AI visibility alerts without false positives

The AI visibility alerts with the fewest false alarms are the ones that fire on a trend across repeated readings, never on a single answer: AI engines can answer the same prompt differently each time they are asked, so an alert on one reading mostly reports noise.

Most AEO and GEO platforms now offer alerts: your brand dropped out of an answer, a competitor overtook you, a citation disappeared. The trouble is that the thing being watched moves on its own. Ask ChatGPT, Gemini or Perplexity the same buyer question three times and you can get three different lists of brands. That is answer variability, and it is why an alert that fires on every change trains a team to ignore its alerts within a fortnight.

This guide covers why single-reading alerts misfire, the design choices that cut false alarms, how to test a vendor's alerts during a trial, and how alerts work in Stellarcast, which we build. We publish no false-positive rate for our own alerts, and we would be wary of any vendor that quotes one without saying how it was measured.

Why single-reading alerts misfire

A single AI answer is one draw from a range of possible answers. The model samples its words with some randomness, the search results it grounds on change, engines update their models, and some tailor answers to location. None of that means your visibility changed.

So an alert rule like "tell me when we are not named for this prompt" fires every time the engine happens to leave you out, even when you are named most days. Multiply that across prompts, engines and markets and a team gets several alerts a day, almost all of them about noise. The real drop, when it comes, arrives in an inbox nobody reads any more.

The fix is not a better threshold on the same single reading. It is a different input: many readings, pooled, compared against how much they normally move.

Alert design choices that reduce false alarms

None of these is exotic. Together they are the difference between an alert that says "something moved" and one that says "something moved more than it normally does, for more than one reading, and here is the evidence".

How to test a vendor's alerts in a trial

1. Leave alerts on for two weeks without changing your site. If nothing on your side changed, every alert you get is either a real market move or noise. Count them.
2. Check each alert against the trend. Open the chart behind it. Was the drop one reading, or several? Did it recover the next day?
3. Ask what the alert compared. One answer against the last one, or a window of readings against the previous window? A vendor that cannot say is probably comparing single runs.
4. Ask whether the figure carries a range. A share of voice with no band cannot tell you whether a move is unusual. The P25-P75 range is the middle half of what the readings gave.
5. Check the channels you will actually read. An alert that only lives in a dashboard you open monthly is a report, not an alert.

How Stellarcast alerts work

On every paid plan, Stellarcast runs each tracked prompt on every engine daily, per market, and keeps every reading. Alerts are built on those readings, not on single answers. These are the alert types the product sends today:

• You lost ground. Fires when your share of voice in a market falls by 5 points or more between two consecutive daily snapshots. Each snapshot is the trailing seven-day share, blended across engines and weighted by how much each engine was sampled, so one odd answer is pooled with a week of others. A first reading never alerts, and a day with no readings is skipped rather than read as zero. The alert shows both shares and both dates so you can check it, and a given drop is sent once.
• Lift confirmed (from the Growth plan). Fires when a fix's lift moves from LIKELY or TOO-EARLY into MEASURED. MEASURED needs at least 14 days of readings after the change, a difference outside the range seen before it, and no confounding model update; the alert carries the label, and the range when there is one.
• Market movement. A daily email when an engine's model update, rather than one of your fixes, moved your standing.
• Waiting on you (from the Starter plan). A drafted fix is waiting for your approval, or a connection broke and needs reconnecting.
• Usage limits. At 80% and 95% of your plan's allowance, once each per billing period, so you are never stopped by surprise.
• Weekly digest. One email that rolls up the week's movement, waiting fixes and alerts, so the inbox stays quiet between real events.

Every alert type can be switched on or off separately for email and for Slack, and alerts also appear on an Alerts card in the app. The share of voice behind them carries its P25-P75 range on every screen; the method is in how we measure AI visibility.

What we do not claim: a false-positive rate. The drop alert compares two real snapshots, and it does not carry a range of its own, because a band invented for a day-over-day comparison would be a number we could not defend. Alerts start on the $39 Lite plan; see pricing.