How AI assistants decide what to say about a business

Training. A model is trained on a large collection of text gathered up to a cutoff date. What it learned about your business then is fixed until the next model. It can be months or years old.

Retrieval. Many assistants can also search the web when you ask. They fetch a few pages, read them and write an answer from them, often with links. This is called retrieval-augmented generation (RAG). The pages the answer rests on are called its grounding.

• ChatGPT can answer from training, or search the web and link its sources.
• Perplexity searches the web for most questions and numbers its sources.
• Google’s AI Overviews are written by Gemini models from Google Search results.

• Old training. The model learned last year’s hours.
• Old pages. The assistant fetched a page of yours that you forgot to update.
• Other people’s pages. A directory or review site says something different, and the assistant can’t tell who is right.
• Nothing to find. Your site doesn’t state the fact, or states it only in an image or a script, so the assistant guesses. A confident wrong guess is often called a hallucination.
• Mixed signals. Two of your own pages disagree.

• Your own pages: one clear, current statement of each fact.
• The format: plain text in the HTML, plus schema.org structured data.
• Access: let the crawlers in.
• Other sites: ask them to correct what they list.

What you can’t change: the model’s training, the assistant’s choice of sources, or how it words the answer.

1. Ask the assistant the questions a customer would ask.
2. When it shows sources, open each one.
3. Compare every fact with what is true today.
4. Fix wrong facts where they live, starting with your own site.
5. Ask again in a few weeks.