How to write content for answer engines

To write content for answer engines, often called AEO content, give each buyer question its own page, answer it in the first two or three sentences in the buyer's words, and support the answer with specific facts a reader can check, such as prices, scope, dates, and comparisons. Keep those facts in the page's text and consistent across your site. Skip generic rewriting "for AI": a 2025 benchmark found most such methods largely ineffective and frequently harmful to ranking (opens in a new tab).

This guide explains how answer engines pick the passages they cite, which writing choices independent studies support, and which popular tactics they do not. It is part of AEO HQ's complete guide to answer engine optimization. Statements about how engines work come from the platforms' own documentation. Effects come from laboratory studies and one field study, each labeled. The steps are our recommendations. Sources were checked on September 27, 2026.

Scope and definitions

This guide is for people who plan content marketing for AEO or write and edit a B2B company's website, such as marketers, founders, and subject experts. It covers the text of pages you control. Earning coverage on other sites is covered in the guide to brand mentions, and the evidence on schema markup has its own page.

• Answer engine. A system that answers a question in its own words and names or links its sources, such as ChatGPT search, Google's AI Overviews and AI Mode, Perplexity, Claude, and Microsoft Copilot.
• AEO content. Pages written so that an answer engine can find them for a question and quote them correctly. In this guide, it means ordinary web content held to a stricter standard of clarity and specificity, not a separate format.
• Retrieval. Fetching candidate pages from a search index to answer a question. A page that is not retrieved cannot be cited.
• Reranking. Reordering the retrieved pages by relevance before the model reads them. Pages below the cutoff never reach the model.
• Grounding. Basing an answer on retrieved pages. Google describes retrieval-augmented generation as a technique "also known as grounding" (opens in a new tab) that relies on its core ranking systems to retrieve pages from its index.
• Passage. A part of a page that an engine can use on its own. Microsoft says assistants such as Copilot break content into "smaller, structured pieces" (opens in a new tab) that are then evaluated for authority and relevance.
• Query fan-out. Google's term for "a set of concurrent, related queries generated by the model" (opens in a new tab) to gather more results for one question.

How answer engines choose what to quote

Answer engines do not read a page from top to bottom the way a person does. They decide whether to search, turn the question into search queries, retrieve and rerank candidate pages, and then write the answer from the passages that remain. The table summarizes what helps a page at each step. The sections below give the evidence.

1. The engine decides whether to search

A model can answer from what it learned in training, its parametric knowledge, or search the web first. Anthropic's documentation says Claude searches when a request depends on information that is "current, changing, or outside its training data," including "information about specific organizations, people, or products that might have changed" (opens in a new tab). Those are the questions a company's own pages answer: what it offers, what it costs, and how it compares.

2. The engine turns the question into searches

Google says AI Overviews and AI Mode may use a query fan-out technique, "issuing multiple related searches across subtopics and data sources" (opens in a new tab). OpenAI says that when ChatGPT search uses outside search providers, it "typically rewrites your query into one or more targeted queries" (opens in a new tab). In Google's example, the question "how to fix a lawn that's full of weeds" might fan out to "best herbicides for lawns", "remove weeds without chemicals", and "how to prevent weeds in lawn" (opens in a new tab). A page can therefore be cited for a sub-question it answers well, not only for the question as typed.

Google also warns against the obvious response. Creating separate content for every variation of how people search, including fan-out queries, primarily to manipulate rankings or generative AI responses, violates its scaled content abuse spam policy (opens in a new tab).

3. The index retrieves candidate pages

A page has to be found before it can be quoted. Google says that to be shown in AI Overviews or AI Mode, a page must be indexed and eligible to be shown in Google Search with a snippet (opens in a new tab). Crawlers read the text the server sends: in Vercel's network data from December 2024, none of the major AI crawlers rendered JavaScript (opens in a new tab). Bing's guidelines warn against hiding critical content behind client-side rendering (opens in a new tab).

Words matter at this step, at least for keyword-based retrievers. In SAGEO Arena, a peer-reviewed laboratory pipeline that used a keyword-based (BM25) retriever, rewrites that added technical terms or unusual words caused the largest drops in retrieval, which the authors attribute to a "lexical mismatch" with queries that "typically use common vocabulary" (opens in a new tab). In the same pipeline, optimizing a page's title, meta description, headings, and schema fields gave a 22% higher retrieval hit rate than optimizing the body text alone (opens in a new tab).

