How to Optimize Content for AI Search: Write Passages Engines Can Quote

Optimizing content for AI search (called AEO or GEO, sometimes AI content optimization) means making a passage of your page easy to retrieve, trust and quote inside a generated answer. Two gates apply: eligibility (can the engine fetch the page?) and selection (which eligible passage gets quoted?). Content work pays off at the second gate.

Eligibility is documented, and usually fine. Google needs a page indexed and snippet-eligible. OpenAI says sites opted out of OAI-SearchBot do not appear in ChatGPT search answers; Perplexity and Anthropic say similar. In AEOeye's crawler access study (October 2026; 391 well-known sites, a convenience sample), 1 of 193 SaaS sites blocked any AI retrieval bot in robots.txt, yet 24 of 88 e-commerce and consumer homepages challenged our crawler, so check firewall rules too.

Selection works on passages. Google's AI features are rooted in its core ranking systems, which include a passage ranking system for individual sections, and use retrieval-augmented generation plus query fan-out, related searches on sub-topics. The original RAG paper indexed 100-word chunks; ChatGPT, Perplexity and Claude publish no chunk sizes, so write sections that survive being lifted out.

Probably. A 2025 retrieval study found dense embedding and ColBERT-style models lose 15.6% of performance on average when the relevant information appears later in a passage; BM25 and rerankers were more robust. We found no direct test on ChatGPT or Perplexity, so this is mechanism, not proof.

Engines lift short spans, so the first sentence under each heading should answer that heading in 40–60 words and make sense alone. The length is our rule of thumb, not an engine limit. Anthropic's web search tool returns up to 150 characters of cited text per citation, about one sentence. Two habits help:

• Name the subject. Skip "it", "this approach" and "as mentioned above"; a lifted passage loses its neighbours.
• Phrase headings as questions. Pew Research found AI summaries on 53% of Google searches of 10+ words but 8% of one- or two-word searches (March 2025).

This does not mean chopping a page into fragments: Google says there is no requirement to break content into tiny pieces.

The best-measured elements are statistics, quotations and citations of credible sources. In the GEO study (Aggarwal et al., KDD 2024), each lifted a rewritten page's visibility by 30–40% on one metric and 15–30% on another, while keyword stuffing gave little to no improvement.

The setup: 10,000 queries; a research engine on GPT-3.5-turbo answering from the top five Google results; one source per query rewritten by an LLM and compared with its original.

The limits: a 2023 research engine, not Google's AI features or today's chatbots, measuring share within an answer, not retrieval. Google's guide never mentions these tactics; it says unique, non-commodity content will likely matter more than any other suggestion in it. Read the study as evidence for credibility, not a formula.

What to do:

• Put a real number, its source and its date in one sentence (anatomy of a quotable passage).
• Attribute quotes to the person who said them, and link primary sources.
• Add what only you have: your data, pricing and first-hand tests. Never let a tool invent a figure.

Schema markup helps machines parse a page and tell entities apart; it is not a citation switch and cannot replace quotable text. As of October 2026, Google says structured data is not required for its AI features and no special markup is needed. Use it because it is accurate and cheap; structured data for AI covers types and examples.

• Correlation exists. GEO-16, an observational study of 1,100 cited B2B SaaS pages, found metadata and freshness, semantic HTML and structured data most associated with citation. Its authors warn that confounders such as brand reputation may explain this.
• FAQPage is no longer a Google rich result. Google stopped showing it on May 7, 2026. Write FAQs for readers.
• Adoption is not effect. In the same study, Organization markup was on 129 of 176 SaaS homepages (73.3%): table stakes. FAQPage was on 13 (7.4%), but on 18 of 42 AI-visibility tool vendors (42.9%, aeoeye.com included). That shows what the AEO industry believes, not what works.
• llms.txt follows the same logic. The proposed file is on 108 of 180 checkable SaaS sites (60%), Google says it ignores such files, and we know of no major engine documenting its use for answers. See what is llms.txt.

Build the page as a stack of self-contained answers: a short answer first, one question per heading, tables where facts run in parallel. In the last column, Documented means an engine's docs say so, Measured means a published study found it, and Judgment means the call is ours. For a fuller audit, use our AEO audit checklist.

Yes for structure, editing and gap-finding; no for facts. "AI content optimization" also names this other job; only optimizing for AI engines decides whether they quote you. Google says generative AI can help with research and adding structure to original content.

• Good uses: restructuring a draft answer-first, listing sub-questions it leaves unanswered and tightening sentences.
• Bad uses: inventing statistics, quotes or sources (models predict likely words rather than retrieve facts, so any figure a tool adds is a guess until you source it), and mass-producing pages, which Google says may violate its scaled content abuse policy.

Publish fewer, better pages: one strong page per buyer question, with sub-questions as sections rather than near-duplicate pages. Google says a high quantity of pages does not make a site more relevant and warns against separate pages for every query variation or fan-out query. If two pages would answer the same buyer prompt, they are one page.

On 2026-09-24 we retired 234 of our blog posts and merged 27 duplicates, after Search Console showed about half of our 496 posts were unrelated to AI visibility and had no clicks in 90 days. On 2026-10-03 we merged another 165 duplicate-intent and research pages into 42; this guide replaces five of them. It is our position, not a measured result: the merges are days old and we have no citation data yet.

• Give readers a reason to click. Pew found Google users clicked a regular result in 8% of visits when an AI summary appeared, against 15% when none did. Answer first, then offer what a summary cannot: your data, a tool, the full comparison.
• Measure by asking. Re-ask the same buyer questions monthly. Google's Search Console report counts impressions in AI Overviews and AI Mode, and our free audit asks three buyer questions to ChatGPT.