not sure which client story is closest to your business? Take the quick check. Pick one and the matching section opens, with a question to put to any agency.

Which Dublin AI SEO agencies did we compare?

We compared 6: BeFound SEO, Arekibo, Hatch Digital, Wolfgang Digital, Webjuice and Rank Dough. Rank Dough is ours and based in Loughrea, Co. Galway, not Dublin, so it is disclosed and checked on the same evidence as the rest.

Source: agency websites and case study pages, read 29 October and 2 November 2026; Ahrefs Site Explorer, estimated organic visits a month, pulled 2 November 2026. Awards as published by Arekibo on 21 October 2026.

What happened to their clients’ organic traffic?

We pulled estimated monthly organic visits for each client’s whole domain from Ahrefs Site Explorer on 2 November 2026. Every chart runs to October 2026, so the last 12 months compare October 2025 with October 2026.

Source: Ahrefs Site Explorer, estimated organic visits a month, subdomains included, pulled by Rank Dough on 2 November 2026.

3 of the 8 sites grew over the last 12 months. First Accounts went from 20 to 277 visits, Beacon Recruitment from 41 to 485 and Capcon Engineering from 153 to 446. The other 5 fell: Millies by 4%, Hostelworld by 23%, Aer Lingus by 41%, Mater Private by 46% and Pure Green by 52%.

The three that grew are small sites that started from under 160 visits. The five that fell are the bigger ones. That is a pattern in 8 sites, not a finding, and I have not tested why.

First month in chart is January 2022, or the first month the site had traffic (Beacon Recruitment, December 2024). Estimated organic visits a month, Ahrefs Site Explorer, pulled 2 November 2026.

One method note. Ahrefs estimates traffic from rankings and search volumes for the whole domain, so it can disagree with a client’s own analytics. Where the two differ, the agency’s figure may still be right, but it needs dates and a domain to be checked.

What do BeFound’s clients show?

Capcon Engineering is an Irish firm that designs and installs siphonic drainage for data centres, hospitals and airports. BeFound took on its SEO in May 2022, when the site had 151 visits a month. It had 571 in August 2026 and 446 in October. On three-month averages that is 111 to 488, 4.4 times, so BeFound’s “about 4x” holds.

It was not a straight line. Capcon fell from 369 visits in March 2025 to 119 in November 2025, then climbed back. In a niche this small, a few rankings move the whole chart.

First Accounts is an online accountancy in Maynooth. BeFound started in December 2025, when the site had 18 visits a month after more than two years of decline. It had 277 in October 2026. BeFound says Google clicks rose 4x in 10 months and that AI assistants now name the firm for “best online accountant in Ireland”. We did not test the AI claim.

What do Arekibo’s clients show?

Mater Private Network is a private hospital group. Arekibo built a Medical Conditions Hub with content attributed to its consultants. Arekibo reports 117% more whole-site traffic, 620% more hub traffic and 500+ AI Overview appearances between mid 2023 and early 2026.

Ahrefs shows more than the claim for that window: 18,214 visits in June 2023 and 54,198 in February 2026, 3.0 times. The site peaked at 58,383 in October 2025. It has fallen every month since March 2026 and had 31,710 in October, 46% below a year ago.

The award covers the rise, not what came after. I have not tested why the fall happened. AI Overviews answering health questions on the results page is one candidate, and it is a hypothesis, not a finding.

What do Hatch Digital’s clients show?

Beacon Recruitment is a new recruitment and hospitality consulting business. Hatch named it, built the site and has run SEO and content for two years. The domain first shows traffic in Ahrefs in December 2024, with 1 visit. It reached 589 in August 2026 and had 485 in October.

Hatch’s 10.3x is measured against the first 28 days of a new site, so almost any working site would show a large multiple. The useful part is the shape: under 200 visits for its first 19 months, then a jump from 189 in June 2026 to 468 in July. Hatch’s 146% rise in AI-source traffic can’t be checked in Ahrefs.

What do Wolfgang Digital’s clients show?

Millies is an Irish beauty retailer up against multinational sites. Wolfgang paired a custom GPT with Millies’ in-store beauty staff to write mid-funnel content and reports 58% more sessions and 138% more organic revenue year on year. It gives no dates.

Ahrefs shows a clear step up: 31,942 visits in January 2023 and 60,481 in January 2024. Since then the site has held between 42,407 and 64,328 outside November to January, with peaks of 78,190 in November 2024 and 108,547 in December 2025. October 2026 had 59,783, 4% below October 2025.

What do Webjuice’s clients show?

Pure Green is a US juice bar franchise with 40+ locations. Webjuice says templated location pages, Google Business Profiles and content lifted its organic traffic 400% in one year. No dates are given.

Ahrefs does not show a 400% year for puregreen.com. The best 12 months in our chart is June 2024 to June 2025, 6,949 to 11,053 visits, 1.6 times. The site had 4,658 in October 2026, 52% below a year earlier. The gain may sit in Google Maps or on location pages Ahrefs does not count, but the case study does not say.

What do Rank Dough’s results show?

Rank Dough has no Irish client we can chart yet. The two sites below are my work before Rank Dough, at two Dublin brands, so I do not count them as agency clients.

At Aer Lingus I worked through iProspect from January 2021 to March 2022. Traffic went from 88,779 to 440,720 visits a month, 5.0 times. Travel reopened over the same months, so I do not credit that to SEO alone. At Hostelworld I was in-house from May 2022 to January 2025. Whole-site traffic went from 801,048 to 956,300, up 19%.

Both sites are down on a year ago: Aer Lingus by 41% and Hostelworld by 23%. I left both long before that, and travel sites in our UK comparison fell in the same months.

What do these charts tell an Irish business?

Read together, the 8 charts show five patterns an Irish business should expect from SEO and AI search work. Each comes from the charts above. They are my reading of the data, not tested rules.

Which of these client stories is closest to your business?

Pick the story closest to your business. The matching section opens below, with a question to put to any agency.

How should you choose an AI SEO agency in Dublin?

Four checks follow from the data above. They are my judgement. I have not tested whether they predict results.

1. Ask for the client’s domain and the month the work started. BeFound gave both for its two clients; 2 of the other 4 agencies gave neither.
2. Check the claim in Ahrefs or ask for a Search Console export. Pure Green’s 400% does not show in Ahrefs, while Capcon’s 4x does.
3. Ask what happened after the case study ended. Mater Private is 46% down on a year ago, after the window its award covers.
4. Match the client to your business. A US franchise, a hospital group and a Maynooth accountant need different work, and the closest client tells you more than the biggest number.

Price matters too. Only Webjuice and Hatch publish euro prices, from €499 to €3,500 a month. Ask the others for a range before the first call.

What do people ask about AI SEO agencies in Dublin?

Who is Rank Dough, and what are the limits of this list?

