Before you fact-check AI: what it actually says about brands

Recommended is not the top of the ladder

We measure AI visibility as a ladder of five steps: AI reads your site, cites your pages, mentions your brand, recommends you, and describes you correctly. Most tools, and most marketing teams, stop counting at step four. If AI recommends you, the job looks done.

1. 4
   Recommended

   AI puts you first, or presents you as the best fit.

2. 5
   Described correctly

   What AI says about you is true: the right prices, products, markets and strengths.

But a recommendation built on wrong facts can cost more than no recommendation at all. The buyer arrives expecting the old price, the missing feature or the wrong delivery country, and the sale breaks at the last step. Nothing in your analytics shows why. That is why we treat "described correctly" as its own step, with its own measurement.

What we saw in our first analysis

Every answer an AI model gives about a brand contains small factual statements: "costs X per month", "is available in Y", "integrates with Z". We call these claims. Claims with the same wording count once.

4,010AI answers

754answers that named the brand

31,447cited sources

4 AI models

6 markets4 languages

31 daysof answers

Research range7 Aug 2026 – 19 Sep 2026

We collected 4,010 answers from ChatGPT, Gemini, Perplexity and Google AI Overviews, to buyer questions in six markets: adventure travel, architecture marketing, e-commerce, CRM software, online therapy and AI tools. Together, those answers cited 31,447 sources. AI mentioned the brand in question in 754 of the 4,010 answers (18.8%). That average hides a wide spread: in two markets, AI named the brand in more than half of the answers. In the other four, it named the brand in about 2% to 11% of them. Claims can only come from answers that name the brand, so the claim analysis uses a subset of them:

The analysed answers came from ChatGPT, Gemini and Google AI Overviews, for buyer questions in four of the six markets: CRM software, e-commerce, adventure travel and architecture marketing. On average, one answer contained 3.8 claims about the brand (2,166 claim mentions in 567 answers). Depending on the market, AI made between 48 and 985 distinct claims about the brand.

A claim mention is one claim in one answer. The same claim can appear in several answers, so there are more mentions (2,166) than distinct claims (2,084). 13 claims appear in two AI models, so the AI model rows add up to 2,097 distinct claims, not 2,084.

Few marketing teams could list everything AI says about them. Almost none have checked it.

Three verdicts, not two

To check a claim, you need something to check it against: your own facts. What you sell, for whom, where, and at what price. Every claim is then compared with those facts and gets one of three verdicts:

The third verdict matters. A simple true/false check forces every claim into one of two boxes. In practice, much of what AI says about a brand is neither confirmed nor denied by anything the brand has written down.

The biggest gap: what you never wrote down

When a brand does not state something clearly, AI does not stay silent. It fills the gap from reviews, old pages, comparison articles written by competitors, or its own training. The "not covered" claim in the example may even be true. But the brand did not choose it, and cannot see it unless it looks.

This changes the job. Fixing what AI gets wrong is only half of it. The other half is stating the facts AI keeps guessing about, in plain words, on pages AI can find.

Praised, and wrong

AI models are mostly kind to brands. In our research, 68.8% of the 2,084 claims were positive, 27.5% neutral and only 3.7% negative. That looks like good news in a sentiment report. But a positive claim is not a correct claim:

Sentiment tells you how AI feels about you. It does not tell you whether AI is right. You need both, and you need them per claim, not as one average score.

Each AI model tells its own story

To compare AI models fairly, we grouped the 2,084 claims by meaning into 157 topics, such as a brand's pricing, its ease of use or its delivery area. Then we counted which AI models raised each topic:

About a third of the topics came up in all three AI models. But 41.4% came up in only one. A buyer who asks ChatGPT and a buyer who asks Gemini can hear about different parts of the same brand.

The sources differ even more. The answers behind the claims cited 952 different websites. 83.8% of those websites were cited by only one AI model, 11.8% by two, and just 4.4% (42 websites) by all three. Each AI model builds its picture of your brand from its own set of pages.

