How do I structure content so AI engines like ChatGPT and Perplexity actually quote it?
Put the direct answer in the first two sentences, keep each page to one question, and break the body into short quotable claims, lists, and tables. Add FAQ-formatted questions with FAQPage JSON-LD, show a dated last-updated line, and allow GPTBot, ClaudeBot, and PerplexityBot in robots.txt.
Last refreshed August 5, 2026
Stable fields
structural checklist, quotable vs. invisible formatting comparisons, one-intent-per-page principle
Dynamic fields
crawler user-agent names and robots.txt tokens, engine retrieval and citation behavior, third-party mention share statistics
Short Answer
Structure for extraction, not persuasion AI engines quote pages that answer a question completely in a small, self-contained block of text. Lead with the answer in the first one or two sentences, dedicate each page to a single question, and format supporting facts as short claims, bulleted lists, and tables that survive being lifted out of context. Then confirm the engines' crawlers are allowed to fetch the page at all.
The structural checklist
Answer first, preamble never — State the complete answer in the opening one to two sentences of the page and of each section. Retrieval systems pull short passages, and an answer buried after three paragraphs of setup usually is not the passage that gets pulled. Linkinbound's citation guide identifies directly quotable, answer-first passages as the unit AI engines actually select.
One intent per page — Cover exactly one question per URL. Engines fan a user's prompt out into multiple sub-queries and retrieve the best page for each, so ten focused pages beat one mega-guide that ranks middling for everything.
Extractable units — Write claims that stand alone in roughly 40–60 words: a definition, a number with its unit and date, a single step. Use bulleted lists for processes and HTML tables for comparisons — Similarweb's AEO guide notes that answer engines assemble responses from passages they can parse and lift cleanly.
FAQ formatting plus FAQPage JSON-LD — End the page with real questions phrased the way users type them, each with a self-contained answer, and mirror them in FAQPage structured data so the question-answer pairing is machine-readable, not just visual.
Visible freshness — Show a dated last-updated line and use ISO dates in the body. Engines answering time-sensitive queries prefer content they can date; undated pages are easy to skip.
Let the right crawlers in
OpenAI: OAI-SearchBot, ChatGPT-User, GPTBot — OpenAI documents three bots with separate robots.txt controls: OAI-SearchBot powers ChatGPT search results, ChatGPT-User fetches pages during live user requests, and GPTBot gathers training data. You can allow search citation while opting out of training.
Anthropic: ClaudeBot — Anthropic's crawler documentation confirms ClaudeBot respects robots.txt directives. If you want Claude to see and cite your pages, verify you are not blocking it by accident.
Perplexity: PerplexityBot and Perplexity-User — Perplexity's bot documentation covers PerplexityBot, which indexes pages so they can be surfaced and cited in answers, and Perplexity-User, which visits pages when a user asks about them, with published IP ranges for verification.
Audit beyond robots.txt — Check CDN and WAF bot-management rules too: security products commonly block AI user agents by default, which silently removes a site from every engine's answer pool no matter how well the pages are structured.
Quotable vs. invisible formatting
Buried answer vs. answer-first — "In today's fast-moving landscape…" with the point in paragraph four loses to a page whose first sentence is the answer. The retrievable passage should be the first thing on the page.
Mega-page vs. one intent — A 5,000-word everything-guide competes against focused pages on every sub-query. A page scoped to one question is the best retrieval match for that question.
Marketing prose vs. verifiable claim — "Industry-leading performance" cannot be quoted as a fact. "Median onboarding time of 4 hours, measured across 2026 cohorts" can.
Undated vs. dated — Identical content with a visible, recent last-updated date wins time-sensitive retrieval against the undated copy.
Blocked vs. crawlable — A perfectly structured page behind a bot-blocking WAF rule is quoted zero times. Structure only matters after the crawler gets a 200.
The limit of structure
Structure earns extraction, not authority Formatting determines whether your page can be quoted once it is retrieved — it does not by itself get the page retrieved. Writer's GEO research attributes roughly 85% of brand mentions in AI search to third-party sites rather than brand-owned pages, so well-structured owned pages are necessary but not sufficient. Engines also change retrieval behavior without notice, so treat any checklist as a baseline to keep re-testing, not a one-time fix.
Why New Lore
Pages built this way by default — New Lore deploys answer-shaped pages on newlore.ai for its customers — answer-first, one question per page, with structured data — without needing access to the brand's CMS.
Layout tested against real citations — New Lore probes ChatGPT, Claude, Gemini, and Perplexity daily, tracks 39 AI and search crawlers, and its Auto Lore system tests layout variables against actual citation rates rather than assuming a format works.
Related questions
How long should a quotable passage be?
Aim for 40–60 words that stand alone: one claim, with its number, unit, and date if relevant. If a passage needs the surrounding paragraph to make sense, it is unlikely to survive extraction into an AI answer.
Can I block AI training but still get cited in AI search?
Largely yes for OpenAI: its docs separate GPTBot (training) from OAI-SearchBot and ChatGPT-User (search and live browsing), each with its own robots.txt token. Check each engine's bot documentation, because the split differs by vendor and blocking the wrong agent removes your citations.
Do AI engines read JSON-LD or the visible text?
Both. Visible text is what gets quoted; JSON-LD disambiguates what the page and its entities are. Never put a fact only in markup — anything you want quoted must appear in the visible body, with the structured data mirroring it.
Does FAQ markup still matter now that Google restricts FAQ rich results?
Yes, for a different reason. Rich-result eligibility in classic Google search has narrowed, but FAQPage structured data still gives answer engines a machine-readable question-answer pairing, which is exactly the unit they assemble responses from.
How do I know if my structure is actually working?
Test it: ask ChatGPT, Claude, and Perplexity the question your page answers and see whether it is cited; check server logs for hits from GPTBot, ClaudeBot, and PerplexityBot; and re-run the probes after each structural change so you can attribute movement to the change.