Rank Dough is a small team. Our site names 2 people, Roman Sadowski and David Mason. We trade as Fortunata and work from Loughrea, Co. Galway, not Dublin. We put ourselves in this list because Dublin businesses compare us with Dublin agencies.

I have worked in search since 2004. My Irish record is Aer Lingus at iProspect from January 2021 to March 2022 and Hostelworld in-house from May 2022 to January 2025. Neither was a Rank Dough engagement, so I do not count them as agency clients. This article is published by Rank Dough, which appears in the table, so check every claim against the named domain in Ahrefs.

The limits are clear. Traffic is estimated at domain level from one pull on 2 November 2026. We did not measure sales, leads or AI citations. There is no before-and-after test, start dates are unknown for 3 of the 6 other-agency sites, and each agency is judged on the one or two clients it chose to publish. We also checked Riordan SEO, Wasper Media, VooDoo Digital, Cian Murphy Digital, Dotsquares, Web Tonic and SeoProfy and left them out, because we could not read a named client with a result, or a Dublin base, on their pages.

Sources

• Ahrefs Site Explorer, estimated monthly organic traffic for 8 client domains, pulled by Rank Dough on 2 November 2026
• Arekibo, Best Use of SEO, Digital Media Awards 2026, Mater Private figures
• Google Search Central, Do you need an SEO?

The post Best AI SEO Agencies in Dublin (2026): Compared on Client Results appeared first on Rank Dough Digital.

Rank Dough compared 6 UK travel AI SEO agencies, ours included, and charted 11 client sites in Ahrefs. Only 4 of the 6 had client results we could check, and 9 of the 11 sites charted ended above their first month.

Not sure which of the 8 travel SEO problems matches your site? Take the quick quiz. Pick one and the matching section opens, with a question to put to any agency.

Which UK travel AI SEO agencies did we compare?

We compared 6: Adido, Ink Digital, SALT.agency, north9, Boost Brands and Rank Dough. Rank Dough is ours, so it is disclosed and scored on the same evidence as the rest.

Source: agency websites and trade press, checked 12 October 2026. Years trading counts from founding year to 2026. Team sizes are as stated by each source, and sources conflict for Ink Digital and north9.

What happened to their clients’ organic traffic?

We pulled estimated monthly organic visits for each client’s whole domain from Ahrefs Site Explorer on 12 October 2026. All 11 charts run to June 2026.

Source: Ahrefs Site Explorer, estimated organic visits a month, subdomains included, pulled by Rank Dough on 12 October 2026.

Over each chart window, 9 of the 11 sites ended above their first month. Girls’ Guide to the World grew 5.1 times, Cheaper Than Cars 4.4 times and Grand Prix Grand Tours 3.6 times. Mercury Holidays (0.8 times) and Oliver’s Travels (0.9 times) ended below where they started.

Aer Lingus’s June 2026 figure (196,387) is far below May (579,333) and may be an incomplete month.

Sites tied to Adido, Ink Digital, north9, Hostelworld and Aer Lingus all fell between March and May 2026. That looks like a market-wide pattern, not one agency’s doing. It is a hypothesis, and I have not tested why.

What do Adido’s clients show?

SPL Villas had 667 visits in March 2024, the month its partnership with Adido started on 1 March (Nitravelnews). It reached 1,205 in June 2024, a peak of 1,986 in March 2026 and 1,118 in June 2026. The peak is 3.0 times the March 2024 figure.

Latin Routes shows the opposite. It had 719 visits in January 2025 and 255 in March, after a rebrand and new website that Travel Weekly covered in February 2025.

Adido says it prepared the site for launch and now leads the SEO. I do not have Adido’s start date, so this is a pattern, not proof.

What do Ink Digital’s clients show?

Girls’ Guide to the World is the clearest result. It had 605 visits in January 2024, a peak of 3,226 in February 2026 and 3,061 in June 2026, which is 5.3 times the start. Ink does not publish its start date, so I have used January 2024 as the first month of the chart only.

Mercury Holidays went from 65,333 visits in January 2024 to 140,319 in August 2025, then 50,358 in June 2026. Ink’s case study marks the project completed on 8 April 2025 and reports 181 percent more keyword impressions. Ocean Florida peaked at 32,672 in September 2024 and had 18,307 in June 2026.

What do north9’s clients show?

Rainforest Cruises rose from 28,242 visits in January 2022 to 132,958 in May 2025, then fell to 53,212 in June 2026. north9’s card says organic traffic rose 326 percent year on year but gives no dates. Oliver’s Travels reached 261,623 in June 2024 and had 78,974 in June 2026.

Grand Prix Grand Tours is the most volatile. It had 78,248 visits in May 2022, 9,172 in August 2022 and 99,758 in April 2026.

north9 claims a 16,057 percent traffic rise, and the client quote says about 1,500 visits a month grew to over 20,000. Ahrefs shows no month below 6,028 between January 2022 and June 2026, so a 1,500 start would predate our chart.

What do Rank Dough’s results show?

Rank Dough has one client in the data, Cheaper Than Cars. The other two sites, Aer Lingus and Hostelworld, are my work before Rank Dough.

At Aer Lingus I worked at iProspect from January 2021 to March 2022. Site traffic went from 89,003 to 435,642 visits a month, 4.9 times. Travel reopened over the same months, so I do not credit that to SEO alone. The site peaked at 895,506 in June 2024 and had 196,387 in June 2026.

At Hostelworld I was in-house from May 2022 to January 2025. Whole-site traffic went from 801,060 to 954,250 visits a month, up 19 percent. It peaked at 1,522,154 in May 2025, after I left, and had 883,602 in June 2026.

Cheaper Than Cars is a Rank Dough client. The earliest file I have is from February 2023, when the site had 649 visits. It had 4,619 in June 2026, but I have not confirmed the engagement dates, so I make no claim from it.

What makes travel SEO hard?

Travel SEO is hard because sites are international, built on old systems, database-heavy and seasonal. Eight topics follow, and under each are the items we have seen. Five of the eight come from my own work.

Which of these travel SEO problems does your site have?

Pick the problem closest to yours. The matching section opens below, with a question to put to any agency.

How should you choose a travel AI SEO agency?

Four checks follow from the data above. They are my judgement. I have not tested whether they predict results.

1. Ask for the client’s domain and the dated start of the work. Only 1 of the 8 other-agency sites we charted came with one.
2. Put the agency’s chart next to the market, not next to zero. Sites tied to 5 different agencies or teams fell between March and May 2026, so a flat line may be a good result.
3. Ask how it protects traffic through a migration or relaunch. Latin Routes fell from 719 to 255 visits in two months after its relaunch.
4. Check a percentage claim against Ahrefs. Grand Prix Grand Tours’ 16,057 percent needs a start month nobody published.

What do people ask about UK travel AI SEO agencies?

Who is Rank Dough, and what are the limits of our own record?

Rank Dough is a small team. Our site names 2 people, Roman Sadowski and David Mason. We trade as Fortunata and work from Loughrea, Co. Galway and from 700a London Road, Essex, SS0 9HQ.

