Counts answer “how often”. Recall answers “in what words”
Most of what the agent reads for you is arithmetic: how often you were named, how often you were cited, which pages earned it. Those are the right answers to “how are we doing”. They are the wrong answer to “how is that competitor being described”, or “what reason did the answer give for recommending them”, or “is the claim we published being repeated back”. Those questions are about wording, and the wording is in the answers themselves. Recall searches those answers by meaning rather than by keyword. Ask it for “reasons given for choosing a cheaper tool” and it finds the passages that say that, whether or not any of them used the word “cheaper”.Every passage comes with its receipt
A quote you cannot check is an anecdote. So every passage recall returns carries, on the same row:
The agent is told to quote these rather than paraphrase them. A summarised passage is a claim the platform did not make.
What is searched, and what is not
Recall searches the answers themselves, and nothing else. Not the pages those answers cited, not snapshots of your site, not your own documents. That is a deliberate limit rather than an unfinished one. Indexing every cited page alongside every answer would multiply the store roughly tenfold, for passages that are one sentence of somebody else’s page — and that page is better read directly, which the agent can also do. Answers are searchable from the point the feature was switched on. Older readings are indexed in the background, oldest first, so a long history fills in over the following days rather than all at once.Scope and window
Recall only ever searches your own workspace. There is no way to widen it — not from a question, not from an automation’s settings, not by asking the agent to. It searches the last 28 days by default, and up to 90 on request. If nothing in the window matches, it says so. “Nothing in the last 28 days matches that” is a real answer, and a better one than a paragraph assembled from memory.What it costs
A recall search costs a fraction of a credit — it is one small lookup, not a model call, and it is priced accordingly. It appears on your usage log like any other metered step. Reading your counts and boards stays free.Tables and charts
How the agent draws what it finds, and why a gap is never a zero.