How often do AI search engines get facts wrong?

AI search engines get a large share of facts wrong in the studies that exist. Those studies tested news and research tasks, not brands.

Two figures get quoted most:

• Tow Center (Columbia Journalism Review): over 60% of news queries answered incorrectly across the AI search tools it tested.
• PwC: deep-research agents scored between 39 and 77% factual accuracy.

Both numbers are real. Neither tells you how often ChatGPT, Gemini or Perplexity misstate your pricing, founding year or product names. Nobody has published that census for commercial brands.

So the honest answer has three parts. Errors are common. The rate depends heavily on what you ask and where. And the rate that matters to you is the one for your own brand, which you have to check yourself. If you want a sense of the usual failure points, see which brand facts AI engines get wrong most often.

Tow Center: news citation

The Tow Center asked AI search tools to identify news articles and checked the answers. Over 60% were wrong. The spread between tools was wide: Perplexity's error rate was 37%, and Grok 3's was 94%. That spread is the first lesson. A single "AI accuracy rate" hides a near threefold gap between platforms on the same task. For more on platform differences, read whether some AI platforms are more accurate about brands.

PwC: deep research

PwC tested deep-research agents, which run long multi-step searches. Factual accuracy landed between 39 and 77%. Link validity was above 94%. In other words, the citations resolved and looked relevant, but they often did not support the claim attached to them.

What neither covers

Neither study tested brand facts. News identification and long research reports are different tasks from answering "how much does this product cost". Treat both figures as evidence that errors are common, not as your error rate.

Several things move the rate, and most are outside your control.

• The task. Identifying a source article, writing a research report and stating a price are different jobs. Each has its own failure modes.
• The platform. Google says AI Overviews and AI Mode "may use different models and techniques", and they return different links. Two surfaces from one company can disagree.
• Whether the engine searched. Retrieval fires at the model's discretion on most platforms as of October 2026. Published estimates of how often ChatGPT searches range from 18% to 80.5%, depending on who measured and how. Without search, the model answers from training data frozen at a cutoff. For OpenAI's GPT-5.6 family that cutoff is February 16, 2026. The Instant path carries August 31, 2025. Read whether web search makes answers about your brand more accurate for the trade-offs.
• Query rewriting. ChatGPT search rewrites your prompt into one or more targeted queries before it searches. The engine may not search for what you typed.
• How errors are counted. Benchmarks such as FActScore divide supported facts by the total facts in a response. Others count responses with at least one error. The same answers can produce very different percentages under each method.

Unverifiable is not the same as false

An AI may state your headcount when you have never published one. A checker can flag that claim, but it cannot prove it wrong. FActScore's "not supported" class lumps contradicted claims together with claims that simply lack evidence. Before you act on a flag, read whether every claim a tool flags is actually false.

A citation is not proof

PwC's finding shows links can resolve while the claim is still wrong. Research on how models cite also shows citation behavior can be changed without changing the sources. A cited page is not necessarily the page that shaped the answer.

Refusals and errors

A platform that refuses or fails to answer has not answered correctly. Vectara's hallucination leaderboard reports an answer rate separately for this reason, so a high refusal rate does not pass for accuracy.

Small samples

One wrong answer out of two looks like a collapse. It is a partial run. Judge accuracy across enough answers on each platform.

You cannot control what a model says. You can control what it finds when it looks.

1. Check your brand directly, platform by platform. Ask the buyer questions your customers ask and record each answer separately. The free checker on this site runs up to three buyer questions and gives you a snapshot.
2. Write down what is wrong and where. Note the fact, the platform, and whether the answer showed citations. Citations suggest it searched. No citations means you cannot tell.
3. Fix contradictions on your own site first. If your pricing page and footer disagree, an engine has two answers to choose from. The free check for contradicting facts on your own site finds these.
4. State specific facts plainly in fetchable HTML. A price, a number or a one-line definition gives an engine a sentence it can quote. Make sure it appears before JavaScript runs, since most AI crawlers do not execute it.
5. Trace third-party sources. If the wrong fact comes from another site, find the source of the wrong AI fact and correct it there.

For the wider picture of causes and responses, see the guide to AI hallucinations about brands.

Is there an industry-standard AI hallucination rate? No. Published rates come from specific tasks, such as summarizing documents, identifying news articles or writing research reports. Each uses its own method for counting errors. Some divide supported facts by total facts, others count responses with any error. A rate from one benchmark does not transfer to another task, and none of the well-known benchmarks measure facts about commercial brands.

Does ChatGPT get facts wrong more than other AI search tools? The evidence here does not rank ChatGPT against the others on brand facts. The Tow Center's news study found wide gaps between tools, from 37% wrong for Perplexity to 94% for Grok 3. That shows platforms differ a lot on the same task. It does not tell you which platform is most accurate about your brand. Check each one separately.

Can I make AI engines stop getting my facts wrong? You cannot guarantee it, because the models decide what they say. You can remove the causes in your control. Make your key facts consistent across your site, state them plainly in HTML that loads without JavaScript, and make sure AI search crawlers can reach your pages. Then correct wrong facts on third-party sites the engines cite.

Why does the same question get a correct answer one day and a wrong one the next? Several things change between runs. The engine may search the web one time and answer from training data the next. Your request may be routed to a different model with a different knowledge cutoff. The engine may rewrite your question into different search queries. Logged-in memory can also shape the query. Test logged out and repeat questions before drawing conclusions.