Do AI agents make up facts about my business, or just leave them out?

Is a wrong fact or a missing fact the bigger risk?

A missing fact is the more common risk; outright wrong facts are comparatively rare.

The clearest evidence comes from Finder and colleagues (opens in a new tab). They ran 37,927 agent journeys, each a buyer question about pricing, features or setup, across 1,056 real businesses and four AI agent setups. They then graded answers against facts captured from each business’s own site.

Fact by fact, the shift was toward silence, not invention:

The authors sum it up: poor readability “makes a fact unretrievable, not false.” Site-built answers got 48.3% of asked facts right, against 34.3% for web-built ones about the same business and question.

Read this with care. The authors work for ora, which sells the agent-readiness score the study uses, and they say so. Accuracy was graded on only 131 businesses. On that subset, the difference between the two groups of businesses was not statistically reliable; the gap appears when comparing answers about the same business.

What happens when an agent cannot read your site?

It answers anyway, using other sources, and the answer gets thinner and more hedged.

In about 99% of journeys that hit a dead end on a business’s site, the agent still answered, built from whatever it found elsewhere. Only 56% of runs on hard-to-read sites ended grounded in the business’s own pages, against 78% on readable sites.

The answers also hedged more. Agents said they could not access the business 4.4 times more often when the site was hard to read. Training memory did not fill the gap either: it supplied only 7% to 10% of the finished answer whether or not the site was readable.

So the typical failure is not an agent confidently inventing a price. It is an agent saying it could not find the price, or giving a general answer with the specifics missing.

How often do AI answers state something false outright?

Rarely in the large audits, though how often depends on the engine, the topic and how “false” is measured.

Xu, Iqbal and Montgomery (opens in a new tab) checked 98,020 claims from Google’s AI Overviews (the AI summary at the top of Google’s results) against the pages each Overview cited, over 40 days in spring 2026. Only 2.66% were directly contradicted by a cited page. Another 6.98% were not mentioned by any cited page at all.

That second group matters for this question. A claim no cited page mentions could be made up, or could come from a page the Overview did not cite. The study cannot tell which. Even so, outright contradiction was the smallest failure.

Earlier answer engines did worse. In a 2024 audit, You.com and Perplexity each had around 30% of their statements unsupported by the sources they listed (Narayanan Venkit and colleagues, 2024 (opens in a new tab)). Unsupported is not the same as false, but it shows the risk was higher in older systems.

When an AI gets a business fact wrong, where does it come from?

Usually from a real source, often one the business itself published, rather than from thin air.

Our own studies traced differing facts back to their likely sources:

• Software prices. In our pricing study, 64 quoted prices differed from the official pricing page. For 39 of them the same figure was on another page of the vendor’s own site, and for 21 more on a third-party page cited for the product. Only 4 were found nowhere.
• Phone numbers. In our business facts study, 55 phone numbers differed from the business’s Google profile. 42 of them were on the business’s own website. Only 3 of the 633 numbers given could not be found on the profile, the website or any page the answer cited.

In both cases the wrong-looking fact was mostly an old or alternative fact still live on the web. An assistant that finds two official answers can quote either.

Which facts go missing most often?

The research points to setup documentation, full price lists and opening hours.

• Setup details. In the agent study, 65% of asked setup facts went unmentioned whichever way the answer was built, because setup lives in documentation rather than on marketing pages.
• Full price lists. In our pricing study, answers named a price for 84.2% of the paid plans on the official page on average. The rest were simply left out.
• Opening hours. In our business facts study, 9.4% of answers gave Monday hours that differed from the Google profile, and 12.0% gave none.

Notice the pattern in that last case. For hours, leaving the fact out was about as common as stating a different one. Omission is the main risk, not the only one.

What should you do about it?

Make your key facts easy for an agent to fetch, and remove the old versions that compete with them.

1. Put prices, plans and core facts in plain text on pages agents can load. Agents often cannot see content that only appears after a page’s scripts run, or pages behind bot blocks. See whether agents recommend readable websites for why this matters beyond accuracy.
2. Retire or update old pages when facts change. Most differing prices and phone numbers in our studies were still published somewhere by the business itself.
3. Publish one version of each fact everywhere. Use the same phone number, hours and plan names on your site, your Google profile and directories. How much those outside listings count once an agent is reading about you is covered in off-site mentions versus site readability.
4. Surface setup and documentation facts. Agents rarely reached them, so link them clearly from pages agents do read.
5. Ask assistants your buyers’ questions regularly. Check what is missing, not just what is wrong.

If you want help making your site readable to AI agents, see our generative engine optimization service.

What does the research not tell us yet?

The research shows omission dominates in the settings tested, but those settings are narrow.

• The main agent study comes from a company that sells agent-readiness scoring, uses its own score as the measure, and graded accuracy on only 131 businesses.
• That study did not match businesses on how much content they publish, so thinner sites may explain part of the gap.
• Its sample leans toward software and online commerce, in English. Local services and other languages are untested.
• The AI Overview audit cannot tell whether an unsupported claim was invented or taken from an uncited page.
• No study yet measures how often buyers act on a missing fact versus a wrong one, so the business cost of each is unknown.

Frequently asked questions

Do AI assistants make up prices?

Rarely, in our tests. Of 64 software prices that differed from the official page, only 4 could not be found on any source we could fetch; most were old prices still on the vendor’s own site.

Why does ChatGPT leave out information about my company?

Often because it could not read your site. In one large agent study, only 56% of runs on hard-to-read sites ended grounded in the business’s own pages, against 78% on readable sites.

Is a missing fact really less harmful than a wrong one?

Not necessarily, but it is more common. Answers built off-site were 3.7 times more likely to contain none of the facts the buyer asked for, which can cost a recommendation.

How accurate are Google’s AI Overviews about facts?

In a 40-day audit, 2.66% of claims were contradicted by a page the Overview cited, and 6.98% were not mentioned by any cited page.

Sources

• Finder, Elovic, Shalev and Yosef (2026), AX is the New AEO (opens in a new tab), arXiv:2609.34951.
• Xu, Iqbal and Montgomery (2026), Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact (opens in a new tab), arXiv:2605.14021.
• Narayanan Venkit and colleagues (2024), Search Engines in an AI Era: The False Promise of Factual and Verifiable Source-Cited Responses (opens in a new tab), arXiv:2410.22349.
• Underneath (2026), How faithfully do AI assistants quote software prices?
• Underneath (2026), Do AI answers match a business’s Google profile?