Do AI assistants read British directories, measured across 64 blind readings on 22 October 2026

They read some and ignore most. On 22 October 2026 we put 192 British buying questions to ChatGPT and to Gemini in 32 blind batches each, 64 readings in total, with the United Kingdom fixed by a line written into every question because a location parameter is accepted and then ignored. Counting one host per answer, those 64 readings produced 1,431 host readings across 1,022 distinct hosts, after 9 cited urls belonging to the data provider's own proxy infrastructure were excluded rather than presented as publishers. Checkatrade was cited 0 times. Yell 0. Rated People 0. MyBuilder 0. Bark 0. Trustpilot was cited once, and once more on its uk subdomain. The single most cited British host in the whole set was gov.uk, in 16 of the 64 readings.

The two engines contributed very unequally. ChatGPT's 32 readings supplied 1,343 of those host readings and Gemini's 32 supplied 88, across 86 distinct hosts with no single host appearing more than twice.

What was measured, and what it does not cover

Each batch put six British buying questions to one engine in one session. The subjects were the ones a business owner asks when buying marketing help, choosing a supplier, checking a rule or working out whether assistants name them: agency selection, pricing, vetting, search, paid media, procurement, verticals, cities and rules. The market was fixed inside the prompt text, in the form "I am a business owner in London, the United Kingdom. Answer for the United Kingdom only, and use pounds for any price."

A 33rd batch was run on each engine and is excluded from every figure here. It names our own company inside the question, which makes it a ceiling measurement rather than a visibility result, and averaging it into a blind denominator would flatter us by two readings. So the denominator throughout is 64, and 40 of those 64 are core readings in which no question names any firm at all.

Two limits matter before you read the numbers. The questions were about buying business services, so a zero against Checkatrade is a finding about who gets cited when somebody asks about marketing and supplier selection, and it is not a measurement of what happens when a homeowner asks for a plumber in Leeds. And this is two engines. Perplexity, Claude and Google AI Overviews were not in this reading.

The British hosts the engines actually reached for

The state outranks every commercial British directory

Add gov.uk, business.gov.uk, ico.org.uk and asa.org.uk together and the public sector was cited in more British readings than every consumer directory and agency marketplace in the table put together. This replicates something our earlier work found in a different question set, where UK readings carried gov.uk 57 times and ico.org.uk 28 times while neither appeared in the American set at all.

There is a plain reason for it. The questions that touch a rule have one correct source and the engine goes to it. When a British buyer asks what a supplier must prove, or whether a claim is allowed, the regulator's own page is the answer and nothing else competes with it. A commercial directory has nothing to offer that question.

One British research site was cited 18 times, and it is not an agency

The single top host across the whole British set was whito.co.uk, in 18 of the 32 ChatGPT readings. It publishes research about AI search rather than selling trade services, and the engines treat it as a source. Our earlier United Kingdom work put the same site at 46 citations in a 597 reading pass, alongside rank4ai.co.uk at 39 and aeoscore.co.uk at 32, so this is now the second question set in which the same small British publisher outperforms far larger brands.

Being read is not the same as being chosen, and this site is the sharpest example of the gap we have measured. Counting names in the answer prose, with source lists and link targets stripped out so that a cited url cannot be mistaken for a recommendation, Whito was named in 4 of ChatGPT's 32 readings while being cited in 18 of them. On Gemini it was cited in 1 and named in 1. The firm ChatGPT actually named most often was a different one, Impression, in 10 readings, with Brainlabs in 2 and Rank4AI, Builtvisible and Croud in 1 each. A site can be read constantly and recommended rarely, and a figure that adds the two together describes neither.

Why the famous directories came back empty

• The answer does not need a list page. Asked for suppliers, the engines mostly cited suppliers. Across all 64 blind British readings, 92.9 per cent of the host readings went to a page on a vendor's or an agency's own website, official or platform documentation took 4.5 per cent, news or large publishers 1.0 per cent, social or community 0.8 per cent and review directories 0.8 per cent.
• A profile is thin against an article. A directory entry holds a category, a location and a star rating. The question being asked usually needs a paragraph of reasoning, and the page that contains one wins the citation.
• Login walls and heavy pages cost nothing to skip. Where a listing sits behind an interface rather than in readable text, there is less for an engine to quote.
• Nothing about a directory is specific to Britain. 324 distinct hosts on a .uk domain carried 475 of ChatGPT's 1,343 host readings. British specificity won citations, and the global category directories did not supply it.

What a British business should do with this

The usual advice at this point is to fill in more directory profiles. The measured numbers do not support spending a quarter on that, and we have now measured the review directory class three times and got 0.8 per cent, 0.7 per cent and 0.8 per cent, which is written up in the repeat measurement of the directory finding.

Keep a directory profile where it earns business on its own terms. Checkatrade sends real work to real trades, and a trade platform's fee is justified by won jobs rather than by citations. Treat it as lead supply, price it that way, and do not expand it on the hope of being quoted by an assistant.

Then put the work where the citations actually went:

• Your own pages, carrying the specifics a directory cannot hold. Price, coverage area, what you will not do, the registration number, the insurance cover, the dated case.
• The British facts. Pounds, VAT treatment, the rule you work under, the register you are on. Those are the details that made .uk hosts competitive in this set.
• One page that answers a question rather than describing a service. Across both arms, the pages cited were pages that said something, not pages that listed something.

Questions about British directories and AI answers

Does this mean Checkatrade does not work?

No. It means Checkatrade was not cited in these 64 readings, and these readings asked about buying marketing and business services. The platform's value sits in the enquiries it sends a trade business, which this measurement did not look at.

Why did Gemini cite almost nothing?

Gemini returned 88 host readings across 86 distinct hosts in 32 readings, with no host appearing more than twice, while ChatGPT concentrated hard on a short list. That pattern has held across every reading we have taken: the same question produces a short entrenched source list on one engine and a flat spread on the other.

Is Trustpilot worth it if assistants ignore it?

Yes, for the job it actually does, which is reassuring a buyer who already has your name. The measurement here is about discovery and citation, and a review platform is read by people at a later stage than that.

Should we put our company number and insurance on our own website?

Yes. Everything a trade platform verifies about you is something a cautious buyer and an engine both want to read, and on your own page it is text rather than a badge.

How often does this need re-running?

Monthly is enough to work from current data. These counts are a single date, 22 October 2026, and the engines change their source mix without announcing it.

Can we reproduce this ourselves?

Yes, and it takes an afternoon. Write six questions a British buyer would type, put the market in the question text rather than trusting a setting, ask them on both engines in fresh sessions, and record which domains came back with each answer.

The reason we publish the zeros

Our own company was named in the prose of 0 of all 64 blind readings. Our domain was cited in 2 of ChatGPT's 32 and in 0 of Gemini's 32, and the two that landed were both pages carrying an original British measurement rather than anything describing our product. That is the whole argument of this page turned on its author: the thing that got read was a count with a denominator on it. AI Knows Us sells the reading itself, across ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews, so we have an obvious interest in you caring about these numbers, which is exactly why every one above is printed with the number of readings underneath it.