Answer engine optimization: how to make your pages easy for AI answers to quote

Answer engine optimization, or AEO, is the work of making your web pages easy for AI answer engines to find, understand, and quote when someone asks them a question. Answer engines are the tools that reply with a written answer instead of a list of links: ChatGPT, Gemini, Perplexity, Claude, Microsoft Copilot, and the AI Overviews and AI Mode in Google Search. When one of them answers a buyer's question in your category, it builds the reply from a few passages on a few pages and credits those pages as its sources. AEO is how you make one of those passages yours.

This matters because the written answer is often where the reader stops. Pew Research Center tracked the browsing of 900 U.S. adults in March 2025 and found that when a Google search showed an AI summary, people clicked a regular search result on 8 percent of visits, against 15 percent when no summary appeared. They clicked a link inside the summary itself on just 1 percent of visits (Pew Research Center). If the answer at the top doesn't mention you, a good ranking underneath it is worth less than it used to be.

Most AEO advice arrives as a long checklist. This guide takes a narrower view, built on what the answer engines' own documentation and the published studies actually show. Answer engines find pages through search, and then they quote passages rather than whole pages. For most sites, that makes AEO an editing job on pages you already have: putting the answer where an engine looks for it, in a sentence that still makes sense when it's lifted out of the page. The sections below cover the names, how the engines choose what to quote, the page structures that get quoted with before-and-after examples, and which popular tactics the evidence doesn't support.

SEO for AI goes by several names, most often generative engine optimization (GEO), answer engine optimization (AEO), LLM SEO, and AI SEO, and they all describe the same work of getting AI answers to find, cite, and recommend your business. Wikipedia's entry treats them as one practice under several labels and notes that usage varies across practitioners, vendors, and publications "with no consensus definition" (Wikipedia).

Each name came from a different corner of the industry:

• Generative engine optimization (GEO) comes from a 2023 research paper by Pranjal Aggarwal and colleagues at Princeton and IIT Delhi, later presented at the KDD 2024 conference (Wikipedia). That paper's experiments are still among the few controlled tests of what makes a page more visible in AI answers, and this guide draws on them below.
• Answer engine optimization (AEO) is the older label. SEO practitioners used it for featured snippets and voice assistants before ChatGPT arrived, then stretched it to cover AI chat answers (Outrigger). HubSpot uses AEO as its umbrella term for all of this work (HubSpot).
• LLM SEO, sometimes written LLMO for large language model optimization, frames the same work around the language model inside the answer engine and how it represents your brand (Outrigger).
• AI SEO is the plain-language umbrella, meaning simply "SEO, but for AI." Semrush markets its tools under this name (Wikipedia).

Some companies draw lines between the terms, but they draw them in different places. Semrush says AEO covers getting your existing pages pulled into AI answers as people ask questions, while GEO covers how a brand appears in what a model already learned from its training data (Semrush). HubSpot says AEO aims at a direct answer in one place, such as an AI Overview or a voice assistant, while GEO aims at broader visibility across many AI sources (HubSpot). Each of these companies sells tools under its preferred name, so the label tells you less than the method behind it. Outrigger, which sells AI visibility measurement, gives buyers the same advice: ask a vendor "what signals they move and how they measure it" (Outrigger).

Most marketers sidestep the debate. When the growth agency Fractl surveyed 343 U.S. marketing decision-makers, 81 percent said they still call their AI search work SEO. The new terms show up mostly among executives: 28 percent of C-suite marketers said they use GEO and 17 percent said AEO, roughly three times the rate of individual contributors (Search Engine Land). This guide uses AEO, but everything in it applies whichever name your team prefers.

AEO builds on SEO, because the major answer engines find the pages they quote through a search index, and a page that can't be found there never gets the chance to be quoted. Google says its AI Overviews and AI Mode are "rooted in our core Search ranking and quality systems" and retrieve "relevant, up-to-date web pages from our Search index" before they write a response (Google Search Central). Microsoft says Copilot is powered by Bing's search index (Microsoft Advertising). OpenAI says ChatGPT search rewrites your question into one or more targeted queries and sends them to its search providers (OpenAI). Ahrefs reports that 88 percent of the URLs ChatGPT cites come from its general search index, though its guide doesn't describe how that was measured (Ahrefs).

Google goes further and treats the whole subject as SEO. Search Engine Journal reports that Google's guide to generative AI search defines AEO and GEO and then states that "optimizing for generative AI search is optimizing for the search experience, and thus still SEO" (Search Engine Journal). Google runs the search engine these features sit in, so it has a stake in keeping site owners on its existing guidance. Forrester analyst Nikhil Lai made a similar point from the outside in 2025, as Wikipedia records, arguing that advocates of the new terms "tend to exaggerate SEO and AEO's differences to carve a startup-sized hole in marketers' tech stacks" (Wikipedia).

