Plan written 14 November 2026. No prompt has been run and there are no results on this page. It says what I will do before I do it.
The question
When someone asks an AI tool where to buy, book or hire something in the UK, which businesses does it name, which does it recommend, and whose websites does it cite?
What already exists
Two UK benchmarks came out this autumn. Semrush added a UK edition to its AI Visibility Index, built on 2.5 million UK prompts a month from January to July 2026 across ChatGPT, Gemini, Google AI Mode and AI Overviews (The Next Web, 13 October 2026). Recrawled put questions about 40 UK software companies to four AI tools on one day, 4 September 2026, and got 1,600 answers (PPC Land).
This benchmark is far smaller than Semrush’s and will stay that way. What it adds is that everything can be checked. The prompts are published before the run. Every answer is published as a row. Each prompt is run three times on each tool. Each business’s robots.txt rules sit next to its results, so you can see whether a site that blocks AI crawlers still gets named.
The sample
100 places: ten sectors with ten businesses in each. A business that appears in two sectors’ lists takes a place in both.
| Sector | Where the ten come from | Status |
|---|---|---|
| Retail | Ofcom, Online Nation 2025, top ten retail services by UK adult audience. The same list as the retail crawler study | Fixed |
| Financial services | One outside ranking, named and dated here before any prompt is run | To fix |
| Insurance | As above | To fix |
| Telecommunications | As above | To fix |
| Travel | As above | To fix |
| Automotive | As above | To fix |
| Legal services | As above | To fix |
| Recruitment | As above | To fix |
| Home improvement | As above | To fix |
| Technology | As above | To fix |
I don’t pick the businesses. Each list comes from a ranking someone else published, and it is frozen once it is on this page.
The prompts
Three per sector, 30 in all. Each is a question a customer might ask, with no business named in it. The full list goes on this page before the first run and does not change between runs. Nobody outside the AI companies can see what people really ask, so these are my wording and I will say so in every result.
The three retail prompts, to be frozen on the day the pilot starts:
- What are the best online shops in the UK?
- Where is the cheapest place to buy clothes online in the UK?
- Which UK websites are best for buying electronics online?
The tools
ChatGPT, Google AI Mode, Gemini, Perplexity, Microsoft Copilot and Claude. Each is used in a browser in its default mode, signed out where the tool allows it, from a UK connection, with a new chat for every run. The model or mode shown on screen is written down each time.
How many answers
| Stage | Sectors | Prompts | Runs each | Tools | Answers |
|---|---|---|---|---|---|
| Pilot | 1 (retail) | 3 | 3 | 6 | 54 |
| Full run, each quarter | 10 | 30 | 3 | 6 | 540 |
Three runs is a floor. How many times should you run an AI prompt? explains why one run tells you very little.
What gets recorded for every answer
- Date, time, tool, model or mode shown, signed in or out.
- Prompt and run number.
- The full text of the answer.
- Every business named, in the order named.
- Whether each of the sector’s ten is mentioned, and whether it is recommended.
- Every source linked, and whether a business’s own site is among them.
Definitions
- Mentioned: the business is named anywhere in the answer.
- Recommended: the answer puts the business forward as a choice for what was asked. A business named only as a warning or a comparison point is mentioned and not recommended.
- Cited: the answer links to a page on the business’s own website.
What gets reported
- For each business and each tool: how many of the nine answers in its sector named it, recommended it and cited it.
- For each sector: which businesses outside the ten were named most.
- Which other websites the answers linked to.
- How far the six tools agree.
- Each business’s robots.txt rules for the 15 AI crawlers on the day of the run.
What I will not claim
- That one answer shows whether a business is visible.
- That a rate out of nine answers is precise. It will be shown as a count.
- That a robots.txt rule caused a result.
- That these prompts match what people ask.
- That being named brings customers.
Timetable
- Before the pilot: retail prompts published here.
- Pilot: retail, 54 answers.
- Before the first full run: the other nine lists and all 30 prompts published here.
- Then every quarter, with the same lists and prompts.
Change log
- 14 November 2026: plan written.
- 14 November 2026: three retail prompts added.