AEO in practice: how to get cited by ChatGPT, Perplexity and AI Overviews
Answer engines increasingly resolve queries without a click. Here is what actually changes in how you structure content, schema and claims.
The short answer
To be cited by AI answer engines, structure content answer-first with self-contained claims, implement explicit schema (Organization, Service, FAQPage, DefinedTerm), publish an llms.txt index, allow AI crawlers in robots.txt, and keep factual claims consistent across the site. Measurement requires prompting the assistants directly, since answered queries produce no referrer in analytics.
A growing share of commercial research now happens inside an assistant rather than a results page. The user asks, the model answers, and often no click follows. Optimising for that is a related but distinct discipline from ranking.
What changes, and what does not
Most of classic technical SEO still applies — crawlability, speed, clean architecture, genuine authority. What changes is the unit of success. You are no longer competing for a position in a list; you are competing to be the source a model quotes, and models quote what they can extract cleanly and verify consistently.
Write answer-first
The single highest-leverage change is structural. Lead every page with a self-contained answer of roughly 40 to 70 words that would read correctly if quoted with no surrounding context. No 'we', no throat-clearing, no 'in today's fast-moving landscape'. State the thing.
Then support it. The detail below the answer is what establishes that the answer is credible, but it is the opening block that gets extracted.
Make claims machine-readable
- Organization and ProfessionalService for entity identity, with consistent @id references
- Service schema on commercial pages, with an offer catalogue
- FAQPage where you genuinely answer questions
- DefinedTerm and DefinedTermSet for glossary content — the highest-leverage type for citation
- BreadcrumbList so hierarchy is explicit rather than inferred
Schema must describe what is actually on the page. Marking up content that does not exist is a manual action risk and models increasingly cross-check.
Publish llms.txt
An emerging convention: a plain-markdown file at /llms.txt summarising who you are, your key facts and a curated index of your important pages. It is not a standard with guaranteed support, but it costs an hour to generate from data you already have and gives any agent reading your site a clean, unambiguous starting point.
Allow the crawlers
Check robots.txt. GPTBot, OAI-SearchBot, PerplexityBot, ClaudeBot and Google-Extended are all commonly blocked by default in boilerplate configurations or by well-meaning security teams. If you want to be cited, they need access — and blocking them does not stop training, it only stops attribution.
Be consistent
Models penalise ambiguity. If your pricing is described three different ways across your site, or your founding year differs between your About page and your schema, an assistant has no basis to choose and will often skip you for a source that is internally consistent.
Classic SEO rewards the page that best matches a query. Answer engines reward the source whose claims are cleanest to extract and hardest to contradict.
How to measure it
There is no analytics report for a query that never produced a click. The workable method is manual and repeatable: build a list of the questions your buyers actually ask, prompt each major assistant with them on a schedule, and record whether you appear, in what position, and how you are characterised.
Track the characterisation as carefully as the appearance. Being mentioned inaccurately is worse than not being mentioned, and it is usually fixable by tightening the claims on the pages the model is drawing from.
What not to do
- Do not generate hundreds of thin question pages — scaled content abuse is penalised in both search and answer quality
- Do not stuff schema with claims your page does not support
- Do not write for extraction at the cost of readability; humans still convert, models do not