What can be proved about AI search in the United Kingdom, and the four questions nobody has measured
Six British findings with a denominator attached, and the widely quoted numbers that are targets rather than results.
Published by AI Knows Us · Updated 22 October 2026
Six things about AI search in the United Kingdom can be stated today with a count, a date and a denominator. Four of the numbers most often repeated about it cannot, because no reading exists behind them. The difference matters to a British business deciding where to spend, because the four unmeasured ones are the ones used to sell urgency, and the six measured ones are the ones you can act on this week.
The six British findings that carry a denominator
- The engines name British agencies without being asked. On 22 October 2026, 192 British buying questions were put to ChatGPT and to Gemini in 32 blind batches each, 64 readings, with the market fixed inside every question. Counting names in the answer prose, with source lists and link targets stripped so a cited url cannot pass as a recommendation, 22 of the 64 readings named at least one British agency. 40 of the 64 are core readings in which no question names any firm, and 15 of those 40 still named one.
- The two engines invert each other. ChatGPT named a British agency in 13 of its 32 readings and a tool vendor in 9. Gemini named a tool vendor in 13 of its 32 and a British agency in 9. The firm named most often on both was Impression, in 10 of ChatGPT's readings and 8 of Gemini's, while the tools Gemini reached for were Semrush in 9, Ahrefs in 7 and HubSpot in 7.
- British answers are built from suppliers' own pages. Counting one host per answer, the 64 blind readings produced 1,431 host readings across 1,022 distinct hosts, after 9 data provider proxy urls were excluded. 92.9 per cent went to a page on a vendor's or an agency's own website, 4.5 per cent to official or platform documentation, 1.0 per cent to news or large publishers, 0.8 per cent to social or community and 0.8 per cent to review directories.
- The public sector is the most cited British source after one research site. whito.co.uk appeared in 19 of the 64 readings and gov.uk in 16, with support.google.com in 11, ico.org.uk in 7 and business.gov.uk in 6. A 597 reading pass earlier in October found the same public sector shape, with gov.uk at 57 citations and ico.org.uk at 28 in the United Kingdom readings.
- British cities do not behave the same as each other. Six buyer questions in each of London, Manchester and Birmingham, on both engines, produced 36 readings on 15 and 16 October 2026. Manchester's answers went to small Manchester domains, London's went to UK wide sources with no London named site among the winners, and Birmingham mixed a small local domain with a national newspaper.
- Our own company was named in none of the 64 and cited in two. Named 0 of 64. Cited in 2 of ChatGPT's 32 readings and 0 of Gemini's, both times a page carrying an original measurement rather than a product description. A 33rd batch per engine names us inside the question and is excluded from all of these figures, because it is a ceiling rather than a result.
The four questions with no British measurement behind them
Each of these gets quoted with a confident figure attached. None of the figures comes from a reading.
| The claim you will see | What exists |
|---|---|
| A share of British buyers who consult an AI assistant before contacting a supplier | No survey of British buyers behind it. Our readings measure what the engines answer, not what buyers do before they dial. |
| An AI visibility score for a named British industry, such as property or legal | No per industry rate exists in any of our datasets. The vertical questions we asked were generic, so there is no sector figure to publish. |
| A figure for a British city other than London, Manchester or Birmingham | Three English cities were measured separately. Leeds, Glasgow, Bristol and Edinburgh were not, so there is no number for them. |
| Lost revenue per year from not being named by an assistant | No model behind it that anyone has published with its inputs. A number of this shape is an estimate dressed as a measurement. |
When we asked the engines themselves what page should exist for each British subject, they supplied invented statistics to go with it: a share of AI voice for British commercial property, a cost of a regulatory breach, a retention rate for mid market agencies. Every one arrived with a convincing sample size and a time horizon, and none had a reading behind it. One of the specifications said so in its own words, that a target must not be published as a result.
Why the honest version is more useful than the confident one
A British business cannot act on a national share of buyers. It can act on this: on 22 October 2026 a British question about your category returned a shortlist with other people's names in it, assembled mostly from their own pages. That tells you what to publish. A percentage of buyers tells you only to worry.
The same applies to the two words that get merged. Being cited means your domain appeared in the source list behind an answer. Being named means the answer text said your company. They fail for different reasons, they are fixed by different work, and a page that averages them is describing nothing. In our British set, whito.co.uk was cited in 18 of ChatGPT's 32 readings and named in the prose of 4 of them, while on Gemini it was cited in 1 and named in 1. Those four numbers are useful. Their average is not, and a single blended visibility score would have hidden the most interesting result in the whole reading.
What a before and after can honestly show in the United Kingdom
Our own domain is the cleanest available example of a standing start, and it is published on the proof page from a Google Search Console export taken on 14 October 2026. Over the nine days with data, 4 October to 12 October 2026, aiknowsus.com recorded 5,965 impressions and 3 clicks, with the best day on 11 October 2026 at 2,685 impressions and an average position of 41.6. The domain had been registered days before the window opened, with no history and no links.
That is a discovery result and it is not traffic. Impressions mean the pages are being shown. Clicks follow position, and position takes longer than nine days. Said plainly, it is the most useful thing a nine day window can show, and anybody offering a revenue figure from one is offering something the window cannot contain.
How to measure your own British picture in an afternoon
- Write six questions a British buyer would actually type, with no company name in any of them, including yours. Naming a brand in the prompt contaminates the answer.
- Put the country in the question text. A location setting on a search tool is accepted and then ignored, which we measured directly, so the market has to be fixed in the words.
- Ask them on ChatGPT and Gemini in fresh sessions, and treat the two as separate datasets for the rest of the exercise.
- Record two things separately for each answer. Which domains were cited, and which companies were named in the prose.
- Count one domain per answer. Counting every url occurrence instead gives a number almost twice as large for the same data, so the rule has to be stated wherever the figure appears.
- Repeat monthly, so you are working from current data rather than from one afternoon in October.
The full counting rule, the question design and the first run it came from are set out in the 597 reading benchmark. The British before and after version, and what a short window can and cannot read, is in the note on a British before and after test.
Questions about British AI search evidence
How many British buyers use an assistant before they contact a supplier?
No reading we hold answers that, and a figure for it would be an estimate. What is measurable is what an assistant answers when a British buyer asks, and that is worth running for your own category rather than waiting for a national statistic.
Which British sector is furthest behind on AI visibility?
No dataset we hold ranks British sectors, because the vertical questions in these readings were deliberately generic. A sector ranking requires a question set written in that sector's own words, run the same way for each sector.
Why should the market be set in the prompt rather than in a setting?
Because the setting is accepted and has no effect. A location parameter returned a successful response for every value we tried and the answer came from somewhere else entirely, so the only reliable market lock is a sentence inside the question.
Are these findings about all AI engines?
They are about ChatGPT and Gemini. Two engines, named, on stated dates. Perplexity, Claude and Google AI Overviews were not in these readings, and a product measuring five surfaces does not turn a dataset of two into a dataset of five.
How current does a reading stay?
Treat it as a monthly number. Source mixes move without notice, which is why every count on this page carries the date it was taken.
What we would want checked about us
Our own numbers on this page are a British citation count of 2 in 32 readings on one engine and 0 in 32 on the other. We publish them because a company selling measurement should be the easiest one on your shortlist to audit. AI Knows Us runs this reading for a named business across ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews, and the only thing worth buying from anybody in this category is a count you can repeat yourself and get the same answer.