Google describes its own systems differently. It says AI systems "can understand synonyms and general meanings," so site owners need not capture every variation of how people search (opens in a new tab). Our reading of both sources: use the words your buyers use, once, in the places that carry the most weight (the title, the H1, the first paragraph, and the headings), and do not stuff in variants. That serves keyword-based and meaning-based retrieval alike.

4. Reranking favors early answers and focused pages

After retrieval, a reranker orders the candidates, and only the top ones reach the model. In SAGEO Arena, "placing the answer early in the document yields higher reranking scores, whereas restructuring that displaces the answer to later paragraphs results in significant rank drops" (opens in a new tab). The reranker also favored additions that matched the question's information need and penalized additions that widened the page's scope (opens in a new tab). One automated rewriting method lost 22.35 retrieval ranks because it lengthened and diluted the documents it rewrote (opens in a new tab).

Bing's guidelines point the same way. They say to "place essential information near the top of the URL" (opens in a new tab) and avoid long introductions, and that "URLs focused on a primary topic are more likely to be selected for grounding results" (opens in a new tab).

5. The model chooses what to use and cite

With the candidates in front of it, the model picks which passages to use. Three findings are consistent:

• Relevance comes first. In an ACL 2024 study, models relied heavily on how relevant a page was to the question, while largely ignoring features people find persuasive, such as scientific references or a neutral tone (opens in a new tab) (peer-reviewed).
• Specific facts help; formatting alone does not. In 252,000 controlled trials across six models, topical relevance and list position were the biggest drivers of being cited first; an explicit price and a recent date also helped consistently, while formatting-only edits had little impact (opens in a new tab) (peer-reviewed; laboratory; the authors work for a marketing software company). Internal contradictions and hedged language were smaller, secondary disadvantages (opens in a new tab).
• Extractable evidence is used more. Across 602 prompts on ChatGPT, Google, and Perplexity, the cited pages that most shaped answers were richer in definitions, numerical facts, comparisons, and procedural steps (opens in a new tab) (preprint; descriptive, so it shows association, not cause).

Engines also misquote. In a 2023 audit of four generative search engines, only 51.5% of generated sentences were fully supported by their citations, and 74.5% of citations supported their sentence (opens in a new tab) (peer-reviewed; the engines tested have since changed). Bing asks that facts and definitions be explicit and that key statements not rely on implied content (opens in a new tab). Our reading: a fact stated plainly gives an engine less room to get it wrong.

Why rewriting pages "for AI" rarely helps

A widely quoted result comes from the 2023 paper that introduced generative engine optimization (GEO). It reported that rewriting methods could raise a page's visibility "by up to 40% in generative engine responses" (opens in a new tab). A 2026 review of 45 studies found those gains "valid within its experimental setting but conditional on a source already being present in a fixed context" (opens in a new tab) (preprint). Later tests of the same kind of rewriting were mostly negative:

• A NeurIPS 2025 benchmark found statistically significant ranking gains in only 3 of 54 cases, and adding statistics lowered rankings in 19 of 24 settings (opens in a new tab). Moving a page to first place in the model's context gave far larger gains than any rewriting method (opens in a new tab) (peer-reviewed).
• In SAGEO Arena, existing optimization methods often degraded visibility at the retrieval and reranking steps (opens in a new tab) (peer-reviewed).
• The 2026 review found that generic heuristics transfer poorly, citation-oriented rewrites can impair retrieval, and no reviewed technique showed a stable, longitudinal, cross-platform causal effect on organic discoverability (opens in a new tab).
• Rewritten pages can be detected. One detector identified GEO-optimized pages with an F1 score of 0.944 and estimated that 8.90% of 10,095 pages in Google and Gemini results were GEO-optimized, rising to 16.36% of pages modified in 2026 (opens in a new tab) (preprint). A proposed defense cut the success rate of seven GEO rewriting attacks from 50.32% to 6.20% by demoting rewritten documents (opens in a new tab) (preprint).

Google tells site owners they don't need to "write in a specific way just for generative AI search" (opens in a new tab). Our reading of the evidence: engines reward pages that are more relevant and more specific, not pages that sound optimized.

Steps

These steps are AEO HQ's recommendations. Each notes the evidence behind it.

