UK AI Visibility Report, Q3 2026
First edition. What 6,901 cited sources across four AI engines show about how UK local businesses are found — and what our own corpus cannot yet tell you.
This is the first edition of a report we intend to publish quarterly. Everything in it comes from our own corpus of AI answers about UK local-service businesses. Every figure states its date and the command that produced it, and where a number would be misleading we have left it out and said why.
Read this first: the corpus is small
59 audits. 629 recorded answers. 193 questions. Three monitored businesses. Measured 31 August 2026.
That is not a market study, and we are not going to present it as one. The businesses are concentrated in vehicle and roadside services, in London and Leeds. Anything below describes this corpus, not the UK economy.
We are publishing it anyway for two reasons. The engine-level behaviours are structural rather than sample-dependent — how an engine formats a citation does not change because we asked it about a different trade. And a report that states its limits is more useful than a confident one that hides them.
1. Perplexity cites nearly three times more than ChatGPT
Source entries recorded per engine across the corpus:
| Engine | Source entries | Distinct hosts cited |
|---|---|---|
| Perplexity | 2,852 | 309 |
| Gemini | 1,770 | 197 |
| Claude | 1,265 | 118 |
| ChatGPT | 1,014 | 147 |
6,901 source entries in total. Perplexity cites the most sources and draws from the widest set of hosts by a wide margin. If you are choosing where to start, its citation list is both the largest opportunity surface and the easiest to read.
2. Gemini hides every publisher behind a redirect
1,770 of 1,770 Gemini source entries are Vertex AI grounding redirects. Not most. All of them. The same measurement on 21 August returned 1,625 of 1,625.
A grounding redirect is a Google-hosted URL standing in front of the real destination. Any tool that reads the publisher's hostname straight out of the citation URL therefore sees zero Gemini citations — not fewer, zero — and reports the other engines as though they were the whole picture. Gemini is roughly a quarter of our corpus.
By contrast, ChatGPT exposes 890 of 1,014 sources directly (123 more are search infrastructure rather than publishers), and Claude and Perplexity expose 100% directly.
We found this because a number that should have been non-zero was exactly zero. If a supplier reports Gemini figures to you, that is the question worth asking: how do they resolve the redirect?
Source: signalfound:citation-country-coverage, production, 31 August 2026.
3. Between a third and a fifth of cited sources are not provably UK
For every distinct host cited, we establish whether it serves the UK — by domain where the domain settles it, and by checking the site itself where it does not.
| Engine | Provably UK | Provably foreign | Cannot tell |
|---|---|---|---|
| Claude | 83.1% | 2.5% | 14.4% |
| ChatGPT | 76.9% | 7.5% | 15.6% |
| Gemini | 72.1% | 6.1% | 21.8% |
| Perplexity | 67.3% | 13.6% | 19.1% |
Perplexity cites the most foreign sources of the four, by both measures — it is the only engine in our corpus citing hosts that are foreign by domain alone, and it carries the most that are foreign on inspection.
142 hosts across the four engines cannot be resolved either way. They have a generic domain and a page that does not settle the question. We treat those as unproven rather than assumed-local, which means they are withheld from outreach lists and counted on screen. That is a deliberate cost: it makes our target lists shorter than they could be.
4. Mention rates differ by a factor of nearly three
Across 629 recorded answers, the proportion in which the subject business was named at all:
| Engine | Mentioned | Cited |
|---|---|---|
| Gemini | 52.3% | 42.6% |
| Perplexity | 34.4% | 34.4% |
| Claude | 28.0% | 24.7% |
| ChatGPT | 19.4% | 13.5% |
Two things stand out. ChatGPT is the hardest of the four to be named in, and it is the one most of your customers use. And Perplexity's mention and citation rates are identical, which fits an engine that cites nearly everything it draws on.
⚠️ These rates are corpus-wide and dominated by a small number of businesses. Read them as engine behaviour, not as a benchmark for yours.
5. The finding we did not expect: the same business, same questions, moved 33 points in six days
One monitored business — a roadside-assistance firm in London — was checked 28 times over six days in August 2026, on the same three questions.
- Lowest score: 17. Highest: 50. Mean: 34.9.
- First reading 32, last reading 42.
- Nothing was done to the business in between. No new listings, no site changes, no campaign.
The full range is nearly three-quarters of the mean. A supplier who checked on the right day could show a 47% improvement, and one who checked on the wrong day could show a 47% decline, from the same business doing nothing.
We think this is the single most useful thing in this edition, and it argues against our own product's most saleable framing:
- A single AI visibility check is a snapshot, and snapshots of a volatile signal are close to worthless for judging performance. Ours included. A free check tells you what an engine said once. It cannot tell you whether you are improving.
- Before-and-after claims across a handful of readings are not evidence. If an agency shows you a score that rose after their work, ask how many readings sit behind each end of that comparison. Two points on a signal with this much variance mean nothing.
- The trend needs enough readings to see through the noise, which is exactly why monitoring exists and why we will not sell a one-off check as proof of anything.
⚠️ This is one business, one vertical, one six-day window. It is enough to establish that the variance is large; it is not enough to quantify it generally. We will report the same measurement next quarter with more subjects.
What this report cannot tell you
Stated plainly, because the gaps matter as much as the figures:
- Nothing about market share. No engine publishes how many people asked anything, so no denominator exists. Anyone quoting you an AI market share has modelled it.
- Nothing about recommendation as a distinct outcome. We record whether a business was mentioned, whether it was cited, and every source the answer drew on. We do not currently score "was this business recommended" — extracting it reliably from free text is harder than it looks, and we would rather omit a number than publish one we cannot defend.
- Nothing generalisable about verticals. The corpus is concentrated in vehicle and roadside services.
- Nothing about Google AI Overviews. We sample ChatGPT, Claude, Gemini and Perplexity. Overviews are not among them.
Method
Questions are commercial-intent sentences generated from each business's service and location, qualified to the UK. Each is asked of four engines through DataForSEO; the full answer text and every cited source are stored. Publisher countries are resolved per host and cached globally. Corpus figures come from signalfound:citation-country-coverage and direct read-only queries against production, run 31 August 2026.
No customer is named. Publisher hosts are public web pages and are counted, not attributed to any business.
Next edition: Q4 2026, with the volatility measurement repeated across more subjects.