Retrieval, not ranking
A search engine returns a list of documents and lets the user choose. An assistant retrieves passages, synthesises an answer and cites two or three sources. Your competition is not the page ranked above you — it is the passage that answers the question more cleanly than your passage does.
This is why a page that ranks page-one can be invisible in AI answers, and why a modest page with tight, self-contained sections gets cited repeatedly.
What a chunk-ready answer looks like
Take the heading 'How long does a HubSpot migration take?'. A chunk-ready answer states the range, the variables and a qualifier, in one paragraph, with no pronouns pointing at earlier paragraphs and no 'as we discussed above'.
The test: copy the paragraph into a blank document. If it still answers the question for someone who has never seen your site, it can be retrieved. If it does not, it will be skipped.
The structural checklist
- H2s phrased as the questions people actually ask, not as marketing labels.
- One answer paragraph of 50–150 words immediately under each question heading.
- At least one table or numbered procedure per page — both are unusually extractable formats.
- Specific, dated facts. 'Four to eight weeks' beats 'fast'. '70,000 tickets in 2025' beats 'large migrations'.
- FAQ block with three to six genuine questions and FAQPage schema.
- Content present in the server-rendered HTML, verifiable by disabling JavaScript or reading the raw response.
Schema that earns its place
| Schema type | Use on | What it disambiguates |
|---|---|---|
| Organization | Site-wide | Who you are, where, and what you are known for |
| Service | Service pages | What is offered, to whom, in which area |
| FAQPage | Any page with Q&A | Question–answer pairing |
| Article | Blog posts | Author, publish and update dates |
| DefinedTerm | Glossary entries | Definitional intent |
| BreadcrumbList | All pages | Site hierarchy and topical grouping |
Corroboration beats assertion
Models weight claims that appear consistently across independent sources. A capability you only assert on your own site is weaker than one that also appears in a solution provider directory, a review platform and a customer's write-up.
Practical version: keep your directory listings, review profiles and public case study details describing the same facts in the same words. Inconsistency reads as uncertainty.
How to measure it
- 01Write 20 to 40 questions a buyer would ask an assistant before hiring you.
- 02Ask them in ChatGPT, Perplexity, Claude and Google AI Mode monthly, and log who gets cited.
- 03Track referral traffic from assistant domains separately in analytics — volume is small, intent is exceptionally high.
- 04Track branded search volume, which tends to move before referral traffic does.
Common mistakes
- Client-side rendering the answer content. If it is not in the HTML response, most AI crawlers never see it.
- Padded introductions before the answer. The chunk that gets retrieved is the one that starts answering immediately.
- Marketing-language headings ('Our approach') where a question would work.
- Schema that describes something different from the visible text.
- Judging AEO by traffic volume rather than by citation share and conversion quality.