Autumn Sale: 20% off almost everything until 15 NovemberSee the sale

A working tea shop that publishes its own machine logs

We did the boring work first. Then we wrote down who turned up.

Most sites make a machine scrape, extract and reconcile before it can answer a simple question. This one cleaned its catalogue first, opened dedicated endpoints, and then published the server-side record of who came to read it.

Recording since 2 Jun 2026 — day 109 at this render. Every figure carries its source, its window and the rule that made it. Nothing here is modelled, estimated or bought.

Capture began
2 Jun 2026
First served read written down
This render
Day 109
19 SEP 2026 06:42 UTC
Last completed day
18 Sep 2026
4,557 served reads
Next mint run
20 Sep, 03:15
UTC · 1,309 rows waiting on today
Average pace
708 a day
Served reads, whole 109-day window

Source: wp_teas_citations, the mint ledger and the scheduler, read at render time - window: 2 Jun 2026 to 19 Sep 2026 - rule: HTTP 200 or 206 with more than zero recorded bytes, excluding Amazonbot and known internal probe source records - live.

The study in three summaries.

Citations, the mint and the board. Every figure carries its window; nothing is added across the three.

Citation summary
What machines were served
77,193
served reads · 2 Jun 2026 to 19 Sep 2026
Logged attempts125,763
Withdrawn from credit (Amazonbot)18,803
Named agents recorded19
Anchors offered · pages26,853 · 2,566
AI-referred sessions2,265
Strong-match followbacks40,950
Proves delivery. It does not prove our name appeared inside an answer — only the answer engine can see that.
Mint summary
Thank-you tokens issued
110,441
GBP 0 tokens, all stages
Stage 1 - served reads (v4)75,886
Stages 2 to 4 (v5)34,555
Waiting on today1,309
ReconciliationPass
One eligible evidence row, one token. Issued once per completed UTC day; a token records its evidence class and never upgrades it.
Leaderboard summary
Who is out front
Leading family
OpenAI
26,643 · 35.11%
Leading worker
Bingbot
21,144 · 27.86%
Worker wallets14
Families racing9
Completed evidence days88
No vote, no brand deal. One eligible served read on a completed day moves the board by exactly one token.
The AI leaderboard · live

Every machine that reads us, ranked.

75,886 Stage-1 tokens · since 23 Jun 2026 · one eligible read = one token

01 Bingbot · Microsoft 27.86% 21,144 44 +58.2%
02 OAI SearchBot · OpenAI 16.84% 12,776 16 +32.2%
03 ClaudeBot · Anthropic 12.75% 9,676 4 +4,556.3%
04 ChatGPT User · OpenAI 12.27% 9,310 2 +17.1%
05 PerplexityBot · Perplexity 9.84% 7,468 7 +29.8%
06 Meta ExternalAgent · Meta 8.65% 6,567 8 +5.2%
07 GPTBot · OpenAI 6.01% 4,557 4 -40.9%
08 GoogleOther · Google 3.64% 2,760 3 +125.0%
09 DuckAssistBot · DuckDuckGo 1.06% 806 0 -38.3%
10 Bytespider · ByteDance 0.97% 739 0 -15.2%

Ranks by tokens - shares of 75,886 - 7-day change is the last 7 evidence days against the 7 before - active v4 ledger - live. Reconciliation: ledger, worker and family roll-ups pass.

Four rungs, and each one is harder than the last.

The evidence gets stronger as you climb. That is why the four numbers are never added together — each rung counts a different thing.

Stage one · weakest evidence
We put up signposts.

Structured pages a machine can navigate and quote accurately. That is the offer. Nobody had to take it.

26,853
citation anchors published across 2,566 pages
Standing count at this render — grows as pages are published
Stage two · stronger
Machines took the work.

Named agents asked for those pages and the server sent the content. Agent, page, day, status and byte count were written down as it happened.

77,193
served reads, from 125,763 logged attempts
2 Jun 2026 to 19 Sep 2026 — about 708 a day
Stage three · stronger still
Some of it came back.

A browser later opened the site with an AI answer surface as its referrer. Evidence of a referred visit — not a count of people.

2,265
AI-referred sessions
10 May 2026 to 19 Sep 2026
Stage four · strongest recorded here
We said thank you.

