Room 03 of 08 · agents and arrivals

Who read us.

Every named machine agent this instrument has recorded, what each was actually sent, and the browser sessions that arrived afterwards from AI answer surfaces.

An agent name is a user-agent claim, not a verified organisation. What protects the figures is behaviour: an agent that breaks the published rules is withdrawn from credit with its history kept visible.

25 NAMED AGENTS1 WITHDRAWN FROM CREDIT2,679 AI-REFERRED SESSIONS61.6 GB SENT TO MACHINES

Section 01

Every named agent, and what each was sent.

One row per agent, every column from the same table and the same window, so the row sums are meaningful. Byte counts are the content actually shipped, not page counts. Amazonbot's row is kept in full and credited nothing.

Agent	Attempts	Served	No content	Days seen	Anchors	Data sent (MB)	Last seen
bingbot	45,317	38,718	6,599	139	6,534	15,656.7	5 Nov
perplexitybot	31,794	29,326	2,468	138	4,285	10,139.7	5 Nov
amazonbot withdrawn from credit - history kept in full, nothing credited	21,796	18,803	2,993	45	4,303	7,271.5	3 Aug
oai-searchbot	21,567	21,111	456	140	5,320	6,082.0	5 Nov
claudebot	19,596	13,839	5,757	140	5,120	10,301.8	5 Nov
gptbot	14,361	8,381	5,980	141	5,572	2,539.0	5 Nov
meta-externalagent	14,278	9,448	4,830	136	6,483	3,926.7	5 Nov
chatgpt-user	12,835	12,008	827	145	2,276	4,562.8	5 Nov
googleother	7,589	3,601	3,988	102	3,118	1,377.6	31 Oct
duckassistbot	2,113	2,064	49	135	1,138	737.1	5 Nov
bytespider	1,578	1,480	98	116	210	450.2	5 Nov
ccbot	102	79	23	18	96	19.9	4 Nov

The 25 agents together: 193,061 attempts, 158,903 served, 61.6 GB sent.Served here counts the raw rule (status + bytes); the credited study figure of 140,100 additionally removes the withdrawn agent.

All 25 named agentsthe long tail, same columns, same window

Agent	Attempts	Served	No content	Days seen	Anchors	Data sent (MB)	Last seen
claude-user	27	27	0	17	25	9.3	5 Nov
google-extended	23	4	19	13	17	5.3	9 Oct
cohere	19	2	17	13	15	5.5	2 Nov
youbot	19	4	15	15	15	4.2	30 Oct
perplexity-user	13	1	12	10	12	1.4	4 Nov
grokbot	10	2	8	7	7	0.0	26 Oct
meta-externalfetcher	6	1	5	4	6	0.0	4 Nov
mistral	4	1	3	4	4	2.3	24 Sep
anthropic-ai	3	1	2	2	3	2.2	1 Aug
claude-web	3	0	3	2	3	1.0	1 Aug
mistralai-user	3	2	1	2	2	0.7	25 Oct
xai-grok	3	0	3	1	3	0.0	9 Oct
generic	2	0	2	1	2	0.5	5 Oct

Source: wp_teas_citations grouped by agent - window: 2 Jun 2026 to 5 Nov 2026 - bytes are SUM(bytes_served) - live.

The coverage cross-check - this instrument against the raw access loguntil 7 Aug 2026 capture was gated to AI surfaces, so the raw log sees far more; both are published so the shortfall can be measured

Agent	Recorded by this instrument	Recorded in the raw access log	Coverage
meta-externalagent	14,278	93,113	15.3%
gptbot	14,361	91,267	15.7%
amazonbot	21,796	27,856	78.2%
claudebot	19,596	10,164	192.8%
chatgpt-user	12,835	8,921	143.9%
oai-searchbot	21,567	8,848	243.8%
googleother	7,589	4,338	174.9%
bytespider	1,578	3,016	52.3%
perplexitybot	31,794	2,270	1,400.6%
google-extended	23	22	104.5%

Across the frozen log: 3,432,235 requests in total, of which 250,590 were named AI crawlers, 146,034 Googlebot and 21,801 bingbot.Instrument counts are live; the log column is a frozen 7 Aug 2026 reference. A ratio above 100% means capture has outgrown that frozen day.

Section 02

The sessions that actually arrived.