I have worked in search since 2004. My travel work is Aer Lingus at iProspect from January 2021 to March 2022, Hostelworld in-house from May 2022 to January 2025 and Cheaper Than Cars as a Rank Dough client. The first two were not Rank Dough engagements, so I do not count them as agency clients.

We have traded for 1 year and we are not travel-only. Our site lists 3 awards from my earlier work, and I have not linked the winners lists yet, so I do not count them here. This article is published by Rank Dough, which appears in the table, so check every claim against the named domain in Ahrefs.

The limits are clear. The traffic is estimated at domain level from one pull on 12 October 2026. The search results check covers 3 UK searches in one snapshot. There is no before-and-after test, and 7 of the 8 other-agency sites have no known start date.

Sources

• Ahrefs Site Explorer, estimated monthly organic traffic for 11 client domains, and Ahrefs SERP Overview for 3 UK searches, pulled by Rank Dough on 12 October 2026
• Google Search Central, How to move a site
• Nitravelnews, SPL Villas partners with Adido, start date 1 March 2024

The post Best UK Travel AI SEO Agencies (2026): Compared on Client Results appeared first on Rank Dough Digital.

Claude can handle most of the hands-on work in SEO, from keyword research and content briefs to technical audits, publishing and Search Console analysis, once you connect it to your tools and give it written rules. Rank Dough tested every step of this guide on its own site between 5 and 9 October 2026, and took one new page from an 80-keyword research pull to a Google-indexed URL in a single day.

The guide covers which Claude features matter for SEO, how to set them up, how to use Claude for each SEO job, the prompts to start with, 12 hacks and the limits. Boxes marked “From our build” show what happened when we did it ourselves, with real screenshots.

What can Claude do for SEO, and what can’t it do?

Claude can do the research, writing, analysis and repetitive build work in SEO. It reads data you give it or pulls through connectors, writes to rules you set, runs code to process files, and can publish drafts to a CMS through its API.

It can’t replace judgement. It doesn’t know your margins or which customers matter most, it can be confidently wrong about numbers it hasn’t checked, and without connectors it has no live data at all.

Source: Rank Dough’s working view from client and in-house SEO work, October 2026.

What are the best Claude SEO hacks?

These 12 come from running the workflow, not from a list of tips. We scored each one on impact and on the effort to set it up; start with the bubbles in the top-left corner. Each hack is explained in context in the sections below.

Which Claude features matter for SEO?

Ten Claude features do the work in an SEO workflow. Each one replaces a step you would otherwise do by hand: exporting a file, opening a tab or re-explaining your brief.

Source: Claude features as used by Rank Dough, 5 to 9 October 2026. Custom connectors via remote MCP are available on Free, Pro, Max, Team and Enterprise plans, per Anthropic’s help centre.

How do you set up Claude for SEO?

Set up four things before your first SEO task: data, projects, rules and memory. It takes an afternoon, and every later task gets faster because you stop re-explaining.

1. Connect your data. In Claude, go to Settings, Connectors, and add the tools you already pay for. A tool without an official connector can often be added as a custom connector by URL.
2. Create one project per site. Put the brief, target markets, past audits and recent exports inside it, so answers come from that site’s context.
3. Write your rules as skills. Page structure, tone, banned phrases, what needs a source, who approves a publish. A skill loads these rules every time the task comes up.
4. Turn memory on and correct it. When Claude gets a preference wrong, say so once; it carries the correction into later chats.

How do you do keyword research with Claude?

Use Claude to pull, sort and question keyword data, not to invent it. Without a connected tool or an uploaded export, its search volumes are guesses.

1. Give Claude your seed topics: your services, what customers ask in emails and calls, and your own category language.
2. Pull volume, difficulty and the top 10 for each seed through a connector, or upload an export from your keyword tool.
3. Ask Claude to group the keywords by search intent and map each group to one existing or new page.
4. Check the search results for every keyword you plan to target before you commit.
5. Add the follow-up questions each keyword leads to (cost, time, minimums, comparisons), even at zero volume, because buyers and AI tools ask them next.

Source: Ahrefs Keywords Explorer and SERP Overview through the Ahrefs connector, United States, pulled 5 and 9 October 2026. Keyword difficulty is the figure the connector returned; the Ahrefs app can show a different, newer figure.

How do you write content briefs and articles with Claude?

Let Claude assemble the brief and the first draft, and keep the facts and the final edit with a person. A good brief from Claude lists the target keyword, the questions the page must answer, what the current top results cover and what your page can add that they don’t.

1. Pull the People Also Ask questions and the headings of the top results for the keyword.
2. List the facts only you have: your prices, your data, your results, your process. These are what make the page worth citing.
3. Draft against your skill rules: answer first, one idea per paragraph, every heading a question.
4. Fact-check every number against its source, including titles, meta descriptions, image alt text and FAQ answers.

How do you plan content and avoid cannibalisation with Claude?

Ask Claude to scan your own site before it plans anything new. A plan built without knowing what already exists will happily commission pages you already have.

Source: WordPress REST API scan of rankdough.com (29 pages, 19 posts) and Yoast focus keywords, 9 October 2026.

Keep the plan in an artifact rather than a chat message. Claude can build a sortable dashboard where every row shows its keyword, volume, status and live URL, and you can update it as pages go live.

How do you build and publish pages with Claude?

Let Claude build and publish drafts through your CMS’s API, and keep the publish decision with a person. On WordPress, an Application Password lets Claude create drafts, set categories and upload images without your main login.

1. Copy the structure of a page that already works on your site, so the new page matches without a redesign.
2. Open with the keyword in the H1 and one fact sentence: who you are, where, and what you do.
3. Create it as a draft, read the preview, then publish.
4. Add internal links from the menu and related pages only once the page is live.

Location pages need care. Two near-identical pages for two countries is a doorway page, so each version needs facts that only apply to that market.

How do you run technical SEO and UX audits with Claude?

Give Claude the crawl data and let it do the sorting and joining. Upload your crawler exports and it can list non-indexable URLs, missing or duplicate titles, redirect chains and orphan pages, then join each issue to its Search Console clicks so you fix what has traffic first.

Claude can also check the live site. In its own sandbox it can run a headless browser that opens public pages, takes screenshots and measures what actually renders, such as font sizes, image weights and whether content needs JavaScript to appear.

For a UX review, ask Claude to walk the site like a buyer, page by page, on desktop and at phone width. The question for each page is whether a prospect understands the offer, trusts it and knows what to do next.

Ask for the output as a report you can reopen and share: a verdict and score first, then an annotated screenshot per page, each finding with its fix and an evidence label, and the fixes in order.

How do you analyse Search Console data with Claude?

Connect Search Console, or upload exports, and ask questions in plain English. With a connector, “is this URL indexed?” and “what dropped against last month?” each take one sentence.

