Read like a machine. Write like a person. Prove it from the logs.
AI assistants such as ChatGPT, Claude and Perplexity answer in place. Most people never click. So the thing to measure is not the visit but the answer: which assistant fetched which of your pages to reply to whom, and whether it named you. LLMs Bureau is built around that one reading. Here is each instrument, labelled by the access it needs.
The Monitor
We put the questions your buyers ask to ChatGPT, Claude, Perplexity and Gemini and record who gets named, in what position, and who is named instead of you.
Named or not, per question, per assistant, with history from day one.
The Tag
A script served from our domain, no cookies, no browser storage. It sees the visitors an assistant sends, the pages they read, phone taps and form sends.
Visitors from assistants, and enquiries attributed to the assistant that sent them.
The Edge Tag
Crawlers do not run scripts, so a page tag never sees them. A small snippet at your edge reports each crawler hit as user agent, path and status, never an IP.
Answers served per assistant per page, and which index crawlers have fetched which pages. WordPress plugin and PHP middleware: coming.
Change orders
We read your site the way an assistant does, score it, and write each fix as a numbered copy-paste block for your developer. The method stays ours; the result goes on your site.
Applied, verified or pending, with the reading each order was meant to move and whether it moved.
Answers (a person just asked)
- ChatGPT-User, ChatGPT
- Perplexity-User, Perplexity
- Claude-User, Claude
- Meta-ExternalFetcher, Meta AI
- DuckAssistBot, DuckDuckGo AI
- MistralAI-User, Le Chat
Index builders
- OAI-SearchBot, ChatGPT search
- Claude-SearchBot, Claude search
- PerplexityBot, Perplexity
- Bingbot, Bing, Copilot, DuckDuckGo
- Googlebot, Google, AI Overviews, Gemini
- Applebot, Apple
- Amazonbot, Alexa
Training crawlers
- GPTBot, OpenAI
- ClaudeBot, Anthropic
- Google-Extended, Gemini
- Meta-ExternalAgent, Meta
- Bytespider, ByteDance
- CCBot, Common Crawl
Index builders hold the page. Live fetchers read it for a person. We report both and never mix them.
Numbers with names on them.
Nine panels. Each reading carries its date range and its source. Inferred questions are marked likely. Nothing is a model's guess dressed as a fact.
Answers served
Per assistant, per page, per day, from the Edge Tag. Each row shows day, assistant, page and a "likely asked" inference; the column header literally reads Likely asked.
Named or not
The Monitor runs: per question and assistant, named yes or no, position in the answer, competitors named, with a 30 and 90 day trend.
Crawler coverage board
Which index crawlers have fetched which of your pages and when they last did, labelled "pages fetched by each index's crawler" (a log shows fetches, not index membership), with amber lamps on gaps: not fetched by Bing, blocked at the edge, JavaScript-only page.
Visitors from assistants
From the Tag, by assistant, landing page, pages read, phone taps, form sends; UK share; no cookies, no IPs stored.
Enquiries attributed
Each enquiry with the assistant that sent the visitor and the page they arrived on, with a field for you to mark the value won.
Change orders
The numbered list, status applied, verified or pending with dates, the reading each one was meant to move and whether it moved.
Search feeds
Bing Webmaster clicks, impressions and crawl, and Google Search Console clicks, impressions, position and impressions inside AI features once we are added as a user. AI-feature impressions are labelled impressions, never citations.
Crawl health
Edge blocks, errors seen by crawlers, pages whose text needs JavaScript, machine-readable files gone stale, rules that deny a named crawler.
Monday briefs
The archive of every weekly brief, with the numbers as they stood that week.
The reading sheet.
One email every Monday at 08:00 UK time, plain text with an HTML twin in the site's type, two phone screens long. The subject line is the week's reading in a fixed form: readings to the date, answers served, pages cited, enquiries. Every figure is yours, from your own Edge Tag, Tag and feeds, with its date range.
- Readings
Answers served by assistant, pages cited, new assistants seen, click-throughs, enquiries attributed.
