LLMs Bureau[1]
Case files / dated / sourced

What the assistants did with our pages.

Two of the files are our own companies, measured from the day they went live, so you see the method on a site with no history, no backlinks and no paid links. The third is a London firm whose own analytics we show with its name removed. Every number carries its date and source. Index builders and live fetchers are reported separately and never mixed.

File 01 / Address-tracing service
The fortnight
10x
answers served a day, once ChatGPT indexed the site
Server logs / answers a day, 29 Sep to 9 Oct against 10 to 13 Oct 2026

Week two up 300% on week one. Read by Amazon Alexa (100%), Meta AI (97%) and Bing (84%). First visitors from ChatGPT on 14 October.

File 02 / Investigation Bureau
Five pages, one answer
98%
of the site read by Google in nine days
Server logs against the sitemap / 5 to 14 Oct 2026

Anthropic 94%, OpenAI 91%. ChatGPT answered from 12 different pages in the first week, once from five pages at a time.

File 03 / a London investigation firm
The £3,300 answer
+207%
visitors sent by ChatGPT, August to September
The firm's own analytics / Aug to Sep 2026

8 tapped the phone number, 3 sent the form, one became a £3,300 instruction. Shown redacted.

How we count

An answer served is a person asking an assistant and the assistant fetching our page, right then, to reply: the user agents ChatGPT-User, Perplexity-User, Claude-User, Meta-ExternalFetcher, DuckAssistBot and MistralAI-User in the server log, with fetches by the same assistant from the same address within two minutes counted as one answer. The page it fetched is the topic it used us for. The question text itself is only visible in Bing Webmaster's AI Performance report and Search Console's AI report, so the questions we show are marked likely. Index builders (OAI-SearchBot, GPTBot, ClaudeBot, PerplexityBot, Bingbot, Googlebot and the rest) are counted separately as pages fetched, because a fetch shows a crawl, not index membership.