1. Collect the questions buyers ask, in their words. Use sales calls, support tickets, demo requests, and search query reports. Bing Webmaster Tools also reports grounding queries, the "key phrases the AI used when retrieving content that was referenced in AI-generated answers" (opens in a new tab). Questions are where AI answers appear: in Pew's March 2025 browsing data, 60% of Google searches that began with a question word produced an AI summary, against 8% of one- or two-word searches (opens in a new tab). In a March–April 2026 survey of 519 U.S. B2B professionals who use AI at work, 61% said they describe their specific use case when researching vendors with AI, and 56% ask for direct vendor comparisons (opens in a new tab) (vendor survey). Evidence: moderate.
2. Give each distinct question one page, and answer its close variants on that page. Bing favors URLs focused on a primary topic (opens in a new tab). Google treats pages made for every variation of a query, when the aim is to manipulate rankings or AI responses, as scaled content abuse (opens in a new tab). Evidence: strong (official guidance).
3. Answer the question in the first two or three sentences. Name the subject in the answer itself, so it still makes sense when quoted alone. Reranking favored early answers in the laboratory pipeline described above. The only controlled field study changed titles to question form and rewrote lead summaries as standalone two-to-three-sentence answers, among other changes (opens in a new tab), and estimated a 1.82-fold rise in ChatGPT referrals relative to unchanged pages (95% CI 1.31 to 2.54), but a placebo test gave p = 0.16, so the authors call the effect "suggestive, not conclusive" (opens in a new tab) (preprint; one site; the authors work for the company that owns it). Evidence: moderate.
4. Use the buyer's words in the title, H1, first paragraph, and headings. Prefer plain words to jargon, for the retrieval reasons above. Do not repeat keywords: the authors of the GEO paper found that keyword stuffing, while "widely used for Search Engine Optimization," (opens in a new tab) offered "little to no improvement on generative engine's responses" (opens in a new tab), and Google's spam policies define keyword stuffing as filling a page with keywords or numbers to manipulate rankings (opens in a new tab). Evidence: moderate for plain wording; strong against keyword stuffing.
5. State specific, checkable facts. Give prices, what is included, turnaround, dates, specifications, and comparisons with named alternatives. Use a number only when it is true, sourced, and answers the question: in the NeurIPS benchmark above, adding statistics lowered rankings in 19 of 24 settings (opens in a new tab). Buyers check what they read. In a January 2026 survey of 1,862 technology buyers, 94% of those who used AI said they fact-check its responses at least some of the time (opens in a new tab) (survey by a review platform), and in the Semrush survey, 71% said they visit a vendor's website after an AI names the vendor (opens in a new tab). Evidence: moderate (laboratory studies and surveys).
6. Make each section stand on its own. Use descriptive headings, numbered lists for steps, and tables for comparisons. Microsoft says headings act "like chapter titles that define clear content slices" (opens in a new tab) for AI. Google says there is no requirement to break content into tiny pieces for AI (opens in a new tab), and it notes that people appreciate pages organized by paragraphs and sections, with headings that give them a clear structure (opens in a new tab). Treat structure as help for readers and retrieval, not as a citation trick: formatting-only edits had little effect on which source models cited, as noted above. Evidence: moderate for retrieval; weak for citation.
7. Keep answers and facts in visible HTML text. Microsoft advises against hiding important answers in tabs or expandable menus, and against relying on PDFs for core information (opens in a new tab). Bing says images and video should not be the sole source of information (opens in a new tab) needed to understand a topic. In a one-page practitioner test, no AI system read facts that appeared only in JSON-LD when it fetched the live page (opens in a new tab) (October 2025). Google asks that structured data match the visible text on the page (opens in a new tab). Evidence: strong for the platform guidance; weak for the test.
8. Date pages honestly, and update them when facts change. A recent date helped in the six-model trials above. Bing asks site owners to "update content when facts or guidance change" (opens in a new tab) and to use freshness signals appropriately (opens in a new tab). Change the visible date only when the content changes. Evidence: moderate.
9. Keep facts identical across pages, profiles, and markup. Bing asks for clear and consistent naming for people, organizations, products, and locations (opens in a new tab). In the six-model trials, internal contradictions counted against a source. Evidence: moderate.
10. If AI tools help you draft, check every claim, and do not mass-produce pages. Google says generative AI can be "particularly useful when researching a topic, and to add structure to original content," (opens in a new tab) but that using it "to generate many pages without adding value for users may violate" (opens in a new tab) its scaled content abuse policy. Bing says large-scale content generated without oversight, quality control, or editorial review may be excluded from indexing (opens in a new tab). Evidence: strong (official guidance).
11. Measure changed pages against pages you did not change. In the field study above, total ChatGPT referrals grew 5.7 times while untreated pages on the same site grew 3.5 times (opens in a new tab), so a simple before-and-after comparison would have credited platform growth to the changes. Track AI visibility as a rate across repeated runs. Evidence: strong for the method.

