How AI
shops for Britain

We ask ChatGPT, Claude and Gemini the questions real UK consumers ask — every week, across 40 categories — and record exactly which brands they recommend and which sources they trust. This is what the answers say.

0 AI answers · 0 brands tracked · measured, never estimated

Legal Services is a monarchy. Accounting Software is a bazaar.

The share of brand mentions captured by each category's top three. In legal services, three brands take 92% of everything AI says; in accounting software, 22 brands split the answers and the top three manage just 30%.

Every category has a king — and 23 thrones changed hands.

The most-named brand in each category's answers. In 23 of 40 categories the AI answer leader is not the market leader we seeded — Citizens Advice outranks Co-op Legal Services in legal services.

Britain has a citation establishment.

reddit.com and youtube.com are cited in 34 of 40 categories — from credit cards to skincare. A small set of trusted referees decides what AI believes, whatever the topic. Each green cell is one category where the domain is cited.

AI reads your website. It cites it 23% of the time.

Across every category, Claude and ChatGPT opened 3,406 brand-owned pages while answering — and put just 795 of them in their cited sources. Solar & Heat Pumps brands convert best (56%); otc medicines worst (0%).

In 8 categories, every model names a different winner.

Consensus is the exception: the models agree on the leading brand in only 12 of 40 categories, and disagree most in skincare. A brand that owns one assistant can be invisible in another — which makes "our AI visibility" a question with three different answers.

There is no single formula. Every category rewards something different.

Percentage-point difference in how often a page gets cited, with a feature versus without, within the pages models actually read. Structured data travels furthest — worth +13pp on average — but read down the columns and the picture changes category by category.

Some categories run on guides. Others run on the crowd.

What AI actually cites in each category, by kind of page. The mix is the brief: it tells you what to publish to be quotable here — and it is not the same twice.

AI wants this month's answer in energy suppliers, and last year's in food delivery.

Median age of the content AI cites, by category — from each page's last-modified date. Where the bar is short, stale content is invisible no matter how good it is; where it is long, evergreen pages keep earning.

Age is measured from each page's last-modified date, which about half of pages publish; categories without enough dated pages are omitted rather than guessed.

Skincare brands build for AI. Hotels brands don't — yet.

How ready each category's brand-owned pages are for AI search, scored on eight structural signals models reward. A low score is not a verdict on the brands; it is the size of the opportunity sitting in front of whoever moves first.

Three AIs, three personalities.

Ask the same 800 questions and the judges behave differently: Gemini names 2.5 brands per answer; ChatGPT keeps counsel with the fewest sources. Winning one model's answer does not win the others'.

Watch the crowns move.

The programme re-runs every week and audits every cited page monthly — the boards below are live, and this report re-writes itself as the data changes.

Method: each category runs a fixed set of 20 unbranded consumer prompts (spread across the buying journey, written from persona and need-state pairings) against ChatGPT, Claude and Gemini weekly. Visibility is the share of answers naming a brand; the read→cited figures use Claude and ChatGPT only, because Gemini does not separate pages it read from pages it cited. Findings are correlational observations of model behaviour at the time of each run. Data updated Tue, 01 Sep 2026 17:27:14 GMT.