Hallucinated claims about your brand
Hallucinated claims about your brand are false statements AI platforms make about you, like a wrong price, a feature you don't have or a closed shop.
Hallucinated claims about your brand are false things AI platforms say about you: a price you never charged, a product you don't sell, a policy you don't have, an owner who isn't yours. They're stated as plainly as the true things around them, so the person reading has no way to tell.
They're worth watching because they hit people who were already interested in you. Someone asking about your returns policy is close to buying. A wrong answer there can lose the sale.
What they look like
For example, say an answer tells someone that Harrow & Pine's espresso machines come with a one-year warranty, when the warranty is three years. Or that Harrow & Pine doesn't offer servicing outside London, when it services across the UK. Neither claim is outrageous, which is what makes them hard to spot.
Some come from the model's memory, with no page behind them. Others come from a real page that's wrong or out of date, like an old review or a listing on a shop that stocks you.
Why you need to keep checking
The same question can get a different answer each time, so a wrong claim might turn up in a few answers a week and not in the rest. Checking once by hand will miss it. You need to read many answers over time and compare each claim with what's true.
What to do about one
First, find where it comes from. If a page says it, fix your own page or ask the site that published it to correct theirs. If no page says it, make sure the right fact is easy to find on your own site, in plain words. And if the platform keeps repeating it, report it to the platform.
In AnswerOrbit
Add your facts in Claims, up to ten checked at once. Every day, we check the claims the answers make about you and mark each one supported, contradicted or not covered. We only mark a claim as contradicted when we can quote the sentence that says it. Inaccuracy drivers shows the pages behind a wrong claim, or the platform's own knowledge. When it's the platform itself, Report to asks for an evidence pack you can send it, and Recommendations can suggest reporting answers that repeat a wrong claim.