What a finding is
Every hour, a pass reads your workspace’s recent readings and checks them against five fixed rules. The rules are arithmetic over the same numbers your dashboard shows, with no model and no prompt being asked. That makes them fast, cheap and consistent: the same window always produces the same findings. What a rule finds is then measured before it is filed, so that sampling noise does not reach your Inbox. See Noise, or a real change.The five rules
A page of yours stopped being cited
A page of yours stopped being cited
A page on your domain was cited on at least three of the last seven days, and on none of the last three — with at least one successful read on each of those three days, so a run of scraper failures never looks like a loss.
A competitor was newly named
A competitor was newly named
A rival is named on a prompt where it was not named before: named on at least two of the last three days, and never named across the seven days before that — counted only over days you actually read that prompt on that engine.
You dropped out of every answer on an engine
You dropped out of every answer on an engine
A page of yours was cited somewhere in the prior seven days, and in none of the last three, on one engine — with every one of those three days carrying a successful read. This is the whole site going quiet on an engine, not one page slipping.
A new domain entered an engine's top ten
A new domain entered an engine's top ten
A domain that was not among an engine’s top ten cited sources is now in it. This only fires once that engine has a full prior week to compare against — an engine you started sampling this week does not report its entire source list as new.
One of your pages changed
One of your pages changed
A page on your own domain changed since the last pass. Not a loss, and never filed as one — it is surfaced because it is usually the answer the other four rules are asking for.
A day that could not be read never counts as a loss
This is the rule everything above is arranged around. Every finding that claims an absence — a citation stopped, a page dropped out — first checks that every day it is claiming silence from carried at least one successful read. A day with three failed reads and a day with no reading at all look identical to this pass on purpose: both are silence, and silence is never evidence of you losing something.Noise, or a real change
A rule says what moved. It cannot say whether it moved by more than sampling moves on its own: a prompt read three times a day can lose a citation for three days on luck alone. So each finding is measured before it is filed.- The arithmetic. The finding’s last three days are compared with the seven days before them, on successful reads only, with Fisher’s exact test. A change significant at p < 0.01 is always filed.
- One decision, on numbers only. A finding that is not significant, and has no Inbox row open yet, is put to a decision: is this a real change, or noise? The decision reads the per-day counts (successful reads, failed reads, answers that named you and did not), the two windows’ rates and the p-value. It never sees a domain, a URL or a prompt’s text.
- Held as noise. Only when the decision is confident the change is not real is the finding held. A held finding is not filed, is not an event, and starts no run. It is held for that hour only: the next pass measures it again, asks again once new readings have landed, and files a change that holds up. Each hold is recorded with the n it was judged on, the number of successful reads across both windows.
One row per deciding source, not per prompt
A source that decided seven prompts is one problem, not seven. Findings are grouped by the source that caused them before they reach your Inbox, so the row reads “affects 7 prompts” and grows in place instead of filing seven near-identical rows. The Inbox stays worth opening: at most three open rows from this pass at a time. A fourth finding bumps the closest existing row rather than raising a new one.What happens when a row is raised
A run may start under your workspace’s approval mode — the same setting that governs everything else the agent does on its own. What a finding is allowed to do next, from opening the row for you to reading further to proposing a fix, follows that setting exactly as a chat or automation would. In Draft and Act modes the run starts on its own, at most once per finding per day. It says why it started, in the finding’s own words: Because example.com stopped being cited. In Ask mode no run starts; the row opens a chat with the question ready for you to send.Every finding is also an event
Each finding also becomes an event that event automations can react to, even when the three-row Inbox cap stops it from filing a row of its own. A finding held as noise does not become an event.
The same event is recorded at most once per subject per UTC day, so a loss that persists does not wake an automation every hour.
Detection and the old Daily check-in
Detection does not pause the Daily win plan; the two run side by side. It only retires an old, untouched Daily check-in automation that still runs the daily-losses procedure, and only when detection is on for the workspace, so the same loss is not reported twice from two places. A check-in you rewrote is left alone.Who has it
Paid workspaces are enrolled in hourly detection automatically, the first time their subscription is active.Findings are read-only in what they claim: a rule never invents a percentage from a single day, and a claim about a page always comes from a page the pass actually read.