Most “AI search” advice is fixing the first list. The answer is decided by the second.
What teams think it is — the noise
llms.txt files
Barely any assistant reads one today. Where it is read, it points at your own content — it does not make anyone recommend you.
Schema / structured data
Helps a page be understood. Does not make it the pick.
Markdown copies of your pages
A retrieval convenience at best. Not a reason to be named.
“AEO” and “GEO” hacks
Reworded headings and FAQ blocks. Models summarise what pages say, not how they are formatted.
“Writing for AI”
The pages that get cited answer the question better, with specifics — not ones stuffed with AI-friendly phrasing.
GEO agencies and services
Most sell the list above. Useful housekeeping, sold as a visibility lever.
These are housekeeping. A few help a page get read; none of them decide whether you are the answer.
What actually moves the answer — the signal
Reviews
G2, Capterra, Trustpilot and the review sites in your category. Models read them and repeat what they find.
Earned media and industry publications
A roundup or a buyer guide on a trade publication is a page the model retrieves by name.
Your own comparison and category pages
A clear “X vs Y” or “best X for Y” page is more citable than a homepage.
YouTube and creator reviews
Video answers mirror what creators already say about the category. Most tools ignore this entirely.
Brand mentions across the web
The more places you are named accurately, the more the model has to go on.
Google rankings
Perplexity leans on search results. Where you rank still feeds the answer.
Topical authority
Being known for one thing beats being mentioned for ten.
Entity consistency
Same name, same category, same description everywhere a model might read it.
Third-party recommendations
Someone the model trusts saying you are a good fit. This is the whole game.
This is what a model retrieves and repeats. It is slower, it looks more like PR than SEO, and it compounds.
A model is reading third-party pages about you, not your meta tags.
Ask an assistant “best project management tool for agencies” and watch what it does: it searches, opens four or five pages, and writes a summary. Those pages are review sites, listicles and comparison articles. Your homepage is rarely one of them, and its structured data never is.
That is why the technical checklist underdelivers. llms.txt mostly is not read, and where it is, it just points at content you already publish. Schema helps a page be parsed, not preferred. “Writing for AI” tends to mean padding a page with phrases that make it worse for the human who has to choose.
The tell: the brands winning AI answers today are usually not doing anything clever technically. They have the reviews, the coverage and the comparison pages.
Work the second list.
Get onto the pages the model already reads
The review sites, directories and “best X” roundups in your category. Being listed accurately — right category, pricing shape, what you sell — does more than any markup.
Make your own comparison and category pages the clearest answer
One real buyer question per page: who it is for, what it replaces, one proof point. Not a homepage boast.
Earn real reviews and real coverage
Slow, unglamorous, and the part competitors cannot copy overnight. This is the moat.
Keep your entity consistent
Same name, category and description on every profile a model might retrieve. Contradictions make you easy to skip.
Then check whether it is working
Which is the one part software is good for.
Prescy is the measurement layer for the second list — not the execution.
It does not write your llms.txt or audit your schema, because that is not what decides the answer. It asks ChatGPT and Perplexity the buyer questions that matter in your category, shows which third-party pages they cite, and whether those pages name you or a competitor.
You get a ranked list of the sources deciding your category and a next step for each — get onto this review site, strengthen this page, pitch this roundup. The work stays yours. Prescy tells you where it is paying off.