Ranking well and being quoted are related, though, and the gap between them is why AEO has a name of its own. An answer engine picks passages, not positions in a list. In the GEO paper's tests, adding cited sources to a page's text helped low-ranked pages far more than top-ranked ones: visibility in the AI answer rose 115.1 percent for the page ranked fifth in Google and fell 30.3 percent, on average, for the page ranked first (Aggarwal et al.). Semrush's comparison of AI answers with Google results found that AI often cites deeper pages on trusted sites rather than the exact pages that rank at the top (Semrush). A useful way to hold the two together is that SEO gets your page into the set of pages the engine reads, and AEO shapes whether the engine quotes your page once it's there. For a side-by-side look at measuring both, see SEO vs GEO: How to See Where You Rank on Google and Where AI Answers Mention You.

An answer engine runs its own searches, reads the pages that come back, and builds its reply from short passages, so the passage that wins is one that answers the question completely on its own. Google calls the first step query fan-out: the model generates "a set of concurrent, related queries" to fetch more results. For someone asking "how to fix a lawn that's full of weeds," Google says the fan-out searches might include "best herbicides for lawns," "remove weeds without chemicals," and "how to prevent weeds in lawn" (Google Search Central). OpenAI describes the same pattern in ChatGPT. A question about cancer drugs that target CCR8 might first go out as "CCR8 immunotherapy drug development 2025," followed by a narrower search such as "CHS-114 conference 2025" once ChatGPT has read the first results (OpenAI).

Fan-out means your page can be found through a follow-up question the buyer never typed. A page that answers the main question and the obvious next questions, each in its own section, gives the engine several ways to reach it.

Once the engine has the pages, it works with pieces of them rather than the whole page. Krishna Madhavan, a product manager at Microsoft Bing, writes that assistants like Copilot break content into smaller pieces, "a process called parsing," evaluate each piece for authority and relevance, and assemble the best ones into an answer (Microsoft Advertising). Anthropic's documentation for Claude's web search tool shows how small those pieces can be: each citation carries the quoted text, up to 150 characters of it, which is about one sentence (Anthropic).

Two independent studies show where on a page those quoted sentences tend to come from:

• Near the top. Kevin Indig, a growth advisor, matched ChatGPT's answers to the source sentences they drew on and found that 44.2 percent of citations came from the first 30 percent of a page, 31.1 percent from the middle, and 24.7 percent from the final third (Search Engine Land). Daniel Shashko of Bright Data studied 42,971 citations and found that in Google's AI Mode and Gemini, 46.1 percent came from the first 30 percent of a page and 74.8 percent from the first half (Practical Ecommerce).
• Whole, short sentences. In Shashko's data for AI Mode and Gemini, every citation was a complete sentence, and sentences of 6 to 20 words made up 92.4 percent of them (Practical Ecommerce).
• Definitions and questions. In Indig's data, cited passages were nearly twice as likely to use definitions such as "X is" or "X refers to," and 78.4 percent of citations tied to questions came from headings. Indig reads this as the engine treating a heading as the prompt and the paragraph under it as the answer (Search Engine Land).
• Named things. Typical English text is 5 to 8 percent proper nouns, and the heavily cited text in Indig's study averaged 20.6 percent. Names of products, companies, and tools anchor an answer (Search Engine Land).

One finding cuts against a popular rule. Within a paragraph, Indig found that 53 percent of citations came from the middle sentences, and only 24.5 percent from the first (Search Engine Land). The answer belongs near the top of the page and at the start of each section, but you don't need to force a quotable line into the opening of every paragraph. Clear sentences throughout matter more.

The pages AI answers quote tend to put the reader's question in a heading, answer it completely in the first sentence under that heading, name their subject in every passage, back the answer with specific facts and sources, and use a table or list where the question calls for one. The examples below come from people who rewrote real pages, along with a few illustrations from Microsoft and Kite.

A heading phrased the way a buyer asks the question gives the engine a direct match for the prompt. Microsoft's example replaces a vague heading, "Learn More," with "What Makes This Dishwasher Quieter Than Most Models?" (Microsoft Advertising). Roman Makarenko, who runs the agency Thunder Marketing Solutions, made the same change across a healthcare client's articles. "Managing Symptoms" became "What Are the Most Effective Ways to Manage [Condition] at Home?", and his rule was that a question 2,900 people search each month earns its own heading (Makarenko on Medium).

The sentence right after the heading should carry the whole answer, so it still works when an engine quotes it alone. Makarenko's before-and-after shows the difference. The original opening read: "[Condition] affects millions of people worldwide. Understanding the various factors that contribute to this condition can help you make informed decisions…" The rewrite opened with the answer: "You should see a doctor for [condition] when symptoms persist for more than two weeks despite home treatment, interfere with daily activities, worsen suddenly, or include warning signs like chest pain, difficulty breathing, or severe headaches." Supporting detail followed in the next two or three sentences (Makarenko on Medium).

Sometimes the fix is adding an answer that isn't on the page at all. Ahrefs noticed that Google's AI Overview gave a passable but indirect answer to "what does Ahrefs stand for," probably because no page on its site answered that exact question. It proposed adding a short question-and-answer block to the page the overview already cited: "Ahrefs stands for 'a href,' the HTML tag that creates hyperlinks in code" (Ahrefs). Semrush suggests keeping that first sentence under roughly 40 words (Semrush). Neither company cites a test for the exact length, so treat it as a guide to brevity rather than a limit.

