Do AI Assistants Recommend UK Local Businesses? Three English Cities, 36 Readings, Three Different Answers

Yes, they name businesses. Asked six buying questions each in London, Manchester and Birmingham on 15 and 16 October 2026, ChatGPT and Gemini produced 36 readings, and 33 of those 36 came back with a list of named sources. The useful part is what happened next: the suppliers named were not the same in the three cities, and they were not even the same kind of supplier. Manchester went to small Manchester domains. London went to UK wide domains with no London named site among the winners. Birmingham mixed a small local domain with a national newspaper. Our own domain was cited in none of the 36.

That is why a single national figure would have been the wrong answer to give you. Three English cities, one country, one question set, three different pictures.

What was measured, and what the three cities returned

The city scan asked six buyer questions per city, on both engines, giving 12 readings per city. It ran on 15 and 16 October 2026. Every reading was usable.

City	Readings	Readings that returned a named source list	Most cited sources in that city
London	12	10	generativeengineoptimizationuk.agency 3, firstmotion.com 3, agentcraft-ai.com 3
Manchester	12	12	mainchester.co.uk 3, popula.co.uk 3, rank4ai.co.uk 3
Birmingham	12	11	get-found.co.uk 3, thetimes.com 3, seocompare.co.uk 2
All three	36	33	0 readings cited aiknowsus.com

Two engines were read and both are named here: ChatGPT and Gemini. The other three surfaces our product covers, Perplexity, Claude and Google AI Overviews, sat outside this scan entirely.

The three cities disagreed, and that is the finding

Manchester behaved the way a business owner expects a local answer to behave. Every one of its 12 readings returned a list, and the top of that list was a Manchester named domain. If you sell in Manchester, there is already a local page holding the slot you want, and you can go and read it.

London did not behave that way. Ten of 12 readings returned a list, and no London named site was among the winners. The engines answered a London question with sources that are British rather than specifically London. For a London business that is a different competitive position: the slot is held by national pages, so the opening is a London specific answer that nobody has published yet.

Birmingham sat between the two. Eleven readings returned a list, one small local domain led it, and a national newspaper was cited in three of the 12. A Birmingham business is competing with a masthead for part of its own city's answer.

Average those three and you get a number that describes none of them. Find out which of the three shapes your own city has before you decide what to publish.

A citation is not the same as being recommended

This matters because the two get reported as one number. What the scan counted is citations: one domain in one reading, deduped, which is the definition every published figure of ours uses. It does not count how many times a business was praised in the prose, and it does not record whether a named supplier was recommended or merely listed.

The size of that distinction is measurable. On the wider run of 16 October 2026, counting one domain per reading gives 20,143 citations across 1,260 readings. Counting raw url occurrences in the same answers gives 36,576. Those two numbers describe the same data, so any visibility figure you are shown needs its counting rule attached before it means anything.

For a local business the working version is simpler. If a domain is cited for your city's question, its page is being read and used. If your domain is not in the list, the engine had no reason to reach for it.

What the wider run says about where those citations go

The city scan sits inside a larger measurement. Across 1,260 readings on 16 October 2026, 82.0% of the 20,143 citations went to a page on a vendor's or an agency's own website. Official platform documentation took 12.3%, news or large publishers 3.2%, social or community 1.9%, and review directories 0.7%.

That 0.7% is the number worth holding onto, because the usual advice to a local business is to fill in more directory profiles. The pages being cited here are overwhelmingly somebody's own pages, and the three cities above make the same point from the other end.

The counting rule, the question set and the first run it came from are set out in the 597 reading benchmark, which is where to go if you want to argue with the method rather than take it from a summary.

What this scan does not tell a UK local business

Three limits, each of them a question we have not answered yet rather than a caveat on the ones above.

The six questions per city were in one category: a buyer looking for help being named in AI answers. There is no restaurant number here, no hotel, gym, salon, trade, automotive or estate agent number. For your own trade the question set has to be written in your trade's words and run the same way.

Three English cities were measured. Leeds, Glasgow, Bristol and Edinburgh were not, so nothing above is a figure for them. For an unmeasured city the national picture is what you have, with the label kept on it.

Two of five surfaces were read. Our product measures ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews; this scan covers the first two.

What a UK local operator can do with this on Monday

Write the six questions a customer in your city actually types, with no business name in any of them, including yours. Ask them on ChatGPT and Gemini in fresh sessions, and write down for each answer whether a source list came back and which domains were in it. Twelve readings is an afternoon, and it tells you which of the three shapes above your city has.

Then read the pages that won. mainchester.co.uk, get-found.co.uk and seocompare.co.uk are not large companies, and the reason they were cited is visible on the page. If the winner in your city is a national page, that is the better position of the two, because nobody has taken the local answer yet.

If what comes back is a list of agencies rather than businesses like yours, the buying question underneath it is a different one, and working out which tier of agency you belong in before you approach anybody is the cheaper conversation to have first.

We should be straight about who counted this. We are the company that sells the measurement and the publishing work that follows it, so the 36 readings above were counted by somebody with a commercial interest in you caring about them. Every count on this page is printed with its denominator for exactly that reason: you can disagree with the conclusion and keep the arithmetic.

What British operators ask after reading a city breakdown like this

Do ChatGPT and Gemini actually recommend UK local businesses?

They name them. In 33 of 36 readings across London, Manchester and Birmingham on 15 and 16 October 2026 the answer came back with named sources, and in Manchester and Birmingham those included small local domains. Whether a named business was recommended or simply listed is a separate judgement that the citation count does not make for you.

How often did an AI answer name a local competitor rather than us?

In all 36 of those readings the domains cited belonged to somebody else. Our own domain was in none of them. That is the baseline we publish about ourselves and it is the one we are judged against.

Is London easier or harder than Manchester?

They are different, not ranked. Manchester already has local domains holding the slot, so you are displacing a page that exists. London returned UK wide sources with no London named winner, so the London answer is open and the work is to publish it first.

Does a mention mean the business was recommended?

Not on its own. A citation records that a domain was used as a source for the answer. A recommendation is a claim in the prose. Log them as two separate columns when you run your own check, because they move for different reasons.

Why only two engines and three cities?

Because that is what was measured on those dates, and a figure for anything else would be a guess with a decimal point on it. The scan is repeatable: six questions, two engines, 12 readings a city, and the next round carries its own date and sample size.

Your city already has a list and you have not read it

Thirty three of 36 readings came back fully formed, with named businesses in them, and customers in London, Manchester and Birmingham are reading those lists today. Three cities, three different lists. There is no British answer to go and check, only your own city's, and the afternoon it takes to read it is the smallest piece of work on this page. AI Knows Us pulls that list on a repeating schedule and writes the page that argues for a different name, and the fairest thing we can tell you about it is that across these 36 readings it had not earned a slot for itself either. Start with your six questions. Everything else follows from what comes back.