From an AI answer to this shop

AI-referred sessions2,679

Pages viewed in them3,067

Product views398

Baskets started29

Checkouts started27

Completed orders4

Each step is its own evidence class, never derived from the step above. The order count is published because it is the truth.

Strong-match followbacks

64,958

lifetime · 3,502 in the last 7 days

A later request that returned to the exact anchor it was offered, at the strictest match confidence. The strongest evidence our own logs can hold that a read had anywhere to land — still not proof our name appeared in an answer.

Candidate rows examined271,627

Matched to an offer66,206

Matched probably62

Source: wp_teas_sessions (10 May 2026 to 5 Nov 2026) and the followback table (17 Jun 2026 to 5 Nov 2026) - two tables, no shared key, never summed - live.

Section 03

Where the arrivals came from.

Referring surface	Sessions	Pages viewed	Product pages
chatgpt	1,006	1,264	210
ddg-ai	934	983	93
bing-ai	529	574	58
brave-ai	68	68	11
copilot	54	58	14
gemini	44	73	8
perplexity	26	26	4
kagi	11	11	0
claude	5	8	0
qwen	1	1	0
grok	1	1	0

AI-referred sessions only; the surface label is the recorded referrer class.

How strongly each arrival is evidencedverified / probable / possible - the grade travels with the session

Evidence grade	Sessions	Viewed a product	Started a basket	Orders
Verified OpenAI 961 · Microsoft 54 · Google 44 · Perplexity 26	1,090	139	4	0
Probable DuckDuckGo 896 · Microsoft 512 · Brave 65 · Kagi 10	1,483	144	3	1
Possible DuckDuckGo 38 · Microsoft 17 · Brave 3 · Kagi 1	59	3	0	0
Unknown OpenAI 45 · Alibaba Qwen 1 · xAI 1	47	19	4	3

Verified: the surface only exists as an AI assistant, or the landing URL carried the assistant's own stamp. Probable: a dual-purpose host, arriving on a citation anchor. Possible: such a host without an anchor. The v5 mint consumes these grades - verified at full weight, probable reduced, possible withheld.

What buyers answered - the customer-declared sourcea separate evidence class, since 2026-09-02, never merged into the referrer counts

What the buyer said brought them here	Orders
google	5
duckduckgo	3
bing	1
chatgpt	1
gemini	1
repeat	1
social	1

These rows sum to 13 answered orders.

Why most recorded sessions are never called human: the session table classifies every row it has ever written as human, so that label cannot be earned from it. Of 1,670,760 recorded sessions, the overwhelming majority are desktop, one-page visits from scraping floods — recorded across 220 country codes and published as bot pressure, not readership. That is why this room ranks nothing by raw session count and treats only the AI-referred subset as separately evidenced.

Section 04

What was sent, and when.

The same source table across rolling and calendar windows. They overlap, so they must never be added. API reads are a subset of machine reads, not a count of completed tool operations.

Machine traffic: month, fortnight and last 24 hours

Cutoff: 2026-11-05 22:48:00 UTC, exclusive. Status: available. Last completed projection: 2026-11-05 22:50:33 UTC. These overlapping periods must not be added.

One source snapshot; changes compare equal preceding periods
Period (UTC)	Machine reads	Reads/day	Change	API reads	AI-referred sessions	Verified-source tier
Rolling 30 days 2026-10-06 22:48:00 to 2026-11-05 22:48:00	31,878	1,062.60	-33.48%	2,938	202	128
Rolling 14 days 2026-10-22 22:48:00 to 2026-11-05 22:48:00	13,523	965.93	-17.57%	1,256	95	52
Rolling 24 hours 2026-11-04 22:48:00 to 2026-11-05 22:48:00	1,091	1,091.00	-3.28%	86	10	7
Calendar month to date (UTC) 2026-11-01 00:00:00 to 2026-11-05 22:48:00	5,414	1,093.74	Not comparable	405	36	23
Previous calendar month (UTC) 2026-10-01 00:00:00 to 2026-11-01 00:00:00	34,385	1,109.19	Not comparable	3,205	226	156

Reads are retained origin records with status 200/206 and positive bytes, excluding Amazonbot and internal probe records. They do not prove that an answer engine consumed or cited the content. A verified-source referral is a source-evidence tier, not a verified person.

The preceding 30-day and previous-calendar-month periods cross an older status-recording correction, so their raw totals do not support a like-for-like percentage change. Historical session labels are preserved; recent classifier and scrape-gate changes limit comparisons of human traffic.

