Attribution is the work of connecting an event to a source. For a shop, that might mean identifying the link a visitor followed or the route associated with an order. It helps explain what the shop can observe, while keeping uncertainty visible.
The shop wants to know whether AI really sells tea, so a page read, a visit and an order are counted separately.
A page read is not a visit, and a visit is not a sale. Page reads, visits and orders are counted separately and published, so nobody has to take the shop’s word for it.
An AI system reading a page, a person visiting from an AI answer and a customer buying are three different events.
So did AI bring that tea order? Sometimes teas.co.uk can answer yes, with evidence. Sometimes the honest answer is “possibly”, and often it is “we cannot tell”. I built the shop so that AI assistants can read it, quote it and send people to buy tea from it, because more and more people now ask an assistant which tea to buy instead of searching for it themselves. That makes it worth knowing which part of the chain is working.
Start with the event
A fetch is a request for information. A server log can show that a response was served. It cannot, on that evidence alone, show what the assistant said to a user.
A referral is a visit associated with a source link. A session is a group of activity counted under the site’s stated rules. A link carrying a source tag can help identify the route; a missing tag does not prove that AI played no part.
An order is a separate commerce event. Counts should say whether they mean placed orders, paid orders or orders after refunds, and which dates they cover.
A single pack shows the difference. Take Yorkshire Tea Original Everyday Blend, 80 tea bags. An AI system may ask for its product page at three in the morning: that is a fetch, and the server log records it. A week later a person may land on the same page from a link in an assistant’s answer: that is a referral. If that person then pays for two packs, that is an order. Three records, kept in three places, answering three different questions: was the page read, did someone come, and did someone buy?
What does conversion mean?
A conversion rate compares a specified outcome with a specified population, such as paid orders divided by eligible sessions. Change the population or time window and you change the meaning of the rate. A result from another retailer or country is context, not a result measured at teas.co.uk.
The crawl-to-refer ratio
This compares page requests with visits attributed to the service making them. A high ratio can describe an uneven exchange, but only if both sides have clear definitions and compatible windows. It is not a count of verified recommendations or an estimate of sales by itself.
teas.co.uk has measured one such ratio. Between 2 June and 15 November 2026, the systems of the AI company that sent the shop the most visitors read about 44 pages for every visitor its assistant sent back. That is an overall ratio for the period, not a record that 44 pages were read for each customer. What it shows is how much reading sits behind each visible visit, and why a tea page has to be right long before anyone clicks on it.
Why not put every number into one total?
Because the same journey can produce several events, and some journeys cannot be connected at all. Adding reads, visits and orders would hide the differences and could count the same activity repeatedly.
teas.co.uk explains its evidence and limits in Under the Hood. When discussing a figure, keep its source, period and definition with it. A dated historical snapshot should remain dated; a live total should be labelled and generated as live.
That discipline belongs to the wider truth entity idea. Clear records should help the shop admit what it cannot trace, as well as explain what it can.
The four steps from a tea page to a tea order
I sum up the whole approach in four steps:
- A website that is easy for AI to read gets read by AI.
- A website that is read by AI gets quoted by AI.
- A website that is quoted by AI gets visited by people.
- People are the ones who can buy.
If the first step fails, the others cannot happen by that route. When an assistant answers a question about tea, it reads on that person’s behalf. I describe it as a scout. The person stays where they are while the assistant goes out, reads shops, guides and data, checks prices and pack sizes, checks whether a tea really is decaf, compares what it finds and reports back in its own words. Only then does the person decide whether to look for themselves or to let the assistant fill a basket for them.
A scout cannot squint at a photo of the back of a box to work out the caffeine. It needs the facts laid out, labelled and sourced. That is why the shop opens its pages to these readers, and why each step is counted on its own.
| Step | What happens | What the shop records | A tea example |
|---|---|---|---|
| 1. Read | An AI system asks the shop for a page | The named system, the page and the time | A guide to decaf breakfast tea is fetched |
| 2. Quoted | The assistant uses the shop’s facts in its answer | Nothing directly: the answer happens inside the assistant’s own app | The assistant says a pack is genuinely decaffeinated and gives its cost per cup |
| 3. Visited | A person follows a link from the answer | A visit from an AI answer, when the source survives | The person lands on the decaf product page |
| 4. Bought | The person pays | A paid order, with a short record of where it came from | A pack of Yorkshire Tea Decaf English Breakfast, 80 tea bags goes out in a parcel |
The customer is always a person, even when the assistant does more of the work. Today an assistant can bring someone a recommendation or a prepared basket; the next stage is an assistant making an authorised purchase for them.
What the shop can see, and what it cannot
On version one of the shop, which traded from 18 November 2025 to 9 May 2026, I could not see a single AI system. Ordinary analytics show people in browsers; they show nothing about an AI reading a shop overnight. When the shop was rebuilt as version two, the measuring instruments were built in with it. I call the method forensic analytics: rebuilding what happened from the evidence left behind.
The shop now keeps a record of everything that asks it for a page. On 15 November 2026 it ran to 28 million rows across 52 tables. About 3.6 million of those rows were “knocks”: one request at one of the shop’s entry points, saying which entry point, when, and who was asking.
| Event | What teas.co.uk can see | What it cannot see |
|---|---|---|
| A page read by AI | Which named system asked, for which page, and when | What the assistant then said, or to whom |
| A quote in an answer | Nothing, unless a person clicks through | Whether the shop was named, or which sentence was used |
| A visit | The source link when it survives | Whether the person arrived after an AI answer if the tag is missing |
| An order | Where the paid order can be traced | The full path when no source tag was present |
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