LLM brand safety: how to monitor what AI says about your brand

In advertising, brand safety means keeping your ads away from harmful content. With AI assistants the risk is different: the engine itself describes your brand, in its own words, to someone deciding whether to buy. You don't control the answer, and you often don't see it.

The IAB's August 2026 framework for measuring visibility in AI includes portrayal as one of four measures, alongside presence, prominence and persuasion, and asks for a hallucination rate as part of it.^[1] Gartner's 2026 market guide names hallucination detection as a core use case for these tools.^[2] In practice, brand safety covers five kinds of problem:

• Wrong facts: prices, plans, opening hours, policies, ingredients, locations or leadership that are simply incorrect.
• Outdated facts: a product you retired, a price from two years ago, a policy you changed.
• Unfair framing: a complaint from years ago presented as current, or a negative comparison that cites a weak source.
• Substitution: the engine recommends a competitor for a question your brand clearly answers.
• Risky adjacency: your brand named in answers about safety, legal or health questions where accuracy matters most.

When AI gets a brand wrong, the brand usually carries the consequences. In Moffatt v. Air Canada, a Canadian tribunal held the airline liable after its own website chatbot gave a customer wrong information about a bereavement fare.^[3] That case was about a company's own bot, but it shows how customers and regulators treat AI-generated statements about a business.

Third-party engines make mistakes too. In May 2024 Google explained that AI Overviews had shown odd and wrong answers, including some drawn from satire and user forums, and described more than a dozen changes it made in response.^[4] Engines improve, but they still synthesize answers from sources of mixed quality, and each answer is generated fresh.

1. Stale training data. Model knowledge has a cutoff. Without live retrieval, an engine may describe your brand as it was.
2. Weak or conflicting sources. Retrieval-based answers cite the pages they find. An old review, a scraped directory or a forum thread can outrank your own page for a narrow question.
3. Your own pages disagree. If your pricing page, help center and press releases give different numbers, the engine has to pick one.
4. Missing facts. If no page states the fact plainly, the engine fills the gap by inference.
5. Randomness. The same question can produce different answers. A SparkToro and Gumshoe study found less than a 1 in 100 chance that two AI answers listed the same brands.^[5]

The last point changes how you measure. One screenshot of a bad answer tells you a problem exists. It doesn't tell you how often buyers see it. Our guide on why AI answers change explains the sampling side in more depth.

1. Write down the facts that matter

List the facts an engine must get right: product names, prices or price ranges, availability by country, key policies, safety and compliance claims, leadership and company details. Publish them on one canonical page, as we do on our own brand facts page, and keep it current. This is your reference for every check.

2. Build a prompt set that includes risk questions

Mix the questions buyers ask ("best {category} for…", "is {brand} worth it") with risk questions: pricing, refunds, safety, complaints, comparisons with named competitors and questions in each market you sell in. Ask in each language natively rather than translating.

3. Sample repeatedly and count

Ask each question several times, on each engine, over several weeks. Report how often each wrong claim appears, with the number of answers behind it. A claim in 2 of 40 answers is a different problem from one in 30 of 40.

4. Grade severity and assign owners

5. Re-check the same questions after each fix

Changes take time to reach answers, and a single clean answer can be luck. Re-run the affected questions and compare the rate before and after, with ranges, before you call it fixed.

There's no button that changes what an engine says about you. Google, for example, says there is no special markup or file needed to appear in its AI features, and that the usual Search fundamentals apply.^[6] What works is changing the sources engines rely on:

• Correct your own pages first, so every page states the same fact in plain text, close to the top.
• Find the cited source. Open the answer and its citations. If a third-party page is wrong, ask its owner to update it, or publish a clearer, better-linked page of your own.
• Update listings and profiles on directories, marketplaces and review sites engines cite for your category.
• Use each engine's feedback tools for answers that are clearly wrong. It rarely fixes a pattern on its own, but it creates a record.
• Keep crawler access open to the pages that state your facts. Our AI crawler checker shows whether major AI bots can read them.

Several AI visibility vendors launched accuracy features in 2026. Bluefish sells AI Accuracy, which checks claims against a store of brand facts,^[7] and Profound includes a FactCheck accuracy score in its answer tracking.^[8] Whichever you consider, ask how many answers each accuracy figure is based on, and whether it re-checks after a fix.

Glossy Data records every answer and the sources behind it, extracts the claims each engine makes about you, and flags changes only when they're bigger than normal variation. Automated wrong-facts checks are coming soon, not live yet. See AI brand monitoring for enterprise brands.