Why does AI make up details like my headcount or founding year?

A language model writes the answer that sounds most likely. It does not check a company register first. When someone asks how many people work at your company and nothing the model can see states the number, it still produces one. The answer reads as confident because the model writes everything that way.

Headcount and founding year get invented often for three reasons:

• Many company sites never state them as plain text. The number sits in an image, a careers widget, or nowhere.
• Third-party pages often disagree with each other, or show a figure from years ago.
• The question looks simple, so the model answers instead of declining.

Often the error points to a gap: no reachable source says the thing. Sometimes it points to a wrong source, or to another company with a similar name (brand conflation). The same pattern drives invented funding, customers or awards. The next move is to find which stage the gap sits in.

An AI answer passes through four stages. A missing fact can fall out at any of them.

• Parametric. With search off, the model answers from training data frozen at a cutoff. OpenAI's model documentation lists a Feb 16, 2026 cutoff for the GPT-5.6 family and Aug 31, 2025 for the Instant path used in ChatGPT. A fact you published after the cutoff doesn't exist for that answer. Nobody controls this stage.
• Access. A crawler has to fetch your page. Vercel found that ChatGPT and Claude crawlers download JavaScript files but don't execute them. If your founding year renders only after JavaScript runs, those crawlers never see it. robots.txt rules are per bot: OpenAI says OAI-SearchBot controls ChatGPT search, while GPTBot covers training only. A firewall or CDN can also block a crawler that robots.txt allows.
• Retrieval. As of October 2026, searching is discretionary per query. ChatGPT decides when to search, and Claude.ai search is a user toggle. Engines also rewrite the question into their own queries, which you can't see.
• Generation. Even with your page retrieved, the model picks what to use. OpenAI's API docs note that retrieved sources often outnumber cited ones.

You control access fully, and your own page content. You don't control retrieval or generation.

Stating a specific fact (a number, a date, a one-line definition) on a fetchable page is the only positive experimental signal for on-page content in our evidence base. A 252,000-trial factorial study (arXiv 2605.25517) found that including explicit price information and a recent timestamp helped consistently. That study tested documents already placed in the model's context.

That limit matters. A critical survey of 45 GEO studies (arXiv 2607.14035) concluded that already-retrieved content can change how it gets used, but no technique shows a stable, cross-platform effect on being discovered. So a clear sentence helps once your page is in the answer's context. It does not pull the page in.

What doesn't hold up:

• Schema markup. Google says structured data isn't required for generative AI search. Ahrefs' matched test found AI Overview citations fell 4.6% after pages added schema. Existing markup does no harm, but it isn't a lever.
• llms.txt. Google says Search ignores it, and no engine documents reading it.
• AI-friendly style or chunking. Google rejects both and says there's no ideal page length.

Two kinds of error look the same in a screenshot. They need different fixes.

Research metrics don't make this split. FActScore's not-supported class covers claims the source contradicts or lacks evidence for, in one bucket. For a brand, splitting them is a practical choice: you can't call a number false in public when nothing public contradicts it.

Also separate hallucinated from outdated. An old headcount from a 2023 press release is a stale source, not an invention (how to tell them apart).

For each error, record:

1. Date and platform.
2. The exact claim, quoted.
3. Whether the answer showed citations. Without them, you can't tell if search ran.
4. Wrong, unverifiable or outdated.
5. The cited source, if any.

1. Write each fact as one plain sentence. Say you run a hypothetical firm called Acme: "Acme was founded in 2017 in Leeds and has 42 employees as of October 2026." Date the headcount so it doesn't read as current forever.
2. Put it in the raw HTML. View the page source with JavaScript off. If the sentence isn't there, ChatGPT and Claude crawlers may not see it.
3. Check crawler access. Confirm robots.txt lets OAI-SearchBot, PerplexityBot and Claude-SearchBot in, and that your firewall doesn't return 403 to them. The AI Crawler Checker requests a page as each crawler and shows which ones get turned away.
4. Make your own site agree with itself. About page, footer, careers page and press kit should state the same founding year and headcount.
5. Fix third-party sources. When engines cite directories or profiles instead of you, getting named correctly there is associated with appearing correctly in answers. That is an association, not a promise.
6. Re-check per platform. The free checker on this site shows what engines say for up to three buyer questions.

Search-off answers won't change until a model with a later cutoff ships. If the error repeats across engines, see why the same wrong fact shows up on several platforms and the guide to ChatGPT having wrong information about your business.