A note on these numbers. An automated grouping step sorts the claims into topics, so the borders between topics are not exact. Two AI models that raise the same topic can still say different things about it. And a website cited by one AI model only may still be known to the others.

This matters for accuracy, because one AI model may quote the current price while another repeats the old one. When that happens, the error is not "AI in general". It usually sits in the specific pages one AI model relies on, or in what that model learned in training. That makes it something you can trace, by checking each AI model separately and looking at the sources its answers cite.

Live page or memory?

AI answers draw on two layers: what the AI model finds on the web right now (Live), and what the model learned in training (Memory). Where a wrong claim comes from decides how you fix it. The answer behind the claim gives the first clue: did it cite sources or not?

In our research, 78% of the answers behind the claims cited at least one source. Together they cited 6,504 sources: 3,785 different pages on 952 different websites. The share differed strongly by AI model:

For most claims there is a page you can trace. For Gemini, almost half of the answers cited nothing, so its claims are harder to trace to a page.

A note on this split. An answer without cited sources does not prove that the AI model used no web pages. AI models also differ in how they show sources, and ChatGPT answers in this sample were collected with web search on. Treat the two cases as directions to investigate, not as exact causes, and do not read the table as a ranking of AI models.

What to do about it

We use the same four steps for accuracy as for every other part of AI visibility:

1. 1
   See

   Write down your facts: products, prices, markets, who you serve. Then collect what AI actually claims about you, across AI models, from real answers.

2. 2
   Understand

   Sort every claim: supported, contradicted, or not covered. Check which AI model says it, and which sources the answer used.

3. 3
   Act

   Fix contradicted claims at their source: your own pages first, then the third-party pages the answers cite. Turn frequent "not covered" claims into clear, findable facts.

4. 4
   Measure

   Keep checking. A fix to a live page can show up quickly. A memory gap closes slowly, over model updates.

Vosapio runs these four steps for you: it collects the answers, extracts the claims, checks them against the facts you enter, and keeps the sources behind every answer. See how Vosapio tracks AI visibility, or compare the plans for brands and the plans for agencies.

For agencies

Run the same check for every client. Show each client which claims AI makes about them, which AI model says them and which sources sit behind them. It is a fast way to open a client conversation with data instead of opinions, and to agree on which pages to fix first.

Method and limits

• Full dataset. 4,010 answers from ChatGPT, Gemini, Perplexity and Google AI Overviews to 297 tracked buyer questions in six markets and four languages (English, German, French and Romanian), collected between 7 August and 19 September 2026 (on 31 days). The answers cited 31,447 sources.
• The 19% is an average, not a benchmark. It pools six markets of very different sizes, and the largest market had one of the lowest rates. Whether a mention rate is high or low depends on the market, the brand and the questions asked.
• Claim analysis. 567 answers from ChatGPT, Gemini and Google AI Overviews to 79 questions, in four of the six markets and four languages. Claims were extracted automatically between 24 September and 10 October 2026. All claim, sentiment and source figures in this article come from these 567 answers.
• How the 567 answers were selected. Only answers that mention the brand can contain claims about it: 754 of the 4,010 answers did. For each brand, the analysis takes the newest answers that mention it and have not been analysed yet, and stops after 500 answers or 500 claims per run. The 567 are the answers that produced at least one claim. This is not a random sample: it leans toward the newest answers and toward the brands that AI mentions most.
• Small sample. Four markets and three AI models in the claim analysis. The numbers describe this sample. They are not an estimate for all brands or all AI models.
• Grouping. The claim totals count claims as the same only when their wording matches, so they are higher than a count by meaning would be. For the AI model comparison, an automated step also grouped all 2,084 claims by meaning into 157 topics.
• Automated extraction. A model extracts claims and their tone from each answer. Some claims will be split, merged or labeled wrongly.
• No accuracy rates yet. The brands had not entered their own facts, so no claim has a verdict. We publish no percentages of right or wrong claims, and we will update this article when we have them.
• Facts come from the brand. A claim can only be checked against what a brand writes down. A brand with few facts gets more "not covered" verdicts.
• No brand names. Our figures are aggregated. No brand behind them is named.