Hallucination

A hallucination is a plausible-sounding statement from an AI model that is false. Google's machine learning glossary defines it as "the production of plausible-seeming but factually incorrect output by a generative AI model" that purports to make an assertion about the real world. For a brand, a hallucination might be a price you never charged, a feature you don't have or an integration that doesn't exist.

Why it happens

A 2025 paper, Why Language Models Hallucinate, argues that "the training and evaluation procedures reward guessing over acknowledging uncertainty": tests score a confident guess higher than an admission that the model doesn't know. Grounding answers in retrieved sources reduces hallucinations, according to Google Cloud, but it does not remove them. Google's own help page for AI Overviews says that "AI Overviews can and will make mistakes."

Hallucination or stale source

Before you call a wrong answer a hallucination, open the pages it cites. If a cited review or comparison says the wrong thing, the engine repeated a source, and the fix is that page: ask for a correction and state the facts plainly on your own site. If no cited page says it, it is more likely a hallucination. The fix then is to make the true facts easy to find and consistent across your site and profiles, so engines have a clear source to ground their answers in.

How to track it

Ask each engine a few factual questions about your brand on a schedule, such as price, plans and integrations. Log every wrong statement with the engine, the prompt and the cited pages, and note whether it repeats. Work first on the errors that repeat across runs and engines, since those reach the most buyers. A wrong statement can also pull down AI sentiment, so read the two together.