Then ask Claude to turn the data into a dashboard you can reopen. The one below mirrors the Search Console performance view for a client’s site: 16 months of weekly clicks and impressions, built in one chat from connector data.

Compare periods by page to see every new or changed URL at once. Pages with high impressions and a weak click-through rate usually need a title that matches the query better.

Source: Google Search Console, rankdough.com property, pulled through the Rank Dough GSC connector on 9 October 2026. Performance data lags about three days.

How do you write titles, meta descriptions and schema with Claude?

Ask for options with limits stated: titles under 60 characters with the keyword near the front, meta descriptions under 155 characters with a number or a reason to click. Then generate FAQ schema from the exact text of the FAQ on the page, so the two never disagree.

Check every one by hand before publishing. Google judges the page, not the tool, and a wrong figure in a meta description is still a wrong figure.

How do you run an AI visibility audit with Claude?

Treat AI answers as a channel you sample, not a ranking you read. In SparkToro’s study of 2,961 runs across ChatGPT, Claude and Google’s AI, the chance of getting the same brand list twice was under 1 in 100, and the same list in the same order about 1 in 1,000. What holds steady is the share of answers that mention a brand.

So Rank Dough built its AI audit around visibility as a rate. Claude researched how practitioners and tracking tools run these audits, checked the evidence on sample sizes, and turned it into a five-phase process: scope, prompt library, baseline sampling, diagnosis, and an action plan where every fix ships with a test.

1. Build a prompt library. 8 to 12 intent clusters with 5 to 8 phrasings each, tagged by funnel stage, market and branded or non-branded.
2. Sample each cluster on each AI surface about 100 times. Standard binomial maths puts that at about ±10 points at 95% confidence. Spread runs over several days, logged out, from the target country.
3. Report each surface and market separately. Show the rate, the run count and the interval. Never a rank position and never one blended score.
4. Diagnose where the brand drops. Recognised by the model, matched to the buyer’s question, retrieved as a source, and finally selected.
5. Tag every number. Observed, own sample, third-party estimate or modelled, so the client knows what to trust.
6. Pair each action with a test. Control clusters, sample size and a readout date, before anyone claims a result.

Our own ChatGPT tests found that ranking first in Google did not mean being cited, and that the pages it cited stated prices, minimums and turnaround in plain sentences. The full method and results are in How ChatGPT chooses which websites to cite.

How do you get screenshots and visuals for SEO content with Claude?

Use real, annotated screenshots for how-to content, so readers can follow the steps. Claude’s headless browser can capture public pages, and Claude in Chrome can capture tools behind your login, such as Ahrefs and Search Console.

Headless browsers stop at bot checks: Google returned a reCAPTCHA and the Ahrefs checker a “verify you are human” screen in our tests. Then run a privacy pass on every capture before it goes public.

What are the limits of Claude for SEO?

• Data is only as good as the source. Our Ahrefs connector and the Ahrefs app gave different difficulty scores for the same keyword.
• It can be confidently wrong. A chart and a claim in this guide’s first draft were both wrong until the fact check.
• Logins and bot checks stop it. Headless browsers can’t pass reCAPTCHA or reach logged-in tools; use Claude in Chrome or your own exports.
• It makes changes fast, including wrong ones. A menu link to an unpublished page went live briefly before Claude caught it on the live site.
• It has no business context unless you give it. Margins, priorities and which customers matter stay with you.

Every limit has the same fix: a person reads the live preview before anything is published, and every number names where and when it was pulled.

What should you set up first?

Start with data, then rules. Connect your keyword tool and Search Console, create a project for your site, and write your page and content rules as a skill.

Then run one small job end to end, such as a single page from keyword to indexed URL, and fix the process before you scale it. Rank Dough runs every workflow on its own site first, and I’ll update this guide with the three-month Search Console numbers for the pages built here.

FAQ

Sources

• Rank Dough build record: Ahrefs pulls, WordPress build and Search Console data for rankdough.com, 5 to 9 October 2026
• SparkToro, New research: AIs are highly inconsistent when recommending brands or products, January 2026: sparktoro.com
• Google Search Central, Google Search’s guidance about AI-generated content: developers.google.com

The post How to use Claude for SEO: the complete guide, built from our own workflow appeared first on Rank Dough Digital.

To get cited by ChatGPT, put one complete fact sentence in the first 50 words of your page and repeat its numbers in the meta description. Rank Dough scored 23 pages across six buying prompts in September 2026: 11 of the 15 pages ChatGPT cited stated their offer in the first 50 words, against 1 of the 8 Google-ranking pages it skipped. Ranking didn’t help. The brand at #1 in Google for both head terms earned zero citations.

I spent three weeks finding out why, across three rounds of testing and two GPT set-ups (the ChatGPT app and a GPT agent workspace). The short answer is that ChatGPT cites pages it can lift a fact from, not pages that rank.

How do you get cited by ChatGPT?

You get cited when ChatGPT can lift a complete fact sentence from the text it already holds about your page. In our September 2026 test every verbatim quote in ChatGPT’s answers came from a page’s opening, a heading or its meta description. Nothing came from body copy further down, and nothing came from schema.

So the job has four parts, in the order the model works through them.

1. Be known. ChatGPT seeds brand names it already knows into its searches. If nobody talks about you on Reddit, in reviews or in editorial lists, you fight for the generic slots. An accommodation brand we track was named in 2 of 168 category answers in October 2026, and 5 of 5 times when asked about by name. The category answers were built from “top 10 websites” articles the brand wasn’t in.
2. Match the words it searches. ChatGPT rewrites your buyer’s prompt and adds criteria: minimum order, turnaround, price, location. State those facts in plain text, not in a PDF or behind a quote form.
3. Get into the pool. Your page has to come back in ChatGPT’s own search results. Google position does not carry over.
4. Win the extract. Put the offer, with numbers, in the first 50 words under the H1, and repeat the numbers in the meta description. 11 of the 15 cited pages did this. 1 of 8 ranking pages did.

Everything below is the evidence for those four steps, and the changes we’d make first.

Source: Rank Dough citation test workbook, 1 September 2026; GPT workspace audit, 7 September 2026.

How did we test which websites ChatGPT cites?

We picked one vertical, custom softball jerseys in the US, where a brand we know ranks #1 in Google for the two head terms. I wrote six buying prompts, ran each in a fresh ChatGPT chat with web search on, and recorded every linked source URL.

Then I fetched every cited page plus every page in Google’s US top 10 for the matching keywords as raw HTML, and scored all 23 pages on 22 attributes: opening content, headings, stated facts, FAQ, schema, speed, robots.txt and more.

Methodology: all citation data comes from ChatGPT runs on 1 September 2026, re-tested with a GPT agent in a separate workspace on 7 September; Google positions come from Ahrefs SERP Overview (US, checked 30 July and 12 August 2026); page attributes come from raw HTML fetches on the same days, which we used as a stand-in for what ChatGPT’s crawler sees. That it doesn’t run JavaScript is our observation from these tests, not something OpenAI documents.

