Disclosure first. We are Proofmend and we sell SEO and GEO services, so read this knowing we operate in the market it describes. What follows is method and measurement, not a pitch for a vendor list.
We sample nine AI engines every week with the questions our clients' buyers actually type. ChatGPT, Claude, Gemini, Perplexity, Grok, Doubao, DeepSeek, Kimi and Qwen, each question in a fresh session, each engine on its default mode, conditions logged next to the observation. One pattern repeats in every round.
Ask an engine about a named brand and it usually answers accurately, often citing the company's own site. Ask the same engine for a recommendation in that category without naming anyone, and the same brand's mention rate collapses toward zero. The two questions travel different retrieval paths. The first pulls official and introductory material. The second pulls pages that already contain a ranked list, because an assistant assembling a recommendation borrows an existing ordering rather than inventing one.
That split is the whole job. Your site controls how you are described. Other pages control whether you appear at all.
What you control and what the engine decides
| Layer | Who owns it | What it changes |
|---|---|---|
| Facts on your own pages | You | Whether the description of you is correct |
| The page types engines quote | Third parties, with your participation | Whether your name enters the candidate pool |
| Judgement pages, reviews, ratings, forum threads | Third parties | How the "is this any good" question resolves |
| Retrieval and ranking inside the assistant | The platform | Everything else |
Only the first row is fully yours. Most disappointment in this field comes from spending the entire budget there and expecting movement in the fourth.
Blocks get quoted, prose does not
When we score pages against the answers that quote them, the shape of the content predicts citation better than its length. Ranked by the lift we measure, cited sources are worth about +115.1%, a stated numeric fact +61.6%, a definition block +57.3%, a comparison table +55.3%, and a clear procedure +41.2%. Two shapes go the other way, pure Q&A padding at about −5.7% and keyword stuffing at about −8.3%.
The reason is mechanical. An assistant composes an answer out of pieces it can lift whole. A paragraph of narrative with no extractable unit inside it has nothing to lift, however well written it is. Source attribution carries the largest lift for pages that start with little authority, which is exactly the position a lesser known brand is in.
Where the answer actually gets built
Three things are worth doing in order, and the order matters more than the effort in any single step.
Start by fixing the entity. One canonical domain, the product category stated in a single plain sentence, the company behind it named, and a disambiguation line if anything else shares your name. Put all of it in static HTML that survives with scripting off, since a fact that only exists after client-side rendering is a fact some crawlers never see.
Then get into the page types that engines quote for your category. Read the citation list under a live answer, group the domains, and you will see the pattern for your market within an hour. Comparison articles, category roundups, developer community posts and encyclopaedia entries dominate in most categories we sample. Publishing another post on your own blog does not move this layer.
Last, hold a recorded baseline. Same questions, same platforms, same modes, timestamps stored beside the answers, at least two rounds before you change anything. Without it, a shift in the answer tells you nothing, because these systems drift on their own week to week.
The part nobody sells you
For "is this vendor trustworthy" questions, the deciding sources are review platforms, rating sites and public discussion, not your content. We have watched an engine settle that question by quoting a site safety rating and a couple of news stories about scams in the category, with the vendor's own material absent from the answer. Content work cannot outrank a judgement page. That gap gets closed by earning third-party records, not by writing more posts.
Boundary
Proofmend does not guarantee rankings, AI mentions or citations, traffic, leads, or revenue. Everything above changes what an engine can find and how correctly it can assemble a description of you. What it does with that material is decided on the platform side, and anyone selling you a guarantee on that side is selling something they do not control.
Method, evidence rules and how we record unknowns are at https://proofmend.com/method/ , scope and limits at https://proofmend.com/ . Figures come from our own August 2026 sampling rounds and shift over time. Materials checked August 23, 2026.
Top comments (0)