Hallucination Rate

What counts as a brand hallucination

A hallucination, in the general sense used across AI research, is any statement a language model generates that is not actually true, regardless of the topic. Hallucination Rate as used here narrows that idea to one specific, checkable case: false statements about a particular brand. That includes a price the assistant states that does not match what the brand actually charges, a feature or product line attributed to the brand that it does not offer, a founding date, headquarters location or executive name stated incorrectly, or a claim about the brand's policies, such as a return policy or certification, that is simply wrong. This is a different failure from a negative sentiment mention, where the assistant states something true but unflattering, and different from a missing citation, where the assistant makes no source claim at all. A hallucination is confidently stated and factually wrong, which is what makes it worth tracking on its own rather than folding it into sentiment or citation checks, since a false but positive-sounding claim about a brand can pass an automated sentiment check without anyone noticing the underlying fact is wrong.

How it works

Checking for brand hallucinations means comparing specific, checkable claims inside an AI-generated answer against verified facts about the brand: current pricing, actual product names, real company history and stated policies. This is closer to fact-checking than to the pattern matching used for mention or citation counts, and it works best when applied to prompts that ask about concrete, verifiable details rather than open-ended opinion questions, since a claim like "the starter plan costs $49 a month" can be checked directly against the brand's actual pricing page, while a claim like "this is a well-regarded tool" cannot be verified the same way. A team building this check typically keeps a short reference sheet of current, verified facts, pricing, product names, key policies, and checks AI-generated answers against that sheet rather than relying on memory. Because AI assistants generate a new answer for each prompt rather than retrieving one fixed response, the same false claim can appear in one run and be absent in the next, or a different false claim can appear instead, which means a hallucination check benefits from repeated runs over time in the same way other AI-visibility metrics do, rather than a single check being treated as conclusive.

Why it matters for AI visibility

A false claim stated confidently inside an AI answer can do real damage to a buyer's decision, whether it understates a price in a way that creates a bad surprise later, or invents a feature the buyer expects and does not find once they look closer. Unlike a low mention rate, which is a visibility gap, a hallucination is active misinformation reaching a buyer at the exact moment they are evaluating the brand, and it happens without the brand having any direct way to correct it in that conversation. Tracking hallucination rate is less about a single number to report and more about catching specific, recurring false claims early enough to address them, often by publishing clearer, more current information in the places an assistant is likely to draw from, such as an up-to-date pricing page or a clearly stated feature list, since assistants tend to generate fewer false claims about brands whose current, correct information is easy to find and unambiguous.

Good practices

• Keep a short, current reference sheet of verifiable facts about your brand, pricing, product names, key policies, to check AI-generated answers against.
• Focus checks on concrete, verifiable claims like pricing and product names rather than subjective opinion statements, which cannot be fact-checked the same way.
• Re-check regularly rather than once, since the same prompt can produce a different false claim, or none at all, on a different run.
• Keep pricing and product pages current and unambiguous, since assistants tend to generate fewer false claims when the correct information is easy to find.
• Log specific recurring false claims separately from a general rate, since one specific wrong fact is what a team can actually act on.
• Treat a hallucinated claim differently from a negative sentiment mention; a true but unflattering statement and a false statement need different responses.

Common mistakes

• Treating hallucination as the same problem as negative sentiment, when a false claim and an unflattering true claim call for different fixes.
• Checking only whether the tone of a mention is positive, without verifying whether the actual facts stated are correct.
• Assuming a hallucinated claim, once caught, will not appear again, when the same prompt can produce a different false statement on a later run.
• Skipping verification on concrete, checkable claims like pricing, which are the easiest kind of hallucination to confirm and the most damaging to a buyer's trust.

Related terms

• Source Attribution: whether a claim is credited to the correct page, a related but separate check from whether the claim itself is true.
• Brand Mention: the underlying event a hallucination check is applied to, since a claim has to mention the brand before it can be checked.
• Sentiment in AI Answers: a tone reading that a false but positive-sounding claim can pass even when the underlying fact is wrong.
• Citation Rate: a hallucinated claim can appear with or without a citation attached, making the two checks independent of each other.

Frequently asked questions