Source: Rank Dough citation test log, ChatGPT app, 1 September 2026; Ahrefs SERP Overview, US, “custom softball jerseys” checked 12 August 2026 and “custom slowpitch softball jerseys” checked 30 July 2026. Prompts are not Google keywords, so each was compared with the closest of the two head terms.

The brand at #1 in Google for both head terms, custom softball jerseys and custom slowpitch softball jerseys, was cited on none of the six prompts. ChatGPT cited 15 other pages instead, 33 times in total because several were cited on more than one prompt, and none of them ranked in Google’s top 10 for those terms. Google’s own AI Overview for custom softball jerseys didn’t cite the brand either: it listed 12 other URLs.

Why does ranking #1 in Google earn zero ChatGPT citations?

Because ChatGPT is not reading Google. It rewrites your prompt into its own search queries, pulls candidates from its own index, and picks sources from the text extracts attached to those results.

Our reading is that Google rank never enters that pipeline. In the 7 September rerun, the #1 Google brand didn’t appear in any of the six retrieval outputs, so it lost before its page was ever judged. That rerun used a GPT agent in a separate workspace, not the ChatGPT app, so treat it as a strong pointer rather than proof.

The page had a second problem as well. The #1 Google result was a shopping grid whose first 1,300 lines of its raw HTML were a cart drawer, navigation and product counts.

The first commercial fact on the page, a 3 to 4 week lead time, sat at line 1,303 of 1,754. A competing page that got cited stated its lead time at line 45.

Position inside ChatGPT’s own result list did not decide selection either. In a follow-up audit the model cited results sitting at positions 12 and 14 and skipped another competitor holding position 1 in ChatGPT’s own result list with three URLs. The #1 Google brand wasn’t in that list at all.

The #1 brand had a third problem we only found in a later audit. Its www address returned fetch errors to the model while the non-www address loaded, and its robots.txt and llms.txt failed the same way. A hostname fault kept it out of a thorough search pass before any page was judged. Check your own: fetch both hosts the way a crawler would, not in a browser that follows redirects for you.

Source: GPT-6 Astra audit, 5 runs, September 2026; verified live against both hosts.

My working model after three rounds, labelled as a hypothesis, is a four-stage funnel: the brand has to be recognised, the page has to match the model’s rewritten queries, the URL has to enter the result set, and the extract has to win selection. In this test, most pages failed at stage two or four, not stage three.

Source: Rank Dough citation test and GPT workspace audit, September 2026; line positions from raw HTML fetches, 21 September 2026.

What did cited pages have that ranking pages didn’t?

We scored the 15 cited pages against 8 pages that ranked in Google’s top 10 but were never cited. Six attributes separated the groups.

Word count did not: cited pages averaged 1,726 words against 2,143 for the ranking pages, and the two shortest cited pages ran 137 and 213 words. Price being shown, Product schema and robots.txt access did not separate the groups either, because every page in both groups already allowed the AI crawlers.

Source: Rank Dough page scoring workbook, 23 pages, September 2026. These are observed gaps in one vertical, not proven causes.

The pattern behind the numbers is extraction. ChatGPT’s answer bullets repeated fact sentences from the cited pages near word for word: a 21-day production claim from one page’s opening, a 10-player minimum and a $109.97 price from another’s meta description, a 3 to 5 week turnaround from a single sentence on a 137-word page.

Pages that gave the model a liftable sentence got quoted. Pages that made it hunt got skipped, and when the model could not find a number it either wrote “request a quote” or, twice in our logs, invented a turnaround figure the page never stated.

How does your page compare with the pages ChatGPT cited?

Tick the traits your key page already has. The checker shows how common each trait was among the 15 pages ChatGPT cited and the 8 Google-ranking pages it skipped, and which gaps to close first.

What does ChatGPT actually read when it cites a page?

A cached text extract, not your live page. Across two audited runs ChatGPT cited 5 of 17 and 6 of 16 returned results while opening zero pages.

Each result carried a stitched extract of the page: headline, intro copy, bullets, prices and FAQ text, with per-page crawl ages running from two weeks to a year. Five things follow from that.

1. Your structured data is invisible at the moment of citation. The extract carried no JSON-LD, no schema.org markup and no meta tags in any audited run. Keep schema for Google; stop selling it as a ChatGPT citation factor.
2. Bot walls block page opens, not citations. One cited page in our test sat behind a password wall and another behind a bot challenge, and both kept earning citations from stale extracts. The same lag works against you: a fixed page stays invisible until the crawler returns, so plan in weeks.
3. Template order looks like a citation factor, though we haven’t tested it directly. Whatever your theme prints first in the HTML is what fills the extract. A cart drawer and a mega menu in front of your content push the facts out of the window the model reads.
4. Internal links still move citations. In one audit the brand’s price list was cited only after the model followed a link from the collection navigation. A page nobody links to from a crawled page doesn’t get an extract. And once you fix a page, expect weeks before the new extract shows: crawl ages in our audits ran from two weeks to a year.
5. It reads your pages against each other. On a branded “is this brand legit” prompt, ChatGPT named an accommodation brand, cited its own site, then told the reader to check carefully before paying. The reason: one page stated a booking fee, another said the service was free, and the footer’s founding year didn’t match the company register. Fix contradictions on your own site before you chase mentions elsewhere.

Source: GPT workspace audit ledger, 7 September 2026; web.run interrogation, 21 September 2026; Rank Dough AI visibility tracking, 168 ChatGPT runs, October 2026.

What does a ChatGPT retrieval pool look like?

About two thirds of every commercial pool we audited was Reddit, Wikipedia and PDFs. Each pool held exactly 12 Reddit results, and that held in 12 of 12 pools across five unrelated verticals. That leaves 7 to 12 slots per pool for supplier pages, and those are the slots you’re competing for.

Who fills the non-Reddit slots changes by vertical. In dental it was provider pages, in standing desks editorial only, in toys retailers. Audit your own category before you decide which page type to build.

One more thing we expected to matter and didn’t. Running the search at thorough depth instead of minimal returned the same pool. Depth changed which entries got cited, not who was in the pool. So entry is decided early, on retrieval; effort by the model only reshuffles selection. One audit, so a pattern, not proof.

Source: Rank Dough retrieval pool audits, GPT-6 Astra rounds 6 to 8, September 2026, five verticals. Counts come from one model surface; the 12-Reddit pattern is observed there, not a documented rule of the ChatGPT app.

How do fan-out queries decide which pages compete?

In neither audited run did ChatGPT search the words we typed. Both times it rewrote the prompt and appended buyer decision criteria: minimum order, turnaround, pricing, youth sizing.

Those appended terms are retrieval keys. A page that states its minimum and turnaround in plain text matches the queries the model actually sends; a page that keeps those facts in a PDF or behind a quote form competes for nothing.