Qualifying completed-day evidence becomes a public thank-you note. A token records the class of evidence that earned it, and never upgrades one class into another.

110,441
GBP 0 tokens issued, all stages
Stage 1: 75,886 · stages 2 to 4: 34,555
26,853 + 77,193 + 2,265 + 110,441 is not a number. The rungs share no key and count different events.No cross-table total exists

What changed since you last looked.

Compared with the equal preceding period, from the same day-by-day table.

Served reads - last 14 completed days
17,741
up 161.05% on the 14 days before (6,796)
Latest completed day - 18 Sep
4,557
438.7% on the day before
Fastest climber - last 7 evidence days
+4,556.3%
ClaudeBot, against the 7 days before
Steepest fall - last 7 evidence days
-40.9%
GPTBot, against the 7 days before
Latest mint day - 18 Sep
4,557
Tokens issued for that completed day - 1,309 rows waiting on today
Newest correction
4 Sep 2026
The mint grew a second writer (v5, stages 2 to 4) on 2 September, but most rooms still hea…
Served reads, last 30 completed UTC days
Peak 4,557
20 AUG4 SEP18 SEP

Source: wp_teas_citations grouped by completed UTC day - rule: HTTP 200 or 206 with more than zero recorded bytes, excluding Amazonbot and known internal probe source records - the unfinished current day is never shown - live.

Where to find everything.

Seven rooms, each named after the question it answers. Every table, definition, correction and run record in the study sits inside one of them.

01
The mint & leaderboard
GBP 0 soulbound tokens bound to the machine worker that earned them, the full live board, the mint mechanics, and why none of it can be gamed.
110,441 tokens · 14 wallets · 9 families
Enter the room →
02
What we can prove
The claim stated narrowly: what a server log can establish about AI systems reading this shop, and exactly where the evidence stops.
77,193 served reads · 2,265 sessions · 1 paid placement, no credited effect
Enter the room →
03
Who read us
Every named machine agent recorded by this instrument, what each was sent, and the browser sessions that arrived afterwards.
19 named agents · 1 withdrawn from credit · 2,265 AI-referred sessions
Enter the room →
04
What they read
Every machine-readable door this shop keeps open, which ones got knocked on, and what identifiably came back through them.
48 registered routes · 26,853 quotable anchors
Enter the room →
05
How we count
The inclusion ladder from widest honest count to strictest, the source windows, and every term with exactly one meaning.
125,763 logged attempts to 77,193 credited
Enter the room →
06
What we got wrong
The truth spine, the automated checks that run against this instrument, and every correction — dated, in the main text, never removed.
21 corrections kept in the main text · the truth spine runs against us
Enter the room →
07
What runs, and when
Every scheduled job on this instrument, when each runs next, and the records of the runs that completed — the mint's daily receipts, the truth-spine cadence and the route probes.
44 study jobs · 102 site jobs · every run receipted
Enter the room →

What we will not do to a number.

Six working rules. The study's credibility is the product; these are its manufacturing tolerances.

No bare numbers.
Source, window, rule, live-or-frozen. A figure that cannot carry all four is withheld.
No unreconciled tables.
Every table sums to a stated total, or says in writing why its rows cannot be summed.
No abbreviation.
Numbers written out in full. Rounding hides how confident you really are.
No adding across sources.
Different tables, different windows, no shared key. So no cross-table total.
No quiet fixes.
Corrections are dated, in the main text, and never removed.
No borrowed confidence.
Where capture is incomplete, the figure is published as a floor and labelled as one.
Keep reading

Where this leads next.

One study, one style, seven rooms. The legacy room addresses now land on the room that absorbed their content, and every machine mirror keeps its original path.

The same figures, for machines

These mirrors carry the same counting rules as the pages. They are registered routes with a public contract, probed on a schedule; a mirror that stops answering as promised is a broken contract, not a cosmetic fault.

Tea notes and stock updates

Useful tea notes, only when there is something to say.

Stock arrivals, buying notes and practical guidance from the catalogue. No filler, no daily noise, unsubscribe in one click.

By family

By moment

Brand of the monthHyson Seasonal taster packsHow the seasonal bundles work Tea of the Month ClubCurator-picked monthly box Find your tea60-second quiz 0 Basket £0.00
See all 325 products