Alerts
You shouldn't have to keep re-checking every prompt to notice something changed. After every tracking run, six detectors do that watching for you — this is where they tell you what they found.
What the Alerts feed is
The Alerts screen (/[org]/projects/[project]/alerts) shows the newest 100 alerts for the selected filter, newest first, under the header line “What changed in your AI answers — detected after every tracking run.” Four filter tabs sit above the list — All, Critical, Notice, Wins — each showing its own all-time project count, so the counts can be larger than the displayed slice. Filtering happens before the 100-row limit, so older wins or notices are not displaced by critical volume. A Mark all read button appears whenever you have unread alerts.
A small badge on the bell icon in the Visibility dashboard header shows your unread count from anywhere in the project.
The six alert types
Every alert is one of six types, each with its own trigger condition checked automatically after a tracking run completes:
| Alert | Trigger | Severity |
|---|---|---|
| Visibility drop | Your project’s (or one platform’s) Visibility Score falls 10+ points below its trailing 7-day average in 2 independent tracking runs, after at least 3 baseline days. | Critical |
| Prompt lost | Your brand goes absent from a prompt it used to appear in, for 3 consecutive runs in a row. | Critical |
| Citation lost | A domain of yours stops being cited on a prompt where it was cited in at least 2 of the last 3 runs, and stays uncited for 2 independent tracking runs. | Critical |
| New competitor | An untracked brand name appears in at least 20% of answered runs over the last 7 days (and at least 3 answers across 2 tracking batches). With model access, a semantic judge must also confirm it is a plausible competitor in context. Keyless mode uses the repeated-name thresholds and deterministic non-competitor exclusions. | Notice |
| Competitor overtake | A competitor’s share of voice exceeds yours for the first time in 30 days and remains ahead for 2 independent tracking runs. | Notice |
| Win | Two independent tracking runs confirm that you regained a prompt after 3+ absent runs, earned a citation you did not have recently, or passed a competitor you were trailing. | Win |
“Win” is really three related detectors folded into one feed — a prompt regained, a citation gained, or a share-of-voice overtake in your favor. These correspond to prompt loss, citation loss, and competitor overtake; there is no separate visibility-drop recovery win, and New competitor has no positive mirror.
Severities and how they’re shown
Every alert card carries a colored left border and a matching status badge for its severity: Critical (rose) — something is actively getting worse and probably needs a fix; Notice (amber) — worth knowing, not urgent; Win (emerald) — something improved. Unread alerts are shaded and carry a small new tag. Opening an alert does not change read state; Mark all read is the action that clears the project’s unread alerts.
Reading an alert card
Each row shows the alert type label, a one-line title (“Visibility down 12 pts on ChatGPT”), a detail sentence with the actual numbers behind it (“Your visibility score on ChatGPT fell to 42.3 on 2026-07-10 — 12.1 points below its 7-day average of 54.4”), and how long ago it fired. The whole card is a link — clicking it takes you to the relevant investigation view: a prompt’s answer with the relevant platform pre-highlighted, or the Visibility dashboard for score-level and share-of-voice alerts. A New competitor alert links back to this feed because its handling action lives on the row. These destinations provide context, but the current UI does not open an exact historical run or render the full stored receipt.
New alerts keep a versioned evidence receipt
Behind the summary, each newly created alert stores the rule thresholds, prompt/platform/competitor scope where applicable, the observed values, its reference window, and the independent confirmation observations. That machine-readable receipt makes the judgment auditable even though the feed currently shows only the title, detail, and relevant destination. Legacy alert rows may predate this receipt format.
Tracking a new competitor from an alert
A New competitor alert carries its own “Track as competitor” button right in the feed — the same one-click action available on the Competitors screen’s untracked-brands list. Using it creates the competitor, backfills its mention history from every answer already captured, and refreshes your visibility rollups. Once handled, the button is replaced with a ✓ Tracking tag so you don’t track the same brand twice.
When alerts run
Detection isn’t a scheduled job you wait on — it runs automatically right after every tracking batch finishes, comparing the new run’s data against your project’s history. Every alert is deduplicated against the specific occurrence that caused it, so re-running detection (or a retried batch) never produces the same alert twice. If a project has no alerts yet, the empty state explains exactly what’s being watched and links you to run tracking to start feeding the detectors.
Alerts need history to fire
Most detectors compare a new run against prior runs — a prompt can’t be “lost” until it was once gained, and a drop needs a trailing average to drop below. Expect your first tracking runs to be quiet on the Critical and Notice tabs; alerts become meaningful once you have a few runs of history to compare against.
Delivery: in-app, email, Slack
Every alert always appears in this feed — that’s the one channel that’s never optional. On top of that, an alert can also be emailed (critical alerts, when email delivery is configured) or posted to a Slack channel (when a webhook is configured and the plan allows it). Rule activation is per type, but delivery is project-wide: one email flag and one Slack webhook are copied across all six rules. Delivery is set per project in Alert settings — this page only covers what the alerts mean and where they take you, not how to configure delivery.
Key takeaways
- Six detectors run automatically after every tracking batch: visibility drop, prompt lost, citation lost, new competitor, competitor overtake, and win.
- Three severities — Critical (rose), Notice (amber), Win (emerald) — each has an aggregate count while the selected tab displays its newest 100 rows.
- Every new alert stores a versioned evidence receipt, while the current card links to a relevant prompt, platform, dashboard, or feed action rather than an exact historical run.
- A New competitor alert can be resolved with one click, right from the feed, without leaving for the Competitors screen.
- In-app delivery is always on; per-type toggles control detection, while email and Slack delivery settings are shared project-wide.
Next step
Configure which alerts fire and where they’re delivered in Alert settings. Or head back to the Visibility dashboard to see the full picture these alerts are drawn from.