Somewhere in a model’s weights is a description of your company. You didn’t write it, you can’t edit it directly, and it’s being served to buyers right now. Here’s what to do about that.
Why the errors happen
Language models assemble a picture of your brand from everything they’ve read: your site, directories, review platforms, news coverage, forum threads, and competitor comparison pages. When sources disagree, the model doesn’t investigate — it weights by apparent authority and repetition.
That means a confidently written competitor page can outweigh your own product documentation, and a three-year-old directory listing can outrank the pricing page you updated last week.
The four errors we see most
- Stale pricing. Almost universal. A retired pricing page or an old review article keeps circulating a number you no longer charge.
- Missing capabilities. Features shipped in the last 12–18 months are frequently described as unavailable, because the sources describing your product predate them.
- Category misplacement. Being described as a competitor to companies you don’t compete with, usually traceable to one influential comparison article.
- Entity merging. Being confused with a similarly named company — common for short or generic brand names, and the hardest to fix.
You fix sources, not models
There is no support ticket for “ChatGPT is wrong about us”. The only durable lever is changing what the model reads next time it crawls.
In rough order of effectiveness:
- Correct your own properties first. Retire or redirect outdated pages rather than leaving them live. Make the current facts prominent and machine-readable.
- Fix third-party listings. Directory entries, review profiles and partner pages. Tedious, high yield — these are heavily weighted as independent corroboration.
- Publish a canonical reference page. A single, well-structured page stating what you do, who you serve, what you cost and how you compare. Something specific enough to be quoted.
- Earn fresh coverage. Recent, credible third-party writing is the strongest counterweight to stale information.
Expect it to take weeks
Corrections propagate at the speed of recrawling and index refresh, not at the speed of your deploy. Six to twelve weeks is normal for a well-sourced correction to show up consistently across engines.
Measure before and after against the same prompt set, or you won’t be able to tell whether anything worked.