You may have lost visibility on AI engines four months ago and only find out today — because no system warns you when your brand stops appearing in the answers. Meanwhile, the customers asking that question are being handed your competitors' names. Setting up an alert system takes less than half an hour and gives you back control: you immediately know what changed and you can act before the problem takes hold.
Set up an alert via Slack or email: when your brand changes position or sentiment on AI engines across ten tracked queries, you get a notification. Thirty minutes to build it, months of advantage over those who notice too late. Here I show you what to set up today so you don’t discover in six months that ChatGPT stopped citing you back in February.
In the earlier articles of this series on measuring AI visibility I talked to you about dashboards and reporting. Now I’m bringing you the piece that makes the difference between “checking the numbers every Monday” and “being notified the very day something changed”. It’s called alert and anomaly detection, and for an Italian SME it’s the difference between reacting within the day and noticing the problem when revenue is already falling.
What I mean by AI visibility alerts
An AI visibility alert system does something very simple: it monitors a fixed set of queries on AI engines (ChatGPT, Perplexity, Gemini, Claude) at regular intervals, compares the results against a history, and notifies you when something changes beyond a threshold you have set.
The four variations worth monitoring are always the same:
- The brand disappearing below a weekly mention threshold
- The appearance of a new hallucination (the AI invents a false fact about you)
- The emergence of a new competitor in answers where you used to stand alone
- A change in sentiment (from neutral to negative or vice versa)
I’m not talking about dashboards. You look at dashboards when you remember to. Alerts come looking for you.
Why it matters, in terms of money
In the field of AI search research there is not yet a public paper quantifying the economic damage of “discovery latency” — the time between the moment the AI stops citing you and the moment you notice. It follows that we have to deduce it from the adjacent principles I have already explained to you in this series.
The principle is this. We know from the articles on E-E-A-T for AI and backlinks as a citation proxy that AI models lean on authority signals that consolidate over time. A sudden disappearance from AI answers is almost always the symptom of a measurable upstream event: a competitor that published a more complete asset, one of your key pages getting deindexed, a hallucination that replaced you with another brand in the model’s knowledge graph.
Translated into practice: every week of delay in spotting the anomaly is a week in which a competitor consolidates its position in the model’s “implicit training”. The operational consequence is that reacting in three days instead of three months changes the order of magnitude of the recovery work.
If you monitor only the classic SERPs, you miss 90% of the phenomenon.
The case of a truffle hunter in San Miniato
Let me tell you a concrete, anonymized case. A small truffle-farming and truffle-hunting business in San Miniato, in the province of Pisa — seven cultivated hectares, direct sales, guided experiences with a truffle dog for food-and-wine tourists. A niche sector, very specific AI queries, zero margin for error because anyone searching “where to buy white truffle San Miniato” or “truffle hunting experience Tuscany” is already ready to buy.
In November 2025 we set up an alert on ten target queries, weekly frequency, notifications via email + Slack to the owner. Total setup: just under an hour, because we had to decide the queries and thresholds together.
In February 2026 an alert fired: on four queries out of ten, the brand had been replaced in Perplexity’s answers by another company in the same area. The sentiment was neutral, no hallucination, the competitor had simply emerged. Investigating, we discovered that in January the competitor had published a long, well-structured guide on the truffle season, cited by two local outlets and by a fairly authoritative gastronomy blog.
Without the alert, we would have found out in May, when the owner would have noticed the drop in experience bookings. With the alert, we had three months of head start to publish a response asset, recover two missing citations and rebalance the presence. By June the split was 60-40 in the client’s favor on the same ten queries.
An indicative test, a single client, a niche sector. But I see the pattern repeat itself: those who have alerts react in days, those who don’t notice problems when revenue drops.
Ten well-chosen queries beat a hundred generic ones.
The setup you can do in thirty minutes
You don’t need complicated infrastructure. To get started, all you need is:
- A list of 10-15 real queries that your customers would run on ChatGPT or Perplexity before buying in your sector. Not brand queries (“who is company X”), but problem queries (“best producer of Y in area Z”, “how to choose W”).
- A fixed frequency: weekly is the minimum, bi-weekly is the maximum acceptable.
- A binary alert threshold: either you’re among the first 5 cited sources, or you’re not. No continuous metrics, no “visibility score” to interpret.
- A notification channel you actually open: the owner’s email + an internal Slack channel. Not a monthly PDF report nobody reads.
For the technical execution of the queries on a fixed schedule you need professional tools — here I’m not suggesting DIY tools, because reliable execution via API has costs and setup that go beyond thirty minutes. What you can do today for free is a manual version: a Google sheet with ten queries, a column for each week, a quick check on Monday morning across the AI engines, a colored highlight when a row changes. It’s rough, but it’s infinitely better than nothing.
Real analysis requires professional tools with automated execution, screenshots, sentiment parsing and historical archiving of the answers.
The most frequent mistakes
I’ve seen many entrepreneurs try it on their own and trip over it. The recurring patterns are four.
Too many monitored queries. A hundred queries look professional, but they actually generate noise: every week some of them swing, and it becomes impossible to tell the signal apart. Ten well-chosen queries beat a hundred generic ones.
Thresholds too tight. If the alert fires every time a position changes, after two weeks you mute it and stop looking. The right threshold is a “qualitative event”: total disappearance, a new player appearing, a hallucination, a sentiment reversal. Not micro-variations.
Alerts with no predefined action. The alert arrives, the owner reads it, doesn’t know what to do, files it away. You must have already written down what happens for each type of alert: who gets notified, within how many hours, what first check is run. Three lines of playbook are enough.
Monitoring Google only. If you monitor only the classic SERPs, you miss 90% of the phenomenon. ChatGPT, Perplexity, Gemini and Claude answer differently from one another: a disappearance on Perplexity may not show up on ChatGPT and vice versa.
The starting audit, two steps
Before thinking about the tool, do two things in half an hour.
First: write down on a sheet the ten queries that a potential customer of yours would run on ChatGPT before buying. They must be problem queries, in your language, without your brand in them. Compare them with the 3-5 competitors the AI cites today in your sector: if the competitors are always the same across all ten, you’ve found your monitoring set.
Second: define the binary alarm rule. Operational example: “if on 4 queries out of 10 the brand doesn’t appear for two consecutive weeks, the alert fires”. A clear threshold, an automatic decision, no interpretation.
Once you have these two elements, the technical implementation is the easy part and gets delegated.
Where does all this lead?
Alerts are the operational layer that makes the rest of the AI visibility work usable. Without alerts, even the prettiest dashboard tells you what happened after the fact. With alerts, the thread of “showing up in AI answers” becomes a continuous process instead of an annual audit.
In the next articles of this series on measuring AI visibility I’ll bring you the structured reporting piece to put on the owner’s desk, the leading indicators of citation loss, and the piece on how to prioritize the queries to monitor by sector.
If you want to understand where it all starts, the link to the authority perceived by the AI is in the article on author entity recognition: without a recognized author entity, the alerts will only tell you that you’re not there, never why.
