ChatGPT, Gemini and Perplexity update the way they work every few weeks — no press releases, no warnings — and every update can wipe out months of work on visibility. Whoever discovers it late ends up starting over from scratch while the competitors who were monitoring have already reacted. There is a minimal system to keep up with these changes without spending hours on it every week, and the difference between those who have it and those who don't shows up in the results within a few months.
The question isn’t “where am I today”. It’s “where will my customers be in 2027?”. Wearable AI, ambient computing, voice-first: the platforms that seem marginal today will be dominant in 18 months. And every time OpenAI releases a new model, every time Google updates Gemini, every time Perplexity changes its source-ranking algorithm, the rules by which your brand gets cited — or not cited — are quietly rewritten.
Let me explain why, without a minimal system for monitoring the evolution of AI engines, you’re applying the strategy of six months ago to a system that has changed its skin twice in the meantime.
The principle: AI systems aren’t finished products, they’re moving surfaces
In my previous articles I talked to you about E-E-A-T for AI and backlinks as a citation proxy. Those levers remain valid as principles. What changes, almost every quarter, is the relative weight that each signal carries inside the model.
In the world of language model and information retrieval research, there’s a documented mechanism: systems that combine generation and retrieval (retrieval-augmented) are periodically recalibrated both on the base-model side and on the side of the document index they draw from. It follows that, for your business, there is no such thing as “optimization done once and for all”. When OpenAI moves from one version to another, when Google rebalances the way Gemini integrates web sources, when Perplexity changes the way it orders the citations at the bottom of the answer, your visibility in AI answers can double or be halved without you touching a single comma on your site.
Translated into practice: if you don’t have a stable observation point, you only notice the change when a long-standing client calls you saying “I was told you no longer show up when people search for me”.
Why Ischia taught me this lesson better than any chart
Over the last eighteen months I’ve watched closely the sector of the historic thermal baths and spas on the island of Ischia: operators like Terme di Negombo and Poseidon Giardini Termali, two of the most recognizable names in the Italian thermal wellness segment. They are brands with real authority, centuries of geological history behind them, excellent reviews and editorial coverage in the major newspapers.
And yet, monitoring natural-language queries like “best thermal baths with sea view in Italy”, “thermal parks open all year round in Ischia”, “historic spas of Southern Italy”, the picture changed visibly three times in little more than a year. In one phase the AI systems cited mainly Wikipedia and aggregated tourism portals. In a later phase they started drawing more from official sites with well-made schema markup. In a more recent phase they gave weight to content signed by an identifiable author and to structured reviews on vertical platforms.
Same brand, same offering, same intrinsic quality. Three different AI citation regimes in twelve-to-fifteen months. Consider this a longitudinal observation, not a study: small sample, artisanal methodology, but the pattern is consistent with what I see with clients in other sectors.
A single drop event is not a trend.
The blind spot of those who don’t monitor: applying yesterday’s strategy to today’s engine
The operational risk is simple. Six months ago you invested in building pages optimized according to a certain principle (for example, short paragraphs in inverted-pyramid style — which remains good practice). In the meantime, the model you used as a reference has changed the way it assesses the author’s authority, giving more weight to recognition of the author as an entity.
You keep measuring the results through the lens of six months ago. You see that traffic holds, maybe impressions in Search Console grow, and you conclude that everything is fine. Except that AI answers — the channel that weighs more and more in B2C and B2B purchase decisions — are citing your competitors who intercepted the new signal.
It’s not a hypothesis. It’s what I saw happen to operators like Negombo and Poseidon in the moments of transition from one regime to another: weeks in which they disappeared completely from queries where they were previously cited, only to come back when the index settled on the new equilibrium.
Open a changelog file (even a Google Doc) with four columns: date, platform, observation, action.
How to build a minimal monitoring system in 30 minutes a month
You don’t need a dedicated team. You need a method. This is what works for me, scalable to any SME.
Official sources to follow. Subscribe to the newsletters of OpenAI, Google AI, Anthropic and Perplexity. They’re free, they come out at variable frequency, and they are the first official signal of a change in model or retrieval logic.
A stable set of target queries. Define 8-12 queries that represent the way a typical customer searches for you or your competitors. For Negombo they could be “thermal parks Ischia”, “thermal baths with outdoor pools Campania”, “historic spas Italy”. Save them in a spreadsheet.
Test before/after every announced update. When a model update comes out, run the same queries on ChatGPT, Gemini, Perplexity and Claude. Note whether you appear, in what position within the answer, with what source. Do the same a week later. The difference is your signal.
Quarterly internal changelog. One document page. Four columns: date, updated platform, change observed on your queries, action decided. That’s it.
This is an entry-level check. The real analysis, on larger volumes and with systematic comparison across engines, requires professional tools. But the entry-level check is already 70% of the value: it gives you the awareness that the ground is moving and in which direction.
The mistakes I see most often
Mistake 1: monitoring only ChatGPT. It’s the most visible, but it’s not the most relevant for many sectors. For local tourism, Perplexity and Gemini with grounding often weigh more in informed decisions. Test on at least three platforms.
Mistake 2: confusing interface updates with model updates. When ChatGPT changes layout or adds a button, nothing changes for your visibility. When it changes model version, it does. Distinguish the two levels.
Mistake 3: changing strategy at every small movement. A single drop event is not a trend. Wait for two or three consistent observations before intervening on the content.
Mistake 4: not documenting. Without an internal changelog, in six months you won’t remember whether that change in the citation pattern came before or after the March update. Operational memory is the real asset.
What to do now, in order of priority
- Subscribe today to the blogs of OpenAI, Google AI, Anthropic, Perplexity.
- Define this week your set of 8-12 target queries for your sector.
- Run a baseline test on ChatGPT, Gemini and Perplexity. Save the screenshots.
- Open a changelog file (even a Google Doc) with four columns: date, platform, observation, action.
- Compare with the 3-5 competitors that AI cites most often in your sector: what do they do that you don’t?
- Repeat the test every 60-90 days or at every official announcement of a model update.
Monitoring the evolution of AI platforms isn’t a technical task, it’s a practice of commercial discipline. It’s what lets you be visible in AI answers twelve months from now and not just today. In the next articles of this series I’ll explain how to strategically compare answers across different AI engines and how to build an alert system for when your brand disappears from a target query.
