A Small UK Agency Against a Large Consultancy on AI Search Optimisation: What the Citation Data Says Each One Can Own
Three British agency domains took dozens of citations each in our readings, and none of them is a large firm. Here is what that result does and does not tell a ten person shop.
Published by AI Knows Us · Updated 17 October 2026
The slots in this market are held by self published pages, which is the one structural fact that favours a small British agency. The 16 October 2026 pass asked 630 buying questions and came back with 1,260 readings from ChatGPT and Gemini. Those readings carried 20,143 citations over 4,947 distinct domains, and 82.0% of them landed on a vendor's or an agency's own website. Official platform documentation took 12.3%. Review directories took 0.7%. Eleven days before, on a question set designed separately, those same two categories came out at 85.6% and 0.8%. Outside the documentation layer the field is held by people publishing their own pages, and a ten person shop can publish.
What follows is the measured version of that argument, including the parts that do not favour a small firm. We sell the measurement and the publishing work, so our own position is printed here rather than left out.
The .co.uk sites already holding British slots
whito.co.uk turned up in 46 of the British readings on 15 October 2026, rank4ai.co.uk in 39 and aeoscore.co.uk in 32. Not one of the three is a large consultancy, and between them they hold more of that field than any firm with a floor of staff on it.
The city scan is sharper still. A separate pass on 15 and 16 October 2026 covered ten cities over 120 readings, and our own count came out at zero everywhere. The sources London reached for most were generativeengineoptimizationuk.agency, firstmotion.com and agentcraft-ai.com, on 3 readings each. Manchester reached for mainchester.co.uk, popula.co.uk and rank4ai.co.uk, also 3 each.
Read the London row twice. A domain whose name is the service and the country, on a .agency extension, took a London slot three times. There is no size requirement in that result. The pages winning do so because the domain and the page are specifically about that city and that question.
What a small shop can own that a large firm finds expensive
On 15 October 2026 we ran 102 separate build spec conversations and asked, each time, what a page would have to hold before the engine quoted it in place of somebody else. The counts are the useful part, because they are requests rather than opinions about style.
| Asked for by | What they wanted on the page |
|---|---|
| 102 of 102 | When it was written, when it was last checked, how often that happens |
| 102 of 102 | Prices quoted in the money the reader actually pays in |
| 101 of 102 | A baseline and a re-reading, with the gap between them stated |
| 99 of 102 | The method, set out rather than summarised |
| 98 of 102 | Each engine written as a name, never as a category |
| 83 of 102 | A stated repeat count, or an error bar on the number |
| 80 of 102 | How many observations sit behind the figure |
| 78 of 102 | The question list handed over so a stranger can run it again |
| 68 of 102 | The failures and the results that went the wrong way |
| 64 of 102 | Who benefits commercially, said on the page |
Five of those ten are decisions a small firm can take on a Tuesday afternoon. Publishing the question set so a stranger can re-run it, printing the failures, naming the commercial interest, dating the page in a way that commits you to re-reading it, and quoting a real figure with its denominator all get easier the fewer people have to approve them. A ten person shop that publishes its own bad number has done something most of the field has not.
The flip side is worth saying to a client before a consultancy says it for you. A large firm brings legal review, indemnities, a procurement process that already holds your client's data handling answers, and the capacity to run across forty markets. None of that is what the engines asked for above, and all of it is what a finance director asks for.
The incumbent list is thinner than it looks, and only on one engine
The two engines behave so differently that a small shop should price them as two jobs. ChatGPT gave 176 of its 306 readings on 15 October 2026 to a single domain, while the best supported Gemini domain managed 19 out of 291. One surface concentrates on a short entrenched list, mostly platform documentation. The other has no concentration at all, so there is nothing there to be displaced from.
Six of the ten most cited domains in the 16 October run are a platform answering questions about itself: developers.google.com in 446 readings, support.google.com in 328, developers.google.cn in 325, blog.google in 309, help.openai.com in 269 and developers.openai.com in 230. Writing a page to outrank Google's own documentation on its own subject is money spent for entertainment. The rows below are reachable: semrush.com at 199, theverge.com at 197 and searchengineinfo.org at 147.
One change between the runs is worth attention. Widen the question set from 102 questions per market to 210 and the four specialist AI visibility tools that had been sitting in the top ten drop out of it completely. The long tail of buyer questions is not yet held by anybody in particular.
The number nobody in this category can give you, including us
You will see claims about how fast a small firm gets a client's first new citation. We publish no median time to first citation, because we have not run the study that produces one. What we hold is our own position, and it is thin. The British half of the 16 October 2026 run gave us 3 appearances in 210 answers on one engine and nothing on the other. Totalled over both runs, that engine returned 921 readings with our domain in zero of them.
Treat a supplier's promised position the way Google treats it in its own guidance: nobody controls what an assistant answers, and a firm offering a number one position is selling something outside its control. What a small agency can offer is the work and the reading: a frozen question set, named engines reported separately, a fixed interval, and every round published including the ones that go backwards. How a firm is paid shapes what it does each month, which is why the fee model deciding how an agency behaves is worth settling before the scope is.
What a ten person shop should actually sell
Four deliverables, each checkable by the client without you present.
- A dated baseline with a numerator and denominator per engine and market, zero columns left in. A pooled figure over two engines is a number about neither of them.
- The question set as a file the client keeps, versioned, with the client's brand in none of it.
- Pages answering one buyer question each, with a date, a stated method and a figure a reader can check. That is five of the ten engine requests above, satisfied by the writing itself.
- A correction log. 68 of 102 build specs asked for failures, and publishing a finding you later withdrew is the artefact a large firm's sign off makes slowest.
Questions a small British agency gets asked about AI search optimisation
Do we need our own scrapers to compete with a larger firm?
No, and the measured field says the lever is publishing rather than collection. Thirty questions on two engines, run on a fixed interval and logged row by row, is a reading a two person team can produce. What decides the outcome is whether the pages you publish carry a date, a sample size, a named engine and a figure a stranger can check. That is writing rather than infrastructure.
How long until an assistant names a client after the work starts?
We will not give you a number we have not measured, and neither should a competitor. Set the re-reading interval before the first result is seen, keep the question set frozen between rounds, and report each engine separately, because a Gemini column reading zero on a first round is where that surface starts rather than a verdict on the work.
Is a manual prompt check enough, or does it have to be automated?
A manual round proves the method and an interval proves the direction. One reading is a photograph. What turns it into evidence is the same questions, the same wording, the same engines and the same session conditions on a schedule somebody will keep, which is the argument for automating the collection rather than the judgement.
Are review directories worth building out for a British client?
They were 0.7% of the 20,143 citations recorded on 16 October 2026 and 0.8% of the 10,652 recorded on 15 October 2026. Set a year of directory profile work beside a quarter of dated publishing and that 0.7% is the figure the comparison turns on.
Somebody is already answering this for British buyers, with no sample size anywhere in it
Put the supplier question to an assistant now and the reply comes back fluent, sourced from pages that name no engine and count nothing. Three small .co.uk domains are already inside that answer dozens of times over, which is the argument for a ten person shop publishing rather than pitching. The company behind this page, AI Knows Us, sells that reading on five surfaces and the writing that follows it, and the two engines named above are the only ones either run touched. The method, the question design and the corrections are in the 597 reading benchmark across ChatGPT and Gemini in three markets, which is also what a prospect can use to check whether we hold ourselves to this.