Full comparison and counting rules · Current snapshot as JSON · Markdown · URSY sponsorship evidence

Recorded machine reads across the exact rolling fortnight. The first and last UTC dates can be partial.

Daily values, source coverage and limitations

Clipped UTC daily buckets — the sum equals rolling 14-day reads
Date	Reads	Bounds
2026-10-22 (partial)	24	2026-10-22 22:48:00 to 2026-10-23 00:00:00
2026-10-23	342	2026-10-23 00:00:00 to 2026-10-24 00:00:00
2026-10-24	647	2026-10-24 00:00:00 to 2026-10-25 00:00:00
2026-10-25	574	2026-10-25 00:00:00 to 2026-10-26 00:00:00
2026-10-26	1,074	2026-10-26 00:00:00 to 2026-10-27 00:00:00
2026-10-27	1,638	2026-10-27 00:00:00 to 2026-10-28 00:00:00
2026-10-28	814	2026-10-28 00:00:00 to 2026-10-29 00:00:00
2026-10-29	1,101	2026-10-29 00:00:00 to 2026-10-30 00:00:00
2026-10-30	908	2026-10-30 00:00:00 to 2026-10-31 00:00:00
2026-10-31	987	2026-10-31 00:00:00 to 2026-11-01 00:00:00
2026-11-01	471	2026-11-01 00:00:00 to 2026-11-02 00:00:00
2026-11-02	1,335	2026-11-02 00:00:00 to 2026-11-03 00:00:00
2026-11-03	1,439	2026-11-03 00:00:00 to 2026-11-04 00:00:00
2026-11-04	1,116	2026-11-04 00:00:00 to 2026-11-05 00:00:00
2026-11-05 (partial)	1,053	2026-11-05 00:00:00 to 2026-11-05 22:48:00

teas_citations: most recent retained observation before cutoff 2026-11-05 22:44:09.335 UTC; source query completed 2026-11-05 22:50:30 UTC. Counts describe retained origin records. An earliest row or a successful query does not prove uninterrupted collection or edge-cache coverage.
teas_sessions: most recent retained observation before cutoff 2026-11-05 22:45:47.000 UTC; source query completed 2026-11-05 22:50:33 UTC. Counts describe retained origin records. An earliest row or a successful query does not prove uninterrupted collection or edge-cache coverage.
Rolling windows overlap and must not be added. All ends are exclusive. Rates use actual elapsed duration.
Named bot families and browser classes are user-agent claims, not authenticated organisations or verified people.
API URL reads are a subset of machine reads, not a count of completed tool operations. Mirrored API trace records are not added.
HEAD and other response-method limitations are exposed separately. Recorded positive bytes are not independent proof of consumption.
Historical session labels are preserved. The 9 September faceted-scrape gate and 12 September unknown-client correction limit human-traffic comparisons.
The September AI-referral projection includes the existing 2 September UTM reconstruction; original archives remain separate.
Read comparisons crossing the 8 August response-status correction are withheld. Earlier raw totals remain visible with partial comparability.
No count here proves that URSY sponsorship caused traffic, a citation, or a sale.

Section 05

Reading and using are different.

From 6 October 2026 the family list also counts xAI (GrokBot, xAI Grok, Grok DeepSearch; marked claimed because xAI publishes no addresses), Mistral (MistralAI User and MistralAI Index, which we can check against Mistral's published addresses, and MistralAI Training), and Anthropic's Claude User and Claude SearchBot.

Reading and using are different. Crawlers read pages. Assistants connected to our MCP server use tools. Both are listed here, in separate tables, and never added together. The assistants that used our tools in the last 30 days:

Assistant	Requests	Tool answers
other	37,190	230
node	6,466	0
script	3,448	7
claude	494	9
unknown	275	0
chatgpt	49	20
perplexity	38	0
copilot	7	0
grok	1	0

MCP server requests by assistant family, teas.co.uk's own tests left out. Directory testers and scripts appear under their own names.

Source: wp_teas_chatgpt_calls grouped by client family; window: last 30 days; live.

Keep reading

Where this leads next.

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

What they read

The doors those agents knocked on, the probe-share honesty split and what came back.

Open the record →

The mint & leaderboard

What these reads earn: the board, the classes and the run records behind them.

Open the record →

traffic/data.md

The dated machine-read windows as a markdown twin, for machines.

Open the record →