AI hallucination

What do wrong AI answers about a brand actually look like?

We read every answer ChatGPT and Perplexity give to "what is Attensira": 66 of them in the 30 days to 4 October 2026, read over the Attensira MCP. The answer at the top of this page is ChatGPT's of 4 October, and its only inline source is our own site. Most of it is right. One sentence is not wrong so much as old. Perplexity's answer the same morning adds two more kinds of error in one short reply:

• ChatGPT: "aimed mainly at B2B SaaS companies and developer-tool companies whose websites are managed in Git-based workflows", and "like an “SEO tool for the AI answer era”"

  What kind of error

  Stale fact: our earlier positioning, faithfully repeated

  Where the fix lives

  Our older pages and the copies of them on the web

• Described Attensira as software that tracks AI mentions and drafts fixes you approve

  What kind of error

  None: accurate as far as it goes, and linked to our pages

  Where the fix lives

  Nothing to fix

• Perplexity: cited the help centre of Attensi, a different company with a near-identical name, as its first source

  What kind of error

  Ambiguity: two real names one letter apart

  Where the fix lives

  Entity signals; see entity SEO

• Perplexity: ended by telling the user their "first question was clear enough", answering a question nobody had asked

  What kind of error

  Hallucination: an invented premise, harmless here

  Where the fix lives

  Nothing on our side; the model filled a gap

None of the stored answers invented a price or a feature. That matters, because the kinds of error above have different fixes, and treating them all as "the AI hallucinated about us" leads to the wrong one.

What is an AI hallucination?

It is a statement a model writes fluently and confidently that is false: an invented price, a feature that does not exist, a quote nobody said, a source that is not there. The model is not lying. It produces likely continuations of text, and a plausible falsehood can be just as likely as the truth.

The 2025 paper Why Language Models Hallucinate by Kalai, Nachum, Vempala and Zhang gives the incentive behind it. Models are optimised to do well on tests, and most tests score "I don't know" the same as a wrong answer. A guess sometimes scores; an abstention never does. So guessing is rewarded. The authors also argue the errors start in pretraining: when true and false statements cannot be told apart from the data, some errors follow from ordinary statistical pressure, not from a bug in one model.

How is a hallucination different from a stale fact?

A stale fact was true once. The model, or the page it retrieved, is repeating it correctly: a price you changed, a plan you retired, an audience you stopped describing. The engine found the old copy and repeated it. The fix is on your side. State the current fact on one page, remove or update the old copies, and wait for the engines to re-read.

A hallucination has no source to fix. The model filled a gap. The fix is to close the gap: make the true fact easy to retrieve and unambiguous, so the model has something better than a guess. See knowledge cutoff for why models without retrieval are frozen at an old picture of you.

Ambiguity is a third case. When two real things share a name, or nearly do, an engine may describe both, or cite the wrong one, as Perplexity did with Attensi above. That is an entity problem: give engines unambiguous signals about which thing you are, on your own site and on independent sources.

Does grounding stop hallucination?

It changes the odds, not the nature of the problem. Grounding and retrieval give the model documents to write from, and Anthropic's guidance on reducing hallucinations recommends having a model extract word-for-word quotes before working with long documents. But a model can still misread a passage, write from a passage that answers a neighbouring question, or blend two sources. The link it attaches records what it read, not that its sentence is correct.

Compare the ChatGPT answer at the top of this page with the Perplexity rows in the table. ChatGPT retrieved our own pages and wrote nothing they do not say, including a positioning line those pages, or older copies of them, once carried. Perplexity also retrieved our pages, but its first source was a lookalike company, and its last sentence came from nowhere at all. Retrieval narrowed the gap in both; it closed it in one. That is the strongest argument for making your facts retrievable and unambiguous: an engine that finds the right you does not have to guess.

What are the objections when an AI gets your brand wrong?

"Hallucinations are random, so there is nothing to do." Many brand errors are not random. They are stale copies or ambiguous names, and both have owners and fixes.

"Correct the AI directly." There is no form for that. You correct the sources it reads: your own pages and the third-party pages it cites.

"One wrong answer is a crisis." Models are not deterministic. Read several answers across days and engines before deciding a claim is persistent.

"A citation proves the sentence." It proves what was retrieved. Check the sentence against the page.

How does Attensira catch wrong claims about you?

Attensira's claims check pulls every checkable statement about your brand out of every answer, quoted verbatim with the links it carried, and checks it each night against the facts you confirmed in your brand profile. A claim counts as wrong only when it contradicts a confirmed fact with at least 80% confidence; with no confirmed fact it reads "nothing to check against", never "wrong". The same wrong sentence repeated across engines and days raises one event.

The check itself only flags. An event automation can act on the flag: trace the sentence to the page it came from and file the fix for your approval. Wrong claims are counted over the Attensira MCP today; the dashboard does not show verdicts yet. The band above is an example of the flow, not a screen. Citation tracking shows which pages each answer about you linked, which is where most fixes start, and the free audit shows which sentences engines write about you today.