The rewrites also seeded brand names the model already knew from training into the searches, which means well-mentioned brands enter the candidate pool by name while unknown brands fight for the generic slots. Our reading, not yet measured: that makes third-party mentions a retrieval problem, not a vanity metric.

You can see this layer first-party and free. Bing Webmaster Tools added an AI Performance report in February 2026 and expanded it in June 2026 with intents, topics and citation share.

Its grounding queries show the reformulated searches the AI ran before citing or skipping you on the Microsoft surface. Verify your site, pull 30 days of grounding queries, and you have a proxy for the searches AI systems run. It comes from Microsoft’s AI, not ChatGPT, so treat it as a guide rather than a mirror.

Source: Rank Dough audit logs, September 2026; Bing Webmaster blog, February and June 2026.

What should you change on your pages first?

Is ChatGPT SEO different from Google SEO?

Yes, at the selection step. Google ranks pages. ChatGPT rewrites your prompt, pulls candidates from its own index and cites from cached text extracts without opening the page. Across our six prompts its citations overlapped with Google’s top 10 zero times.

What carries over from Google SEO: crawlable pages on a working host, plain-text facts, one URL per intent, internal links. What doesn’t: schema (absent from every extract we audited), word count (cited pages averaged 1,726 words to 2,143 for ranking pages) and Google position itself.

“ChatGPT SEO” is the extract and the fan-out terms. There’s no holy grail beyond that, and anyone promising a specific citation is guessing.

Source: Rank Dough citation test workbook, 1 September 2026; web.run interrogation, 21 September 2026.

What stayed unstable across our tests?

Three things refused to settle, and I’d rather tell you than sell you certainty. Listicles got cited in one round, where a competitor’s dated comparison post became the source for a third brand’s recommendation, and got skipped in another round a week later.

Reddit threads entered 12 of 12 commercial retrieval pools we audited but rarely won selection, so pool presence and citation are different outcomes. And only 5 of the 18 URLs from the first run reappeared in the 7 September rerun, which means a single test proves nothing and any agency promising a specific citation is guessing.

Run your prompts across several days, track visible retrieval, citation and accuracy as three separate numbers, and judge changes against that baseline.

Source: Rank Dough citation logs, 1 and 7 September 2026; retrieval pool audits, September 2026.

Limits of this test. The attribute table comes from a single run on 1 September. We scored the HTML as it was on the day, while ChatGPT read extracts up to a year old. And we didn’t change a page and re-test it, so everything here is correlation, not proof of cause.

How can you run this test on your own brand?

Run the same three steps we did: ask, record, compare. It takes an afternoon for a first pass, and the first pass tells you which pages to fix before you invest in tracking.

1. Write six buying prompts. One each for a head term, a niche term, a natural question, a best-of, an audience-specific ask and an attribute such as fast turnaround.
2. Run each in a fresh ChatGPT chat with search on. Record every cited URL, whether your brand is named, and any fact the answer states about you.
3. Pull Google’s top 10 for the matching keyword. Compare the two lists; overlap tells you whether ranking is helping at all.
4. Read the cited pages’ openings and meta descriptions. Note which facts the answer lifted and where they sat on the page.
5. Repeat over several days before judging anything. Only 5 of 18 cited URLs repeated between our runs a week apart.

For tracking over time, we now run every prompt three times per check and average the result, once a week, on ChatGPT, Google AI Overviews and AI Mode, with mentions and citations logged as separate numbers. Our AI visibility audit process explains the set-up.

FAQ

Sources

• Rank Dough citation test workbook: 6 ChatGPT prompts, 23 pages scored, 1 September 2026; second-model rerun ledger, 7 September 2026; retrieval pool audits, September 2026; web.run interrogation, 21 September 2026
• OpenAI, Overview of OpenAI Crawlers: developers.openai.com/api/docs/bots
• Microsoft Bing, Introducing AI Performance in Bing Webmaster Tools, February 2026: blogs.bing.com

The post How to get cited by ChatGPT (and why ranking #1 in Google didn’t) appeared first on Rank Dough Digital.

Yes for ChatGPT, on the evidence so far, but nobody has proved cause. Ahrefs found branded web mentions correlate at 0.664 with AI Overview visibility across 75,000 brands, against 0.218 for backlinks.

Rank Dough’s own test on six custom softball jersey prompts found zero overlap between the pages ChatGPT cited and Google’s top 10. Every study I could find is correlational, so treat mentions on the right pages as the best bet you have, and test it on your own brand.

Why did I put this question to four expert AI profiles?

I wanted more than the usual “mentions are the new links” posts. So I took one question to SuperMe, a network that builds AI profiles of experts from their talks, writing and notes, and lets you compare several of them on one question.

I asked the AI profiles of Ethan Smith (Graphite), Kevin Indig (Growth Memo), Alex Halliday (AirOps) and Eli Schwartz the same thing: do unlinked brand mentions get a brand cited in ChatGPT and Google AI Overviews?

Read their answers for what they are. An AI profile summarises what an expert has published and said.

It isn’t the person. Every time I quote one below, I label it “AI profile”, and I’ve only kept numbers I could trace back to the expert’s own published study. Where a profile gave a figure I couldn’t verify, I left it out.

Then I checked their answers against our own test data from September 2026. That’s the part you won’t find anywhere else.

What does the public data say about mentions versus links?

Three published studies carry most of the weight in this debate. They all point the same way, and they all measure correlation.

Source: ahrefs.com/blog/ai-overview-brand-correlation/, semrush.com/blog/the-ghost-citations-study/, airops.com/lp/the-state-of-aeo-report. Checked 6 October 2026.

I pulled each figure from the original study page on 6 October 2026 and dropped any number I couldn’t trace back to its source.

The Ahrefs gap is the headline. Mentions, linked or unlinked, track AI Overview visibility three times more closely than backlink counts.

The AirOps figure tells you where those mentions live: on other people’s sites, not yours. So for most brands, the work that moves AI visibility happens off-site.

What did our own softball jersey test find?

In September 2026, Rank Dough ran a small citation test on a custom softball jersey store. We put six buying prompts through ChatGPT, logged every result it retrieved and every page it cited, and compared that with Google’s top 10 for the same terms. We reran it a week later with a retrieval audit.

Source: Rank Dough first-party test, custom jersey vertical, 1 and 7 September 2026. Small sample, one vertical.

Three things stood out. First, ranking first in Google predicted nothing here.

Second, Reddit showed up in every retrieval pool but rarely got the citation. Third, ChatGPT never searched the words we typed.

It rewrote each prompt and added buyer criteria like “minimum order”, “turnaround” and “youth sizing”.

For a small brand, that last point matters most. The pages ChatGPT pulls are the ones that answer those rewritten queries. If a roundup page already ranks for “custom softball jerseys low minimum order”, a mention on that page with your minimum stated next to your name is worth far more than a link from a big site that never comes up.

Where do the four expert AI profiles agree?

On four points, all four profiles landed in the same place.

