Hallucination

Large language models don’t fetch facts the way a database does; they generate answers by predicting the most likely continuation of the text. When a model lacks enough consistent information on a topic, it can fill the gap with something plausible but false. A fluent answer isn’t necessarily a correct one.

The risk for brands

Hallucinations aren’t limited to general knowledge. A model can get a company wrong too:

• A branch or service that doesn’t exist
• An old address or phone number
• Confusing the company with another one of the same name
• A made-up founding year or founder

Errors like these can send a potential customer to the wrong place.

How to reduce it

Preventing hallucinations entirely isn’t in a site owner’s hands. But you can make it easier for models to find the right information:

• State the basics (name, address, services) clearly and in one place on your site
• Mark them up with structured data
• Use the same details on external sources such as directories and social profiles
• Make sure search-based models can reach your site (RAG)

Example

If someone asks an assistant for a clinic’s opening hours and the clinic’s site doesn’t list them, the model may give another clinic’s hours or a generic guess. Listing the hours clearly and marking them up with openingHours lowers that risk.