A client spent three months publishing pages to fix what ChatGPT said about them. It was never reading their site.
Sharing this because I have now watched three companies make the same mistake and it costs them a quarter every time.
A client came to me in the spring because ChatGPT was quoting pricing for them that was two years out of date. Not slightly off. It was describing a package they killed in 2024, at a price they had not charged since. Their sales team kept getting on calls with people who had already anchored to a number that did not exist.
What they had done about it before calling me was publish. Four new pages about pricing on their own blog, a pricing FAQ, a comparison page. Three months of work. Nothing moved. Not one assistant changed its answer.
The reason is boring but it matters. When you ask an assistant about a brand, most of what it tells you did not come from that brand's own site. It comes from what third parties have written about them, plus whatever is sitting in the model's weights from training. Their blog was not in the retrieval set for that question, so publishing on it was shouting into a room the model was not standing in.
The thing that finally told us what was going on took ten minutes. Ask the same question twice, once with web search on and once with it off. If the wrong answer only appears with search off, it is baked into the model and there is no quick fix. You are waiting on the next training run, and the only useful work is getting the correct version onto enough third party surfaces that the next one learns something better. If the wrong answer appears in both, then whatever it is pulling live is wrong, and that you can go and fix now.
Theirs was the second kind. It was reading two comparison articles on sites they had never heard of, both from 2024, both with the dead pricing. We got one corrected by emailing the author, who was perfectly reasonable about it, and got a newer piece published on a site the same models were already citing for that question. Answers started changing about five weeks later.
What I would take from it: before you write anything to fix an AI visibility problem, go find out what it is actually reading. It is almost never you.
Happy to go into how we worked out which sources it was pulling if that is useful to anyone.