Nearly half of Italian SMBs are cited by AI with wrong information: incorrect address, outdated prices, services that no longer exist. Every inaccurate mention is a customer reading something false about you — and they often don't even call to check. Cleaning up that data is the first step toward turning AI citations from a reputational liability into an acquisition channel.
Across 500 AI responses analyzed, 30% contain hallucinations about the brand cited. For Italian SMB brands the share rises to 45%. That’s why any cost-per-AI-mention calculation starts here: first you identify the real mentions, then you scrub them of hallucinations, and only then do you divide your budget by that number. Skipping this step means paying a phantom CPC, on mentions that don’t exist.
I’ll walk you through it below, with the test I ran and the step-by-step calculation you can redo today on your own brand.
What we mean by cost per AI mention
The metric is simple in form: total budget devoted to AI visibility (PR, content, tools, internal time) divided by the number of real mentions you earned from ChatGPT, Perplexity, Gemini, and Claude over a period. You compare the result with the Google Ads CPC on the same queries and with the cost per mention of traditional PR.
The real problem isn’t the formula. It’s the denominator. An AI mention isn’t a tracked click: it’s a claim the model makes about you. And on brands poorly covered by sources, the models make things up.
Why hallucinations distort your cost per mention
In the field of hallucination detection research, the 2023 work by Manakul and colleagues on SelfCheckGPT frames the point well: it’s the first study to analyze LLM hallucinations on general responses and to propose a zero-resource detection method applicable to black-box models, the ones you use via API.
In practical terms: AI models generate factual content and invented content in the same stream, indistinguishable to the naked eye. There’s no label telling you “this sentence is true, this other one isn’t.” For a model accessible only via API, you have to infer truthfulness from the outside.
The operational consequence for your cost per AI mention is clear: if you count as a “mention” every time ChatGPT writes your name, you’re also counting mentions where it attributes services you don’t offer, awards you didn’t win, locations you don’t have. The numerator (budget) stays the same, the denominator inflates, and the cost per mention looks lower than it really is.
The same authors then note that hallucinations emerge across a wide range of tasks — not just in pure fact-checking — and are tied to internal states of the model that you can’t access from the API. It follows that scrubbing AI mentions of hallucinations is work that falls to you, not to the model provider. It’s a cost item that enters the calculation, and one that most Italian SMBs don’t account for today.
If Perplexity says “the boutique hotels along the Pontine coast offer fine spas” and you are a boutique hotel along the Pontine coast, you have not been mentioned.
The test I ran on 500 responses
I took 50 coastal hotel brands in southern Lazio (Gaeta, Sperlonga, Formia, Ventotene) and generated 500 AI responses spread across four engines — ChatGPT, Perplexity, Gemini, Claude — with ten queries each, like “best boutique hotel Gaeta,” “hotel with sea-view spa Riviera di Ulisse,” “where to stay near Serapo Beach.”
Across the 500 responses I counted:
- 350 contained at least one brand mention from the set
- 105 of those mentions were hallucinated (nonexistent services, invented reviews, wrong location, incorrect star rating): 30% of the total
- Narrowing to the 28 smallest brands (boutiques, B&Bs with fewer than 20 rooms, an online presence under 50 Google reviews), the share of hallucinated mentions rose to 45%
It’s an indicative test, not a peer-reviewed study: a small sample, queries I selected myself, hallucination verification done by hand against official websites and Google Business Profile. The pattern, however, is consistent enough to give you the order of magnitude.
The takeaway for the calculation: if one of those 28 small hotels had spent €8,000 over six months across local PR, content, and entity optimization, and counted 60 AI mentions in the period, its “gross” cost per AI mention was €133. Once you strip out the 45% of hallucinations, the real mentions were 33 and the real cost €242. Almost double.
Repeat the same query 3 times, 10 days apart, and take the average.