The AEO checklist turns these steps into checks with pass criteria.

What the evidence shows and does not show

Antipatterns

An antipattern is a practice that looks helpful but fails or backfires. The full entries, with a test to detect each one, are in Answer engine optimization antipatterns and Generative engine optimization antipatterns.

1. Opening with a story, a history, or a pitch. Rerankers scored early answers higher (opens in a new tab), and Bing asks for essential information near the top (opens in a new tab). Instead, answer first.
2. Running every page through an AI rewriting tool. Rewrites degraded retrieval in a laboratory pipeline (opens in a new tab) and can be detected (opens in a new tab). Instead, edit for relevance and facts.
3. Adding statistics or quotations for their own sake. Adding statistics lowered rankings in 19 of 24 settings (opens in a new tab). Instead, add numbers only when they answer the question.
4. Publishing a page for every phrasing. Google treats this as scaled content abuse when the aim is to manipulate rankings or AI responses (opens in a new tab). Instead, answer close variants on one page.
5. Hiding answers in tabs, accordions, PDFs, images, or scripts. AI crawlers did not render JavaScript (opens in a new tab) in Vercel's data, and Microsoft advises against hidden answers (opens in a new tab). Instead, put the answer in visible HTML text.
6. Stuffing keywords. Google's spam policies define keyword stuffing as spam (opens in a new tab). Instead, use the buyer's words once, in the title, headings, and opening.
7. Refreshing dates without changing the content. Bing asks site owners to use freshness signals appropriately (opens in a new tab). Instead, change the date only for substantive edits.
8. Putting facts only in structured data. Live fetches missed JSON-LD-only facts (opens in a new tab) in one test. Instead, state facts in the text, and mark up only what the page shows.

Checklist

FAQ

What does AEO content look like?

It is a page that answers one buyer question in its first two or three sentences, in the buyer's words, followed by specific facts with their sources, clear headings, and a visible date. The collection of AEO examples shows public pages that follow this pattern. For ChatGPT in particular, see How to get cited and recommended by ChatGPT.

Will AI-generated content hurt my AEO?

We found no study of production answer engines that compares how often they cite human-written and AI-written versions of the same page. Two laboratory studies point away from a penalty for AI wording itself: neural retrieval models tended to rank LLM-generated documents higher (opens in a new tab) (peer-reviewed), and language models preferred options that other language models had described (opens in a new tab) (peer-reviewed). The documented risks are errors, pages that repeat what is already published, and volume. Google asks for "non-commodity content" (opens in a new tab) and warns against generating many pages without adding value, and Bing may exclude large-scale content made without editorial review. Our recommendation: use AI for research and structure, and have a named person check every claim.

Should I break my content into chunks for AEO?

Not into separate small pages. Google says there's no requirement to break content into tiny pieces, because its systems can understand several topics on one page and show the relevant piece (opens in a new tab). Microsoft's assistants parse pages into smaller pieces (opens in a new tab). Sections that stand on their own under descriptive headings suit both.

How long should an AEO page be?

Long enough to answer the question fully, and no longer. Google says there is no ideal page length (opens in a new tab). In one descriptive study, the cited pages that most shaped answers tended to be longer (opens in a new tab), but in a laboratory pipeline a rewriting method that lengthened pages lost retrieval ranks (opens in a new tab). Coverage of the question matters; length alone does not.

Do I need FAQ sections or FAQ schema?

Keep a visible FAQ section where it answers real buyer questions, because readers use it. Google stopped showing FAQ rich results on May 7, 2026 (opens in a new tab), and Google says no special schema.org markup is needed for its generative AI features (opens in a new tab).

How do I update existing pages for answer engines?

Start with the pages buyers already reach. Rewrite the opening so it answers the page's question, update facts and dates, add missing specifics, and remove contradictions with other pages. Keep the address: Bing asks site owners to avoid unnecessary URL changes and to use proper redirects when a change is required (opens in a new tab). Track the changed pages against pages you left alone.

Does the same content work for SEO and AEO?

Mostly, yes. Google says that optimizing for generative AI search "is optimizing for the search experience, and thus still SEO" (opens in a new tab). The differences are in emphasis and measurement, as set out in AEO vs SEO.

How long until answer engines cite a new page?

No study has measured it.

Next steps

AEO HQ's fixed-price packages do not include writing. The AEO Blueprint ($4,995) includes a content strategy and answer-page architecture, and ongoing content work is sold as a monthly retainer. See what each AEO service includes.

Change log

• September 28, 2026: First published.