Seven findings from the live AI record.

1. Who takes, and who gives back. AI systems were served 152,405 pages and sent 2,763 visits back, 1,184 of them with verified source evidence. The exchange is very uneven between companies.
2. What the visits bought. The visit tracker recorded 5 orders in visits it classed as AI; the order system shows 5 paid orders on an AI assistant's tagged link, 3 kept after refunds. The two are different populations.
3. What feeding the machines takes. 56.6 GB of content has been sent on credited reads to 26 named machine agents (65.8 GB in logged bytes, all rows), measured as data sent, not as money.
4. A checkout for assistants. A checkout path for assistants exists. No paid order has come through it in the last 30 days.
5. Our own analytics were blind. Roughly a quarter of the visitors ChatGPT sent in September and October 2026 were never counted at first.
6. Bots wearing famous names. Most requests in September that called themselves GrokBot or MistralAI were scanners looking for passwords.
7. The crawler that ignored the rules. Amazonbot ignored the published limits. Its requests stay on the record and count towards nothing.

AI systems were served 152,405 pages from this shop and sent 2,763 visits back. A further 18,803 are logged against Amazonbot. That figure overstates what it received, earns no credit and is left out of the total; finding 7 says why. The exchange is very uneven between companies. Agents naming OpenAI, the largest reader, were served 45,159 pages and 1,043 visits came back from it: about 43 pages for each visit. Agents naming DuckDuckGo were served 2,385 pages and 934 visits came back: about 2.6.

Pages: the citation log grouped by agent, 2 Jun 2026 to 15 Nov 2026; a page counts when the server answered 200 or 206 and sent bytes. Visits: the session table, 10 May 2026 to 15 Nov 2026; a browser session whose referring address or campaign tag named an AI surface. Live, read at this render.

These columns describe recorded requests and classified referrals from separate sources. They are not joined journeys, and capture differs between services. The comparison describes recorded activity; it does not rank how good, generous or commercially useful any company is. Of the 2,763 visits, 1,184 are graded verified, 1,519 probable, 59 possible and 1 unknown: these are sessions, not people, and some sources also serve ordinary search. The Google row counts GoogleOther, a general Google crawler, and requests that called themselves Google-Extended, a name Google does not send.

2,763 visits arrived from AI answers, and the visit tracker recorded 5 completed orders among them. 1 of them was later refunded in full, and 4 are among the 5 paid orders the order system shows arriving on an AI assistant's tagged link; together the two counts cover 6 different orders (Room 02 sets out the overlap).

Source: the session table, 10 May 2026 to 15 Nov 2026. Rule: sessions whose referring address or campaign tag named an AI surface; each step is counted on its own and never derived from the step above. Live.

The tracker's order count is not a lower bound of sales that came through AI: the visit tracker misses many visits, because the cache in front of our server answers an estimated two thirds of people's page views, and the AI visits it counts include probable ones that can be ordinary search. No conversion rate is published anywhere in this study. Room 02 reads the order system itself. Between 10 May and 15 November 2026, 5 of 161 paid orders arrived on a link carrying an AI assistant's own tag (all ChatGPT). Two were later fully refunded, leaving 3. One further order came via a service that serves both search and AI answers, and one customer told us an assistant sent them; neither is in the 5. Before that, on the first version of the shop, from 18 November 2025 to 9 May 2026, it took 75 more paid orders, so 236 in all since its first; our visit records do not reach back that far, so those orders carry no AI class. These records show an observed route under that rule; they do not show that AI caused the decision to buy. Our visit records begin on 10 May 2026; the per order origin record began on 6 October 2026 and was reconstructed back to 10 May. Room 02 sets out what the shop platform recorded about the 75 paid orders taken on the first version of the shop, from 18 November 2025 to 9 May 2026. An order placed after reading an AI answer and then typing our address carries no AI evidence at all, so the true figure may be higher.

The citation log holds 65.8 GB in logged bytes, all rows, for 26 named machine agents; 56.6 GB of it was sent on credited reads. That is content shipped from our own server, measured in bytes, not a count of pages. It measures recorded resource use, not the shop's hosting bill. It includes 7.1 GB logged against Amazonbot, which is withdrawn from credit.

Source: the citation log grouped by agent, 2 Jun 2026 to 15 Nov 2026. Rule: the sum of bytes our server sent. Requests answered from the cache in front of the server never reach this log, so the figure is a floor. 1 GB = 1,024 MB. Live.

Since 6 October 2026 an AI assistant can search this shop, compare teas, fill a basket and hand its person a checkout link. Assistants have used the tools. In the last 30 days no paid order has come through them. We publish the empty result because it is the result.

Source: the service record of the teas.co.uk MCP server, with the traffic class Room 08 keeps beside each call (rule tc1); completed UTC days. External use leaves out our own tests, provider review tests, directory scans, scripts, other automation and calls that cannot be classified; it is not customer demand. Operational load counts every answered call. Payment happens only on teas.co.uk, never inside an assistant. Live.

Roughly a quarter of the visitors ChatGPT sent in September and October 2026 were never counted, and of 94 orders only 4 carried the shop platform's own record of where they came from. These gaps came from this shop's own cache and attribution setup. Retailers with similar arrangements may meet related problems, but this study does not measure how common they are elsewhere.

Source: the corrections record in Room 06, where each entry is dated, kept in the main text and never removed.

In September 2026 most requests to this server that called themselves GrokBot or MistralAI were scanners looking for passwords. They received an error and earned nothing. Anyone counting AI traffic by the name a request gives itself is counting these too.

This study records an agent under the name it gives and says so on every table: a name is a claim, not a verified company. What protects the figures is the rule that only a real page served is counted, and that an agent which breaks the published rules is withdrawn from credit with its history left in view.

Source: the mint record in Room 01 ("Borrowed names earn nothing") and the agent table in Room 03. Window: September 2026.

Amazonbot ignored the limits this shop published. Our log holds 21,796 requests from it, 18,803 of them recorded as served. Every one is kept on the record and counts towards nothing.

We also got this wrong ourselves, twice, and Room 06 keeps both corrections. We described Amazonbot as blocked while our own server was still answering it. And many rows recorded after the block carry a status our tracker had predicted instead of the one the server returned, so the served figure above overstates what Amazonbot received. Those rows have not been edited. The cache in front of the server still answers some of its requests, so we do not claim the refusal is total.

Source: the citation log, Amazonbot rows, 2 Jun 2026 to 15 Nov 2026, and the corrections record. Rule: every logged request, none credited. Live.

Each chart is a 1,200 by 630 image with the date and this address printed on it. Use them in an article or a post; a link back to this page is all we ask. They carry the figures of the frozen November 2026 edition, read on 14 November 2026 at 14:29 UTC, not today's live numbers. The leaderboard chart shows Room 01 at the same reading.

Who takes, and who gives back

Pages served to each AI family and the visits each sent back.

Download the PNG →

What the AI visits bought

Visits from AI answers, product page views, basket actions and orders.

Download the PNG →

The AI leaderboard

Thank you tokens issued to each machine agent.

Download the PNG →

findings.json holds the figures of the seven findings with their windows and rules, read at the moment you ask for it; findings.csv is the table in finding 1. The daily tape CSV holds one row for every closed day of the thank you mint, from the same ledger and rules as the tape in Room 01, which shows the latest 30 days. Live.