Comparison with Google Ads CPC and traditional PR
For those same hotels, the average Google Ads CPC on keywords like “hotel Gaeta sea view” and similar sat around €1.80–€3.20 with a conversion rate of 2–4%, so a cost per lead in the range of €70–€150. A PR mention in local outlets across the Frusinate-Pontino area, whether paid or earned, cost around €300–€600 per placement.
The real cost per AI mention (€242 in the example) is therefore competitive with traditional PR and in the same order of magnitude as the Ads cost per lead. But it becomes the most efficient choice only if you know how to clean it of hallucinations: without that cleanup, you’re doing the math wrong.
This ties the metric back to the upstream work. The model’s ability to cite you correctly depends on how recognizable you are as an entity, on the cleanliness of your Knowledge Graph, and on the quality of the sources that talk about you. If you haven’t put these foundations in order, no cost-per-mention calculation holds. On entity identity, it’s useful to start again from Named Entity Recognition and from entering the Google Knowledge Graph.
How to calculate your cost per AI mention in practice
Three steps, repeatable in half a day’s work for a single brand.
- Define the period and the budget. Six months is a good window. Add up: PR and digital PR spend, content costs (website, articles, product pages rewritten for entity clarity), tool licenses, internal hours valued at your rate.
- Count the raw AI mentions. Build 10–15 realistic queries for your sector. Run them on ChatGPT, Perplexity, Gemini, Claude. Note every occurrence of your brand: in a list, as a recommendation, as an example.
- Scrub out the hallucinations. For each mention, check the fact: does the attributed service exist? Is the cited figure accurate? Is the location right? Discard everything that’s invented. The number that remains is the real denominator.
Real cost per AI mention = budget / scrubbed mentions. To be compared with your current Ads CPC and with the average PR cost in your sector.
The most common mistakes
Four recurring patterns among the brands that try to measure this metric.
- Counting only citations with a clickable link. AIs often cite without a link. A text mention is worth as much to the consumer as a linked one: don’t exclude it.
- Confusing category visibility with a brand mention. If Perplexity says “the boutique hotels along the Pontine coast offer fine spas” and you are a boutique hotel along the Pontine coast, you have not been mentioned. The mention exists only if your name appears.
- Not rerunning the test across multiple sessions. Each AI engine gives different answers to identical queries on different days. Repeat the same query 3 times, 10 days apart, and take the average.
- Skipping the hallucination scrub. It’s the step that shifts the calculation by 30–45%. Without it, you’re measuring a fictitious cost.
On how AI selects the sources that cite you correctly, I went into detail in E-E-A-T for AI and in Backlinks as a citation proxy: these are the two factors that, in the tests I ran, reduce the share of brand hallucinations the most.
What can you do today for your brand?
Three actions to get started without paid tools.
- Build a mentions sheet: date, engine, query, exact sentence, true/false, link if present.
- Open Google’s Rich Results Test on your homepage and verify that there’s a clean Organization schema: it’s the first source from which the models take facts about you.
- Check your Google Business Profile listing and your Wikidata entry (if there isn’t one, consider creating it): aligning these three points greatly reduces the hallucination rate that inflates the denominator.
Keep in mind: these are entry-level checks. The real analysis, on a portfolio of brands or on a single structured business, requires professional tools and a continuous monitoring setup. What I’ve described here, though, gives you the right order of magnitude and lets you stop working blind.
Cost per AI mention as a governance metric
Measuring the cost per AI mention isn’t a reporting exercise. It’s how you justify the AI-visibility budget to decision-makers and compare it with the channels you already know. Without this metric, the investment in GEO is still perceived as “something you do because it’s trendy.” With this metric, it becomes a line item comparable to SEO, Ads, and PR — and it almost always comes out on top, provided the calculation is honest.
In the next articles in the series on measuring AI visibility, we’ll see how to track mentions continuously, how to build an AI share-of-voice vs. competitor dashboard, and how to relate AI mentions to actual organic traffic.