- Compared with last week
One line per reading, up or down, no adjectives.
- What changed
Crawlers that arrived or left, coverage, blocks found.
- What we did
Orders written, orders verified, pieces published.
- What to do next
At most three orders, one sentence each, with a link to the full order.
- One closing line
Whether the curve is turning, in plain words. On Run a named person reads the brief before it goes.
Subject: Readings to 14 Oct 2026: answers served up 300% week on week
READINGS answers served a day up 10x since
ChatGPT indexed the site
ChatGPT, DuckDuckGo AI and Perplexity answering
PAGES CITED home page 75% of answers
pricing, trace a debtor, reconnect with someone,
for solicitors, developers, how it works
COVERAGE Amazon Alexa 100%, Meta AI 97%,
Bing 84% of the site read
CLICK-THROUGHS the first from ChatGPT on 14 Oct,
one straight to the instruction form
WHAT CHANGED ChatGPT's search crawler took the site in by 10 Oct
NEXT one change order, with the exact text to paste
CLOSING the curve turned the day after the crawlEvery figure above is from our address-tracing site's server logs, 29 Sep to 14 Oct 2026. A client's brief carries the client's own figures and nothing else.
We do not control your site. Here is how the fixes happen anyway.
The meter
Watch is the reading. We install nothing on your pages beyond the one script line and the edge snippet your developer adds. The monthly readability check tells you what an assistant sees; the brief tells you what moved.
Written for your developer
The Audit reads your site from outside, as a browser and as the named AI crawlers, and every finding becomes a numbered copy-paste block: the exact text, where it goes, why, which reading it should move and over what window. The method behind the blocks is ours and stays ours; your developer only has to paste. They mark each order applied; we re-fetch within 24 hours and mark it verified with a date, or reopen it with what we found.
By your login or by pull request
On Run we apply the orders through a named editor account on your CMS, or as pull requests your developer merges. You approve each change before it publishes and every change is listed in the dashboard. We also write one dated, answer-first piece a week for your approval, aimed at a question the Monitor shows you losing.
The change orders include the off-site asks only you can do: answer the reviews, claim the listing, earn the mention, because most citations are earned off the site and we say so plainly. Index work that needs no site access we do from our side, with you added as a user on the search consoles, and a weekly check that the named crawlers reach the pages. For a site on managed WordPress hosting with no server access, the first change order is usually "put free Cloudflare in front of the site", which most hosts allow in twenty minutes and which makes the Edge Tag possible.
We never host, proxy, cache or inject into your site. Nothing promises a position or a number of citations; the plans sell the readings and the work, not the outcome.
Questions people ask before they start.
Straight answers. If the answer is that you do not need us yet, we say that too.
Do you need access to my website?
Not to start. Watch needs nothing from you except the Tag line and the Edge Tag snippet, which your developer adds in minutes. Fix writes the changes for your developer to apply. Only Run needs an editor login or a repository, and you approve every change before it publishes.
What is an answer served?
A person asked an assistant something and the assistant fetched your page, right then, to reply. We count it from the server log by the user agent (ChatGPT-User, Perplexity-User, Claude-User and the others), one answer per assistant per address within two minutes. The page it fetched is the topic it used you for.
Why can a page tag not count that?
Because crawlers do not run JavaScript. A tag on the page sees people in browsers. The assistants fetch the HTML and leave. Only the server, or an edge in front of it, sees them. That is what the Edge Tag reports: user agent, path and status, never an address.
Do you store IP addresses?
No. The Tag stores none. The Edge Tag sends none. The free check keeps a daily salted hash of the visitor address for the rate limit only, and the privacy notice says so.
What do you never promise?
A position, a number of citations, or a number of enquiries. The plans sell the readings and the work, not the outcome. Most citations are earned off the site, by mentions and reviews, and the change orders say which of those are yours to do.
Which sites can you work with?
Any site we can fetch. WordPress, Webflow, Squarespace, Shopify, a custom build, a React or Next site. If the text only appears after JavaScript runs, the first change order is to server-render or pre-render it, and we say so before you pay for a month.