1. The page matters more than the domain. A mention helps only if it sits on a URL the engine retrieves for your prompts. The Ethan Smith profile put it plainly: a mention on a big site’s page that never gets cited does nothing for you.
2. “Go win Reddit” is weak advice. All four profiles rejected it. That matches our test: Reddit made every pool and still lost most of the citations.
3. Track mentions and citations as two numbers. A citation means the engine used your page. A mention means a buyer read your name. They move separately.
4. Nobody has proved cause. The Kevin Indig profile said it straight: a cross-sectional chart can’t tell you whether mentions drive visibility or whether strong brands collect both.

Where do the expert AI profiles disagree?

The first split is about links. The Eli Schwartz profile said he tells companies to pause paid link campaigns and has never seen it hurt them.

But he and the Ethan Smith profile both made one exception: brands with little awareness still benefit from earned links. That’s the exception that matters for small ecommerce brands, and it’s the one the big studies can’t speak to, because their samples lean on well-known brands.

The second split is about measurement. The Ethan Smith profile argued that most citation tracking measures a layer that often doesn’t exist.

When ChatGPT answers without searching, there are no citations at all. The answer then rests on how often the model has seen your brand next to your category across the web.

Tracking tools that force a search can show you a very different picture from what a buyer sees. The Alex Halliday profile accepted that point, which narrows what any citation study can claim.

Why does a ghost citation matter if you still get the link?

A ghost citation is an AI answer that links your page as a source but never names your brand. In the Semrush and Kevin Indig study, 61.7% of AI appearances worked this way.

Because buyers act on names, not footnotes. Kevin Indig’s breakdown of a clickstream study of 846,000 US Google sessions from February and March 2026 found that in AI Mode, 74% of users picked the top result and 64% clicked nothing at all. If the answer links your page but names a competitor, the competitor gets the sale.

The Alex Halliday profile described what gets a brand named rather than just linked: the brand name sits right next to the claim. “Brand X is the best option for youth leagues because the minimum order is 10” survives extraction.

“A recent study found the minimum is 10” lets the model lift the fact and drop your name. Our own test agrees.

The sentences ChatGPT lifted whole were the ones that carried brand, product, audience, turnaround and price in one line.

What should a small ecommerce brand do first?

Here’s the order I’d work in, based on the evidence above and our own test. Treat the order as my judgement, not a proven sequence.

1. Map the pages first. Run your real buying prompts several times on each engine. Record every third-party page that gets retrieved and cited. Each engine pulls a different set.
2. Rank those pages by how often they appear. Work down the list from the top.
3. Ask for a sentence, not a logo. Get your brand named next to one specific use case and one number: “best for youth teams, minimum order of 10”.
4. Repeat one claim. Five pages that say you’re best for the same use case beat twenty loose mentions.
5. Fix your own pages for the rewritten queries. Put minimums, turnaround, pricing and sizing in plain text near the top, so third parties have something to quote.
6. Keep earned links in the plan. If nobody knows your brand yet, links still help your own pages win the follow-up questions.

Is a mention on this page worth chasing?

Use this before you spend time or money on a placement. Answer five questions about the page you are considering and you get a verdict, the reason for each answer with a link to the evidence in this article, and the exact line to ask the site to publish.

How would you prove mentions cause citations?

You’d need a test with controls, and nobody I could find has published one. Two of the AI profiles described designs, and they control for different things.

The Ethan Smith profile splits the cited URLs into treatment and holdout groups. The Kevin Indig profile compares the brand against matched brands that get no new mentions, and checks that any lift shows up only on the prompts you targeted.

The test I’d run combines both:

• One low-awareness brand in one narrow category, such as custom softball jerseys.
• Cited URLs split into treatment and holdout within high, mid and low citation-frequency buckets.
• Targeted prompts compared with untargeted prompts for the same brand.
• Two or three matched brands of similar size as controls.
• Mention rate, citation rate and ghost citation rate tracked separately, on forced and unforced runs, several times a week.
• At least three months, with every placement date logged, and tracking kept running after outreach stops to see how fast any lift fades.

A lift that appears only on targeted prompts, only after placement and only for the test brand would be real evidence. Anything less is another correlation. Rank Dough is planning this test on the custom kit vertical, and I’ll publish the results here.

Frequently asked questions about unlinked brand mentions

Sources

• Ahrefs, An Analysis of AI Overview Brand Visibility Factors (75K Brands Studied): ahrefs.com/blog/ai-overview-brand-correlation/
• SuperMe AI profiles of Ethan Smith, Kevin Indig, Alex Halliday and Eli Schwartz, answers collected 6 October 2026 (AI summaries of each expert’s published work, not the experts themselves)
• Rank Dough first-party citation test, custom jersey vertical, 1 and 7 September 2026

The post Do Unlinked Brand Mentions Get You Cited in ChatGPT? What the Data Shows appeared first on Rank Dough Digital.

Rankdough creates non-commodity content by building a data source first and writing second. In this travel example, an agent visits 200 Barcelona hostels through Google Maps, logs six dated observations per property, and that log feeds pages for a query cluster whose head term, “best hostels in Barcelona”, gets 27,000 global searches a month at keyword difficulty 0 (Ahrefs, 4 October 2026).

No competitor holds the same numbers, so Google and AI assistants have a reason to pick our page over the other ten.

What is non-commodity content?

Non-commodity content is a page whose core input a competitor or a language model can’t reproduce. If you could write the page from the top ten results, it’s commodity, however long or well edited it is.

You get a non-commodity input from one of four places:

1. Proprietary data. Orders, outcomes, prices, response times, observations you logged yourself.
2. Direct experience. You did the thing, and you have photos, numbers and the mistakes to show it.
3. Access. Interviews, expert quotes, a community you run.
4. A defended position. A contrarian view with evidence behind it and a name on it.

Google calls the extra value “information gain”. Ethan Smith of Graphite describes the method as converge, then diverge: cover what competitors cover, then add something they don’t have (Maven). Most content teams stop after converging.

Why does commodity content stop working?

Commodity content loses traffic once Google’s quality systems re-evaluate it, and AI assistants have no reason to cite one copy over another. Lily Ray tracked 220+ sites that were public customers of AI content tools and measured each site’s drop from its peak organic traffic.

Source: Lily Ray, published May 2026, Ahrefs data checked against SISTRIX (Search Engine Journal).

Method: Lily Ray identified the sites from public customer lists of AI content platforms and compared each site’s organic traffic with its peak in Ahrefs, checked against SISTRIX.

AI writing the words didn’t sink those sites. Volume did: thousands of pages carrying nothing the web didn’t already have, at a scale Google’s scaled-content systems notice.

How do you find a data source nobody else has?

Start with the question your audience asks that no ranking page answers with a number. Then find a source you can turn into that number at a cost competitors won’t pay.

