Content formats AI systems parse best
Two pages can contain identical information and perform completely differently as AI sources - because extraction favors structure. This lesson is the practical pattern library: what gets cited, what gets skipped, and how to retrofit what you already have.
The properties extraction rewards
Four patterns recur across everything that gets cited well. Answers stated up front: a section that opens with its conclusion ("Linen runs about half a size large in most cuts; size down if between sizes") extracts perfectly; one that builds to it across four paragraphs extracts poorly. Journalists call it the inverted pyramid; it's now machine-readability advice. Self-containment: each section comprehensible without the rest of the page - because retrieval often surfaces fragments, not full pages. Pronouns referring to a previous section's subject die in extraction. Explicit facts over implication: numbers, names, and units beat adjectives ("165 g/m² midweight linen" over "substantial feel"). Honest headings: a heading that names what its section answers ("How should a linen shirt fit?") is a retrieval target; a clever one ("The Drape Debate") is a missed connection.
Formats that consistently win citations
The format ranking, from audit experience across assistants: Q&A/FAQ structures - near-perfect extraction shape, question as heading, direct answer first (add FAQPage schema and it's machine-legible twice). Comparison content with tables - "X vs Y" queries are fanout staples, and a clean table is the densest fact structure a page can offer. Specification blocks - structured attribute lists on product pages (your metafields, rendered - Phase 1's theme lesson completing its arc). Definitional and how-to content with stepped structure. And the one that surprises people: genuinely opinionated expert content - assistants synthesizing "which should I choose" answers cite sources that commit to recommendations over sources that neutrally describe. Hedged content reads as low-information. Your practitioner knowledge, stated plainly, is a citation asset.
Retrofitting, not rewriting
The good news: this is mostly restructuring, not new creation. The retrofit pass for an existing guide or collection page, in order of ROI: front-load the answer in each section, convert buried question-answers into an FAQ block, table-ify any comparison currently living in prose, break wall-of-text into honest headings, and de-pronoun section openings. An afternoon per key page. Prioritize by the audit you just ran - pages on topics where you're absent from AI answers get retrofitted first.
One boundary worth stating: everything above serves human readers identically - scanability is extractability. The moment a tactic diverges from that (blocks of robot-directed text, invisible content, keyword incantations), you've crossed from formatting into the snake oil this phase warned about. The test stays the same as it's been since Phase 2: are we informing the reader, or gaming the parser? Only one of those survives model updates.