What actually changed
For twenty years the goal was a position. You wanted to be one of ten blue links, and the user did the rest of the work. Answer engines removed that step. The assistant reads the candidate pages, writes the answer itself, and attributes a handful of sources.
That means two brands can rank identically in classic search and have wildly different outcomes in AI search. One gets quoted and linked. The other is read, summarised, and never named. The metric that matters is no longer position — it is citation share: of all the sources an assistant credits for questions in your category, how many are you?
See your own citation share first
Takes under a minute. No account needed for the preview.
How engines choose a source
The mechanics differ per engine, but the shape is consistent. A question is decomposed into sub-questions. Candidate documents are retrieved. Passages are scored for how directly and confidently they answer each sub-question. The answer is composed from the best passages, and those documents get cited.
- Retrieval favours pages that are indexable, fast and readable without JavaScript.
- Passage scoring favours specific, self-contained statements — not marketing prose.
- Composition favours sources that agree with each other, which is why third-party mentions matter.
- Attribution favours the source that contributed the most quotable language.
The seven levers
Across the retail audits we run, nearly every gap traces back to one of these seven.
Entity clarity
The model has to be able to say, in one sentence, what your brand is and what it sells. Most homepages bury this under a campaign slogan.
Answerable pages
A page that answers one question completely, in plain language, near the top, is quotable. A page that requires scrolling and inference is not.
Structured data
Product, FAQ, Organization and Breadcrumb markup is how machines confirm what your prose implies. Missing markup makes you a weaker source than a retailer who has it.
Third-party corroboration
Assistants prefer claims that appear in more than one place. Retail listings, review sites and trade press are part of your citation surface, not a distraction from it.
Freshness signal
Dated, maintained pages beat undated evergreen pages when the question implies current information — price, availability, comparison.
Crawlability for AI agents
Rendering everything client-side, blocking agents in robots.txt, or gating specs behind interactions removes you from the candidate set entirely.
Comparison content
A large share of commercial prompts are comparative. If you never publish an honest comparison, someone else's version becomes the cited one.
How to measure it
Pick the twenty questions a buyer actually asks before purchasing in your category. Ask them across every engine. Record who gets cited. Repeat weekly. That is the whole method — the difficulty is doing it consistently and comparably, which is what the audit automates.
Citation share
Your sources as a percentage of all cited sources.
Answer coverage
Share of category questions your site answers cleanly.
Entity accuracy
Whether the model describes you correctly, unprompted.
Five expensive mistakes
- 1.Treating AEO as a keyword exercise. Questions, not keywords.
- 2.Blocking AI crawlers reflexively, then wondering why competitors are cited instead.
- 3.Publishing comparison pages that only flatter you — models discount one-sided sources.
- 4.Leaving product specs in images or PDFs where nothing can extract them.
- 5.Measuring once. A single snapshot cannot tell you whether a fix worked.