Does improving AI visibility work? What a British before and after test can read, and the slots it cannot move
On 15 October 2026 gov.uk was cited in 57 UK readings and ico.org.uk in 28, and neither appeared in the American set. Scope a British test below those slots or most of your effort lands where nothing moves.
Published by AI Knows Us · Updated 16 October 2026
A British before and after test works, and it works best when it is pointed at the part of the answer that can actually change. A day earlier, on 15 October 2026, 102 buying questions per market produced something the British readings held and the American ones did not: two institutional sources turning up again and again, gov.uk in 57 British readings and ico.org.uk in 28, with no appearance in the United States set. On 16 October 2026 we ran a wider set, 210 questions per market, 420 UK readings, and our own domain was cited in 3 of the 210 ChatGPT readings and 0 of the 210 Gemini readings. Put those together and you have the shape of a British test: a layer of public sector sources that will keep the top of certain answers whatever you publish, and a set of slots underneath it where a measured change is readable.
What was read in Britain, and on which day
London was named inside the British prompts of that earlier run, which also covered the United States and India. What AI assistants actually cite across 597 readings of ChatGPT and Gemini in three markets holds the whole thing. The 16 October run widened the question set to 210 per market in 21 batches, 630 in total, aimed at eight directions the first run missed, including procurement depth, questions that name a competitor, and questions asked before the buyer knows the category exists. The British cut of the first run, with its own counts for the small agency domains that won it, sits in the UK cut of that benchmark.
| Reading | Date | UK detail |
|---|---|---|
| First run | 15 October 2026 | A shared core of questions in every market plus thirty written only for Britain. Two institutional domains recurred here, at 57 readings and 28. |
| Second run | 16 October 2026 | 210 questions per market. 420 UK readings, all usable. Our domain appeared as a source in 3 ChatGPT readings and in none on Gemini. |
| Both runs, our own domain | 15 and 16 October 2026 | 3 of 597, then 6 of 1,260. Across both, 921 Gemini readings and not one citation. |
The second set was built to break the first one, not to confirm it, and it did not break it. Agreement between two question sets designed independently is worth more than either set on its own, and it is the reason the British figures above can be planned against rather than merely noted.
The British layer you will not displace, and why that is useful
gov.uk at 57 readings and ico.org.uk at 28 are not competitors. They are the sources these engines reached for when a British question touched registration, data, consent, complaints or anything a regulator publishes on. A British agency page explaining what the Information Commissioner requires is bidding against the Information Commissioner. That is budget spent for sport, and it is the single most common way a UK test comes back flat: half the frozen questions were never winnable, so the average could not move even when the work was good.
The useful half of that finding is what it tells you to do instead. Sit above the institutional source rather than against it. A page that says which regulator page applies, names it, and then answers the British commercial question the regulator does not answer, is sitting next to the source the engine already trusts rather than trying to replace it. The same run showed that 82.0% of the 20,143 citations recorded on 16 October 2026 went to a vendor or agency own site, so the slots below the institutional layer are held by self published pages and are open to anyone publishing better ones.
How to scope a British test so the result is readable
- Sort your frozen questions into two piles before you start. Questions whose honest answer is a government or regulator page, and questions whose honest answer is a supplier page. Measure both, report them separately, and judge your work on the second pile.
- Put Britain in the question, not in a setting. A country setting can report success and still return an answer about somewhere else, so write the United Kingdom into the sentence and then read the answer for pounds and British context before you count it.
- No brand in the question, including yours. Supply the name and all you learn is whether the engine can repeat it. Ask the way a British buyer asks, with no firm in the sentence, and the answer shows you who gets put forward instead.
- Report the two engines separately. On 16 October the same British questions produced 3 citations of our domain on ChatGPT and 0 on Gemini. A single blended British score would have averaged those into a figure describing neither engine.
- Fix the interval before the first measurement. Publish the window with the result so it is part of the finding rather than a decision made after the numbers arrived.
What goes on the British pages in between
The build spec conversations run alongside the readings are specific about what a British page needs. Every one of the 102 asked for a price in the reader's own currency, which here means a figure in pounds with the VAT position stated and the month it was true, because a monthly figure that leaves VAT ambiguous answers a different question from the one the buyer asked. Reading a supplier's number for what sits inside it is its own skill, and how to read a London rate card for the overhead instead of the output takes one apart line by line.
Keep every claim inside what the page can prove on its face. The CAP Code governs British marketing claims and a complaint about one can be put to the Advertising Standards Authority, which is reason enough to attach the working to any comparison you publish. A dated measurement with its denominator printed is the cheapest form that evidence takes, and it happens to be the same thing these engines kept citing. 83 of the 102 specs asked for a confidence interval, a variance or a repeat count, and 80 asked for the sample size.
What these British readings do not establish
No work was done at these questions between the two runs, so the pair is a reading of a steady British baseline rather than the result of anything. The intervention study, with a dated rate on each side of the change, is in progress. Everything above belongs to ChatGPT and to Gemini and to no other surface: Google AI Overviews, Perplexity and Claude were not asked, so nothing here reads on them, and the five surfaces we cover for customers are never collapsed into these two. The trade level British rate is absent too, because the questions aimed at particular trades were worded generically rather than named.
Questions about scoping a UK AI visibility test
How long does it take for a British AI visibility figure to change?
Set the interval in advance, measure on the same frozen questions, and publish the window beside the result. Our own UK reading is taken again on the same set with a fresh date and denominator printed next to it.
Our citations went from one answer to three. What does that license us to say?
That the count rose by two readings, with both denominators printed beside it. If the denominator also grew between the two measurements, the count and the rate tell different stories and both belong in the report. Give it in readings rather than percentages and nobody has to guess at the sample behind it.
Can the change be attributed to the work we paid for?
Split the frozen set in two and work on only one half. Measure both halves on both dates, and the gap that opens between them is the part belonging to the work. The untouched half earns its keep twice, because it also shows you the weeks an engine moved on its own.
Which interventions moved a British number?
What the two runs establish is which pages the engines used, and in the British set those were small British agency sites and British institutional sources. A frozen British set is the instrument that tells you which of your own pages did the work. No published benchmark, ours included, can substitute for it, because it is not asking your questions.
How many British questions should a frozen set contain?
Enough that the report does not swing on one answer. A single reading out of 210 is worth roughly half a percentage point, so a British set of that size holds still; a set of twenty will not, and sets that small are why so many British reports look dramatic.
Where a British test usually goes wrong, and the fix
The common failure is not a weak intervention. It is a frozen question set that quietly included a dozen questions whose honest answer was always going to be a government page, which flattened the average and made good work look like nothing. Sort the set first, and the same test suddenly reads. We sell this measurement and the publishing that follows it, so take the British figures above with that in mind: they come from our own run, in which our own domain was cited in 3 of 210 British ChatGPT readings and in none of the 210 on Gemini. Everything above carries its fraction and its day, and the identical questions go out again next round.