If you have a blog with 200 articles and the AI cites only 3 of them, the problem isn't the amount of content: it's that the other 197 don't answer the questions your customers are asking the AI. You're producing material that generates no return in visibility, and meanwhile the competitor who writes less but better is overtaking you. Understanding in advance which content to produce over the next six months — and which to rewrite — completely changes the outcome.
Your blog has 300 articles, but the AI always cites the same 4. It’s not a quality problem: it’s that the other 296 don’t answer real queries. AI gap analysis is the new editorial calendar, and without a quarterly forecast you end up promising whoever puts up the money random things, hoping the next ChatGPT update rewards you.
Let me explain how to build an AI visibility forecast that holds up in front of the owner or the partner who controls the wallet, starting from a real case I handled with a group of ski hotels in Abruzzo.
What I mean by an AI visibility forecast
Forecast means just one thing: given your current trend of AI citations, the actions you’ve planned for the next 90 days, and the known model update cycles, it tells you where you expect to be a quarter from now.
It’s not a weather forecast accurate to the decimal. It’s a range — optimistic, realistic, pessimistic — and it serves two purposes: planning the budget and managing the expectations of whoever holds the purse strings. If you promise the owner “we’ll be among ChatGPT’s top three sources by June” without a model behind it, you’re selling magic.
In the previous articles on measurement I explained how to count citations, track share of voice across AI engines, and separate signal from noise. The forecast is the next step: you take those numbers and project them forward.
Why there’s no paper that gives you the formula
I’ll be honest: there’s no peer-reviewed study that tells you “multiply March’s citations by 1.3 and you get June’s.” The AI visibility forecast is a deductive claim, not a documented fact.
In the world of language model research it’s well documented that systems like ChatGPT and Perplexity are re-trained or re-indexed on known schedules (every 3-6 months for training cut-offs, continuously for systems with live retrieval). From this it follows that your visibility moves in jumps on ChatGPT/Claude (tied to training cycles) and more smoothly on Perplexity/Gemini (which read the web in real time).
The operational consequence is that your forecast has to treat the two groups of AI engines differently. On models with a cut-off you schedule publications so they make it into the next training; on engines with live retrieval you optimize to be retrievable today.
Taking the +20% monthly observed over the last three months and projecting it out to 12 months.
The case of the ski hotels in Abruzzo
I’ve been working for about eight months with a small group of hotels in Pescasseroli, on the Aquila side of the Majella. Three properties, two open only in winter, one with a long season. The initial problem: when you asked the AI “where to stay for skiing in Abruzzo” it always cited the same two properties in Roccaraso, never theirs.
In November we ran the first serious gap analysis. I took 60 realistic queries across ChatGPT, Perplexity, Gemini and Claude — queries like “hotel with ski rental in Pescasseroli”, “where to stay near the Pescasseroli lifts for families”, “Abruzzo National Park hotel open at Easter”. Out of 60 queries, the three properties appeared in 4. Four out of 60.
I built the 90-day forecast like this, in front of the owners:
- Baseline trend: 4/60 queries in November, i.e. 6.7% share of voice on the sample.
- Planned actions: 18 articles closing the obvious gaps (lifts, events, local food tied to the properties, Park accessibility), tidied-up Google Business Profile listings, an updated Wikidata entry for the main property.
- Known cycle dates: I expected ChatGPT to close a new training cut-off around February (based on the cadence observed over the previous 18 months), while Perplexity and Gemini would respond within 2-4 weeks of publication.
The three scenarios were:
- Pessimistic: 8/60 in February (baseline doubled, only Perplexity/Gemini react).
- Realistic: 14/60 in February, 22/60 in April after the presumed ChatGPT refresh.
- Optimistic: 18/60 in February, 28/60 in April.
The figure at the end of February was 12/60. By early April, 19/60. The range landed between realistic and pessimistic, and the owners didn’t feel betrayed because the number stayed inside the stated forecast.
Let me also tell you the limits of this test: a sample of three hotels, a single sector, a single territory. It’s not a study. It’s an operational case showing that a forecast built with discipline beats any “gut-feeling” promise.
Always present the range to whoever decides on the budget, never the single number.
How you build your forecast in practice
You need very little, but you need discipline. Here are the four ingredients.
Measured baseline. Choose a fixed sample of queries (40-60 realistic queries for your sector, written the way a customer would write them, not the way you would) and count them every month across the four main AI engines. Without a constant baseline you don’t have a forecast, you have opinions.
Inventory of planned actions. Articles going out, new product pages, updated Google Business Profile listings, a tidied-up Wikidata entry, expected link earning. Each action has a date and an expected weight (high/medium/low).
Known-cycle calendar. Note the presumed cut-off windows of the main models — you don’t have official dates, but you can observe the historical cadence of the last year and a half. On engines with live retrieval (Perplexity, Gemini with grounding) you factor in 2-4 weeks of latency from publication to citability.
Three scenarios, not a single number. Pessimistic (only live retrieval reacts), realistic (retrieval + part of the new training), optimistic (both react at maximum). Always present the range to whoever decides on the budget, never the single number.
The mistakes I see most often
Linear forecast. Taking the +20% monthly observed over the last three months and projecting it out to 12 months. It doesn’t work that way: AI citations jump in line with training, not in linear growth. You hit a wall in the first quarter without a jump.
Confusing traffic and citations. The organic traffic forecast is not the AI visibility forecast. ChatGPT and Claude often don’t generate clicks — they generate brand mentions. You’re tracking two different things, and the forecasts live on two separate tracks.
Promising the single number. “By June we’ll be in Perplexity’s top 3 results for query X.” A promise that specific burns your credibility: the AI doesn’t work with stable positions the way Google does. Promising share-of-voice ranges is honest and defensible.
Ignoring competitors. Your forecast is incomplete if you don’t look at what the 3-5 brands the AI cites today in your sector are doing. If they’re publishing three times your content, the “optimistic” scenario becomes unrealistic.
Operational audit before promising the owner anything
Two steps, half a day’s work.
First: take 40 realistic queries for your sector (written by a customer, not by you) and measure them today across the four engines. How many cite your brand? Under 10% you’re in the building phase, between 10% and 30% you’re in the consolidation phase, above 30% you’re in the maintenance phase. The phase determines what kind of forecast is realistic.
Second: compare your editorial output over the last 90 days with that of the 3-5 competitors the AI cites for your key queries. If they publish 3x your content on relevant queries, your “optimistic” can’t predict overtaking them in 6 months. Adjust the expectations of whoever signs off on the budget now, not later.
These are entry-level checks: serious analysis requires professional tracking tools and far larger query datasets, but to get out of “promising in the dark” they’re more than enough.
Where the forecast takes you in the AI visibility thread
The forecast isn’t an analyst’s exercise: it’s the document that turns your AI visibility measurement into a strategic conversation with whoever holds the wallet. If in this series you’ve learned to measure share of voice, citation count and brand mention rate, the forecast is the piece that puts them in a time sequence and ties them to the budget.
In the previous articles I talked about the weight of backlinks as a citation proxy and the implicit reference weight: both enter the forecast as weighting variables for the planned actions. Having your Google Knowledge Graph entry sorted out, for example, carries a “high” expected weight in the realistic scenario.
In the next articles in the series I’ll take you into longitudinal citation tracking and revenue attribution to AI mentions: two pieces that close the loop between forecast, measurement and a ROI you can defend in front of whoever funds the project.
