We tried to visit 30 of the UK's biggest retailers as an AI agent. 19 wouldn't let us in.

Ask an AI assistant to find you a decent air fryer under £80 and it will go shopping on your behalf: visit retailers, read product pages, compare prices, come back with an answer. That only works if retailers let it in. So in July 2026 we scanned 30 of the UK's biggest retailers with AgentVisible, identifying ourselves honestly as an automated agent, the way well-behaved bots have done since robots.txt was invented.

Two-thirds of them refused the connection outright.

The numbers

Scans used our published methodology; every published score links to a live profile showing exactly what was found, where, and when.

The bouncer problem

None of this is malice. Retail bot defences were built for scalpers, scrapers and card-testers, and they do that job well. The trouble is they cannot tell a stock-sniping bot from a customer's shopping assistant, so they bounce both. A blocked agent does not appear in anyone's analytics as a lost sale. It appears as nothing at all, while the assistant quietly recommends a competitor that answered.

The same scan window shows this is a policy problem, not a technical one: software companies we scanned score anywhere from 50 to a perfect 100 on the identical checks. The web can be made agent-readable today. Most of UK retail has simply chosen locks over legibility.

The bit nobody expects: Lakeland

The best agent-accessible retailer in Britain is not a supermarket, a department store or a fashion giant. It is Lakeland, the Windermere kitchenware company, at 83 out of 100 — and not by accident. Lakeland publishes an llms.txt telling agents where its useful resources live, and serves clean Markdown versions of its pages when an agent asks. That is deliberate, modern plumbing that most of the FTSE-scale names on this list have not touched.

This is what makes the agentic channel interesting for challengers: the giants' bot walls do the losing for them. A mid-size retailer that opens the front door properly is, to an AI assistant, bigger than Tesco.

Reachable is not the same as legible

John Lewis and Ocado deserve credit: they let an identified agent in. But both score 20, because beyond the doorway there is nothing addressed to agents at all — no llms.txt, no Markdown, no machine-readable description of what they sell or how buying works. Reachable is step one. Legible is where the recommendations get won.

What we did not do

We did not disguise the scanner as a browser, rotate residential IPs, or execute JavaScript to fight through challenge pages — the tricks a hostile scraper would use. That is the point of the measurement: it shows what a polite, identified agent experiences. Some retailers may allowlist specific commercial assistants behind the scenes; where that is true, we cannot see it, and their customers' agents that are not on the list cannot either.

The six "challenge page" sites let our scanner connect but served interstitial or near-empty pages, so their profiles are under review rather than published with numbers we do not stand behind.

If you run a retail site

Three moves, roughly a morning of work, none of which weaken your bot defences against actual abuse:

1. Decide your agent policy on purpose, rather than inheriting whatever your WAF vendor's default is. State it in robots.txt.
2. Publish llms.txt and serve Markdown for your key pages — our seven signals checklist walks through it.
3. Scan yourself and see what an agent actually experiences.

Run a free scan and see where your site stands. No sign-up, just a URL.

Fair game, open ledger

Every published number is reproducible from a stored snapshot with a visible timestamp and evidence URL. Scores cannot be bought; no retailer here has any commercial relationship with us. Corrections and rescans are free — the route is on every profile page, and improvements get credited in the next monthly cycle. AgentVisible is built by Pixelshed.