AI Visibility
Mention alerts watch the web. They cannot see what an assistant tells a buyer who never visits a page.
Start MonitoringClassic brand monitoring works by watching published pages for your name. That model breaks down when the thing shaping opinion is a generated answer that exists only in one conversation, is never published, is never crawled, and differs slightly every time it is produced.
Last updated
A mention alert fires when your name appears on a page it can crawl. When a buyer asks an assistant which tool to use and the answer names three competitors and not you, nothing is published, nothing is crawled, and no alert fires. The conversation that decided the shortlist leaves no trace. For a growing share of buying research this is now the primary channel, and it is precisely the channel classic monitoring cannot observe even in principle.
Since the answers are not published, monitoring means producing them yourself: run a fixed set of prompts against each platform on a schedule, parse the responses the same way each time, and store the results. That is what makes it monitoring rather than a one-off audit. The discipline that matters is holding the prompt set constant. Adding prompts because they seem better changes the measurement, and the resulting movement is your edit, not the model's behaviour.
Presence alone is a noisy signal because model non-determinism produces movement every run. The events that justify interrupting someone are narrower.
Any monitoring product that computes a score faces a problem it rarely admits: when the scoring logic improves, every stored score moves. Reporting that as the customer's trend is straightforwardly wrong, and it is the most common way these dashboards mislead. We stamp an engine version onto each result and suppress the comparison across a version change, saying so on the page and in the email, rather than letting an improvement to our code appear as a drop in your visibility.
Monthly suits most brands. Retrieval results shift weekly, but weekly checking mostly measures noise and generates alert fatigue, and each check costs real model calls. Move to weekly around a launch, a rebrand, or a competitor's funding announcement, when the underlying reality is genuinely changing faster than usual. Otherwise, monthly checks with a stable prompt set produce a trend line you can actually act on.
| Mention alerts | AI monitoring | |
|---|---|---|
| What it watches | Published, crawlable pages | Answers generated on demand, never published |
| How it gets data | Crawls and indexes what exists | Runs a fixed prompt set and records the responses |
| Trigger | Your name appears somewhere new | Your position, sentiment, or rivals change across runs |
| Blind spot | Every private conversation with an assistant | Anything outside the prompt set you chose |
| Cost per run | Effectively zero once indexed | Real model calls, which is why cadence matters |
| Main risk | Noise from irrelevant mentions | Mistaking model variance for a real trend |
No. Alerts index published pages. An assistant answer is generated in a conversation, never published, and never crawled, so there is nothing for an alert to find.
Monthly for most brands. Weekly around a launch or a competitor event. More often than that mostly measures model non-determinism and costs real money per run.
A sustained drop across several checks, a shift to negative sentiment, a new competitor appearing in your answers, or a recurring factual error. A single-run fluctuation should not.
Yes, and you should. The same prompts that reveal your position reveal theirs, and a rival newly appearing in answers you used to own is often the earliest warning you will get.