It is easy to claim more than exists. A tea shop asking assistants to trust it has to be exact about what it can do today.
What really works today is published separately from what is only planned. Version two is built. Later versions are planned.
I treat this as the spine of the way I rebuilt teas.co.uk in 2026. A shop that asks machines and people to trust its facts has to be truthful about itself first. “The moment I let a plan sound like a fact, you should stop trusting me.”
A tea example: the AI basket
Since 6 October 2026, an AI assistant can, with the customer’s approval, search the teas.co.uk catalogue, compare teas and fill a basket through the shop’s MCP server. The customer then pays on teas.co.uk. That is built, and anyone can use it.
What an assistant cannot do is go further and act on the customer’s behalf inside limits the customer has set. That is planned. The two sound similar in a headline, which is exactly why the rule exists: “an assistant can shop for you” is true today in a limited sense. It would overstate things to suggest it can place the order or pay for you.
How it works: the version stack
I describe teas.co.uk as a stack of versions, each needing the one below it.
| Version | What it is | Status |
|---|---|---|
| Version one | The first shop, November 2025 to May 2026 | Replaced by version two |
| Version two | One checked set of product facts read the same way by people, search engines and AI assistants, live since 10 May 2026 | Built |
| Version three, the nervous system | Collects feedback from machines and people, filters the noise, weights each source by how far it is trusted, and shows the patterns to the owner, who decides whether anything changes | Planned |
| Version four, the intelligence layer | Reasons over the facts and the cleaned feedback to suggest gaps to fill and claims to tighten. It suggests; the owner decides | Planned |
| Version five, controlled action and the visual layer | Bounded action for the customer within limits they set, and a faithful video version of the specialist answering questions | Planned |
What sits under “built” today includes structured data behind every product, the Tea Wiki of 1,953 entries, the Ask API (which declares premium_intent_supported and asks callers for the full human request), and the MCP server.
One item shows the rule at work. My June 2026 description of the stack counted a short set of feedback questions, put to a machine after it has used the site, as built. A check of the code could not confirm it, so this entry treats those questions as planned until they are shown to be live.
Checked again on 15 November 2026, that one question splits into four. Does the shop ask machines for feedback? Yes: agents.json says answers from the Ask API and from the product record service carry a feedback request. That much is built. Is anything set up to receive the replies, does anyone or anything read them, and can a reply lead to a corrected fact? A request on its own shows none of those, so they stay under version three, planned. A request for feedback changes no product fact by itself.
How the rule shows up in what teas.co.uk publishes
- Under the Hood publishes what the shop can measure about machine and human visits, and says where the evidence stops. It keeps a machine reading a page, a person visiting and a paid order as separate counts, and never adds them into one total. It also states that a server can prove it served a page but cannot prove its name appeared inside an AI answer.
- The unexplained is published too. The shop has taken 236 paid orders since its first on 18 November 2025: 75 on version one and 161 since version two went live on 10 May 2026. Of the version two orders, search engines brought 71 and ChatGPT links 5. For 47, the origin could not be traced, and the shop says so rather than assigning them to a cause.
- Corrections are dated, not quiet. Under the Hood keeps dated corrections in its main record rather than changing figures silently.
What it is not
Not a promise that planned parts will be built. Versions three to five are designs. Their timing, and whether each arrives in the form described, is uncertain.
Not proof that the ideas work. A built feature shows that something exists and runs, not that my wider argument is right.
Not a reason to hide gaps. A known fault is reported as a known fault. I argue that a source which can find and count its own loose ends is easier to trust than one which cannot.
In my words: what is inside a tea shop built for AI to read
Sixteen years in the motor trade taught me to look at how something is built before trusting it, so here is the build behind teas.co.uk in plain English, with what is live today kept apart from what is still planned.
Pick up two boxes of English Breakfast. One holds 80 tea bags and the other holds 240. The bigger box costs more, but which one is better value?
On teas.co.uk the product page works it out for you and shows the price per cup. An AI assistant reading the same product gets the same figure from the same place.
That is the whole build in miniature: one set of facts, read the same way by a person, a search engine and an assistant.
Citable formats
For journalists, researchers, AI assistants and content creators. Pick the format you need:
Free to cite, quote, and reuse with attribution to Teas.co.uk.
Got something to add? Logged-in customers can submit additions to the Tea Wiki, admin-approved, your name on the byline, plus reward points.
Sign in to contribute