ChatGPT names you through one of two routes: it browses and cites a page, or it recalls you from training. Optimize both. Make your pages retrievable and quotable for the browsing route; make your description consistent across independent sources for the memory route. Then test with the actual prompts your customers use.
How ChatGPT arrives at a recommendation
When someone asks ChatGPT for a recommendation, one of two things happens. Either it searches the web, reads a few results and writes an answer with links — or it answers from what it already knows, with no links at all. Which route it takes depends on the question, the settings and whether the model judges the topic to be time-sensitive.
Both routes end in the same place: a short list of named businesses. Your job is to be eligible on both.
Browsing versus memory, and why both matter
The trap is optimizing only for browsing, because it is the measurable one, and then discovering that the highest-intent questions get answered from memory with a competitor's name.
The playbook
Check robots.txt and your firewall. Blocking OpenAI's crawlers removes you from the browsing route entirely — the one route you can actually influence quickly.
Not “Services”. “How much does an AI visibility audit cost”, “Do you work with agencies outside Kansas City”. Title it as the question, answer it in the first paragraph.
Assistants strongly favour sources that contain the concrete detail the asker wanted. Vague pages get skipped even when they rank.
People ask assistants to compare. If the only comparison of you and a competitor was written by that competitor, that is the source being read.
Industry directories, association listings, local press, well-moderated forums, review platforms. These appear disproportionately in cited results.
Name, address, hours, service area, pricing, credentials — one page, plain text, marked up, updated. Everything else points at it.
Same wording, same list, logged. Answers move without warning; only a repeated test tells you which way.
Prompts worth testing every month
Use your own category and city. The point is to ask the way a customer asks, not the way a marketer searches.
Record whether you were named, in what position, what was said, and which sources were linked. That log is the whole measurement programme.
What does not work
Prompting the model to remember you. Anything you tell ChatGPT in a chat affects that chat. It does not teach the model.
Keyword stuffing. Extraction rewards clarity, not density. A stuffed page reads as low quality to exactly the systems you are courting.
Hidden text aimed at models. Instructions buried for crawlers are treated as manipulation and risk your eligibility outright.
Buying a spot. The cited body of an answer is earned. Anyone guaranteeing placement is guessing or lying.
Common questions
Does having a lot of traffic help?
Indirectly. Traffic correlates with the mentions and coverage that do help, but the model is not reading your analytics.
Should I block AI crawlers to protect my content?
That is a legitimate strategic choice with a cost: you stay in the answers via other people's descriptions of you, with no ability to correct them.
How fast can this move?
Crawl-access and page fixes have shown up in browsing answers within two to three weeks. Memory-route changes take months.
What about the other assistants?
The mechanics are close enough that the same work serves Perplexity, Claude, Copilot and AI Overviews. Start with What is AEO for the general version.
The free check asks six engines about your business and shows you the answers.