AI agent search
Where the search indexes an agent calls put your brand, which is not where the assistants put it.
Every other page under Measure reads a chat assistant: a question was asked, an answer came back, and the answer named somebody. An assistant that can search does not work that way. It calls a search index, gets a ranked list of URLs back, and writes from what it was handed — and the indexes it calls are not the ones a person uses.
This page is that layer, measured on the same questions as the rest.
Why more than one index
Three of those indexes are measured directly — Brave, Exa and Tavily — and they disagree.
Over thirty tracked questions across three brands, Brave and Exa found a brand eleven times between them and agreed on two of those eleven. A single index would have reported four fifths of this wrongly — in either direction. So each is drawn on its own, and none is averaged into another.
These are separate readings, not one score
There is no combined number on this page on purpose. Averaging an index that returns you at #1 with one that does not return you at all produces a figure that describes neither, and hides the only thing worth knowing: which of them an agent happened to call.
Summary
One panel above the table, folded to its header line with Hide (the page remembers). Its header still reads the headline: found for N of M questions, and how many no search finds you for. Open, it has three columns:
- Found by any search — the share of questions at least one search returns you for, and a bar splitting them into found by every search, by some, and by none.
- By search — a row per search: who uses it, its hit rate, and a bar. A new search is a new row.
- In your place most often — the sites that most often sit in the top three where you are missing, as “N of M” questions a search missed you on. Competitors you track are in orange.
Reading the table
One row per question, the ones no search finds you for first.
- A column per search. The header carries its mark, how many questions it
finds you for (
Exa 3/5), and — on hover — who uses it. Claude’s use of Brave is reported (Anthropic lists Brave as a subprocessor), not stated in Claude itself; Exa names Cursor among its customers on its own site; Tavily ships as a ready-made tool in LangChain and grounds JetBrains’ Junie. A search added recently readsnot yetuntil your brand’s weekly turn. A cell is your place in that search’s results (#4), or a dash. - Ranks instead — the sites that came back where you did not. Competitors you track are in orange. These are the pages to read to find out why.
Open a row for the full results each search returned, your own result tinted.
If a search has none of your pages at all, a line above the table says so: no question can return a site it does not hold, which is a crawling problem, and Crawlability is where it starts.
How often it runs
One project is checked each scheduler tick and each project is checked weekly, so a newly tracked brand waits its turn rather than being checked immediately. Projects are ordered by age — oldest first — so a queue, not a draw.
If the deployment has no index key configured the page says so rather than reporting zero of twenty, which is a finding it has not earned.