For solo travellers in Barcelona, the unanswered question is concrete: what does the walk from the night bus to my hostel look like at 2am? Every safety guide repeats the same city-wide advice. None measures the last 400 metres.

Three tests decide whether a source is worth building:

1. It answers a real question with a number. Minutes, counts, shares, dates.
2. It costs effort to collect. If a competitor can copy it in an afternoon, it’s commodity next quarter.
3. It gets better with time. A dataset you refresh builds a history that nobody can rebuild later.

Google Maps passes all three. The imagery and routing are public, but nobody has scored 200 hostel approaches against a fixed checklist, dated every image and checked a sample by hand.

How did Rankdough build the Barcelona agent?

Rankdough built an observation agent with a traveller’s profile: someone arriving alone, late, with a bag. The profile decides the checklist, and the agent runs the same checklist at every property:

1. Time the walk from the nearest night-bus stop and metro entrance at a 02:00 departure.
2. Classify the last stretch of street: main street, side street or passage.
3. Look for street lamps and a 24-hour reception sign.
4. Check whether the door is visible from the corner.
5. Record the capture date of every image and flag anything it couldn’t see.

Change the profile and the checklist changes with it. A wheelchair user’s agent counts steps and looks for ramps; a family’s agent checks lift signage and pushchair access. The pipeline stays the same.

What does the dataset look like?

Each visit becomes rows in the agent’s log, and the log is the only thing the agent remembers. The table below is an illustrative example of the format; the Barcelona pilot hasn’t collected real observations yet.

Source: illustrative example of the Walk the Map log format, Rankdough, October 2026.

Agent method: the agent scores Street View imagery against the checklist in two independent passes, times walks with the Routes API, and a person checks every disagreement plus a random sample of 20 properties.

The agent never edits a past row. When it goes back next quarter, it writes new ones, so the log turns into a history: a hostel that gained a reception sign, a street with new lighting, a property that closed. That history is the part competitors can’t buy.

Can readers ask the agent directly?

Yes. The agent answers questions from its log, so readers get a conversation that stays inside the evidence. An illustrative exchange:

You: What’s the walk like from the night bus to Example Hostel B?

Agent: Eleven minutes from the nearest night-bus stop at 02:00. The last stretch is a passage under three metres wide. I couldn’t make out street lamps in the imagery, and I saw no 24-hour reception sign. The images are from November 2024, and no person has checked this one yet.

You: Would you feel safe walking it?

Agent: I can’t tell you that. I haven’t walked it at night, and daytime images don’t show how busy or lit a street is at 2am. If you’re arriving late, the hostel I visited in Gràcia has a main-street approach, a reception sign and a six-minute walk from the bus.

Each sentence the agent says points to a row in the log. That traceability is what you’d want a journalist, a forum moderator or an AI assistant to see before they cite you.

How does the agent collect each observation?

The agent works through one visit in six steps, and every step writes to the same log:

1. Find the property address with the Places API.
2. Look at the street with the Street View Static API: three headings at the entrance and one from the nearest corner, each with its capture date.
3. Score the images against the checklist in two separate vision passes.
4. Send disagreements to a person; agreements go straight into the log.
5. Time the walk from the nearest night-bus stop and metro with the Routes API at a 02:00 departure.
6. Store image date, model version and checker on every row. The agent and the pages read only from those rows.

Where is the line between original data and fabricated experience?

The agent knows what it observed, not how a place feels. That boundary is what makes its answers worth citing.

It can tell you:

• What the approach and the door looked like, and the date of the image.
• Whether street lamps and a reception sign were visible in that image.
• Walking times the routing engine returned, and when it asked.
• Where it couldn’t see anything, which is a finding in itself.
• Whether a person has checked the observation.

It won’t tell you:

• Whether a street is safe, busy or well lit at night. Daytime imagery can’t show that, and it says so.
• Current conditions. Street View imagery can be two to four years old, so every answer carries the date.
• Anything phrased as lived experience. It says “in imagery from March 2025”, never “when I walked there”.

An agent that invents memories produces what any model can produce, so it has no scarcity, and presenting it as testimony runs against Google’s guidance on first-hand, people-first content. An agent that records dated observations produces something only its owner holds.

Which search terms does the data target?

Rankdough doesn’t publish the log as a log. The data feeds pages built for the queries travellers type, and the numbers decide which pages earn the build.

Source: Ahrefs Keywords Explorer, pulled 4 October 2026. Keyword difficulty not reported for “hostels barcelona”.

The table points to three pages, each with a different job:

1. The head term. “Best hostels in Barcelona” is a list page. A list backed by 200 measured night walks is the only one on that results page with information gain.
2. The parent topic. “Is Barcelona safe for solo female travellers?” gets answered with district-level numbers from the log instead of the usual Las Ramblas warning.
3. AI citations. Long questions such as “hostel in Gràcia on a main street near the night bus” go to ChatGPT and Perplexity, where only 12% of cited sources sit in Google’s top 10 (Ahrefs). AI questions average about 23 words against 6 on Google, according to Ethan Smith (Maven).

We checked accessibility terms too and dropped them. At 30 US searches a month for the top query, the cluster can’t justify its own page.

Mentions do the rest. Ahrefs studied 75,000 brands and measured how three signals correlate with AI Overview visibility.

Source: Ahrefs study reported by Search Engine Land.

A dated city report with a published method gives travel forums, hostel associations and YouTube creators something specific to mention. Rankdough treats each quarterly re-run as a new edition and a new reason to cite; we’ll report whether mentions compound once two editions are live.

What does a data source like this cost to build?

One city costs less than a day of consulting. The figures below are estimates until the pilot is billed.

Source: Rankdough build estimate, 4 October 2026, before Google Maps Platform billing.

Before anything goes live, read the current Google Maps Platform terms on storing Places content. The agent keeps its own observations, never Google’s review text. Show Street View attribution wherever an image appears, and give property owners a way to correct an observation, with the image date beside it.

How can you apply this to your own niche?

The same five steps work for any site, with or without an agent:

1. Write down the question your buyers ask that no ranking page answers with a number. For a team-uniform store, it’s real turnaround by sport and month. For a dental clinic abroad, it’s the full cost including flights and nights.
2. Find the source. Your orders, quotes, support tickets, returns, or a public source nobody has cleaned up.
3. Fix the checklist. Same fields, same method, every time. Publish the method.
4. Date everything and re-run it. A history beats a snapshot.
5. Write the pages the queries ask for, with the numbers in the first paragraph and a source line under every table.

Rankdough runs this process for clients whose content has stopped earning traffic or citations. We start with the question, not the keyword list.

FAQs

Sources

• Search Engine Journal: Lily Ray, AI content tools and traffic loss, May 2026
• Google Search Central: Creating helpful, reliable, people-first content
• Ahrefs Keywords Explorer, Barcelona solo-travel and hostel queries, pulled 4 October 2026

The post How to Create Non-Commodity Content for SEO: A Travel Example from Barcelona appeared first on Rank Dough Digital.