A quoted passage travels without the rest of the page, so it should name its subject instead of saying "it," "this," or "as mentioned above." Microsoft's checklist lists it as "Self-contained phrasing: Sentences that make sense even when pulled out of context" (Microsoft Advertising). When the agency Local SEO Services rebuilt ten pages for a client, it rewrote every FAQ answer to make complete sense alone, in three to five sentences, with no references back to earlier sections (New Techno Times).

Naming the subject also puts your brand inside the quote. Kite's page on this work uses a made-up dental scheduling company to show the idea. The heading "Built for busy front desks," followed by three paragraphs about the team's mission, becomes "How long does it take to set up Brightpath?", and the first sentence becomes "Most dental clinics set up Brightpath in one afternoon, and it imports existing patient appointments from Dentrix and Open Dental." That sentence answers the question and carries the product's name wherever it's quoted (Kite).

Specific facts give an engine something concrete to repeat, and vague claims give it nothing. Local SEO Services replaced lines such as "We offer prompt service" with checkable ones such as "we typically respond to new inquiries within one business day," and cut any claim it couldn't back with a real number (New Techno Times).

The GEO paper tested this kind of change directly. Its authors had a language model rewrite source pages in nine ways and measured how visible each page became in AI answers. Adding citations to credible sources, adding quotations, and adding statistics each raised visibility by 30 to 40 percent on one of the paper's two measures. One example added a survey attribution to a sentence about Swiss chocolate consumption and raised that source's visibility by 132.4 percent. Keyword stuffing, a classic SEO tactic, gave "little to no improvement," and on Perplexity it scored about 10 percent worse than leaving the page alone (Aggarwal et al.).

The same property that makes a specific claim easy to quote makes a wrong one spread. Onplana, a project management software company, describes an agent that drafted up to three blog posts a day and committed them to a staging site before anyone read them. The build included a check meant to catch claims the team already knew were false, but that check had never actually run. When the team switched it on, it found the same false claim, about admin-configurable AI guardrails the product doesn't have, in six posts that all cited one page that had already been corrected (Onplana). The setup explains the result: the agent copied a fact from a source that had since changed, and the one safeguard against that wasn't connected. Whoever writes your quotable sentences, person or agent, should take every fact from a current source and check it against that source before it goes live.

Comparisons read best as tables, processes as numbered steps, and groups of options as lists, and the engines quote those blocks cleanly. Microsoft's example sets a long paragraph about a dishwasher's features against a three-bullet list: "42 dB noise level," "Energy Star certified," and "Compatible with Alexa and Google Home" (Microsoft Advertising).

Lilach Bullock, a marketing consultant, tracked 50 pages on her own site against the same 12 prompts in Perplexity every week for 90 days. Only 9 pages were cited at least once, and 7 of those 9 had a numbered list or table within the first 400 words. Pages with long narrative openings "almost never got pulled" (Lilach Bullock). AirOps, which sells AI search content software, compared pages ChatGPT cited with Google page-one results it didn't cite and found that 68.7 percent of the cited pages used headings in strict order with no skipped levels, against 23.9 percent of the others (AirOps). Both are small or correlational findings rather than controlled tests, but they point the same way as the citation-position studies.

Google says you don't need to break pages into tiny pieces or write a special version for AI, and the structures above don't ask you to, because they're the same habits that make a page easy for a person to scan. Google's guide lists "chunking" content among the things you can ignore, says "There's no ideal page length," and says "You don't need to write in a specific way just for generative AI search" (Google Search Central). Google's Danny Sullivan put it more bluntly on Google's Search Off the Record podcast: "We really don't want you to think you need to be doing that or produce two versions of your content, one for the LLM and one for the net" (Search Engine Land).

Mike King, who runs the search agency iPullRank, answered that "chunking and creating content for users are not mutually exclusive" (iPullRank). His point is that a passage covering one idea under a clear heading is good writing for people and easier for a retrieval system to match. In his own scoring tool, splitting one paragraph that mixed machine learning and data privacy into two paragraphs raised the machine-learning half's similarity to a "machine learning" query by 19.24 percent. That measures how closely the text matches a query, not whether any engine cited it.

The best controlled test of structure supports King's reading without supporting tiny pieces. Researchers at the University of Tokyo and three other Japanese institutions restructured 200 articles, changing heading levels, paragraph length, lists, tables, and internal links while keeping the meaning, and ran the originals and rewrites through six answer engines. The rate at which queries cited the content rose from 45.0 percent to 52.8 percent. Their target paragraph length was 150 to 300 words, which is ordinary prose rather than bite-sized fragments (Yu et al.). The study is an unreviewed preprint that tested older engine versions with the authors' own rewriting algorithm, so it shows the direction more reliably than the size of the effect.

The practical reading of both sides is the same: write one version of each page, for people, and organize it so any section can be read on its own.

An answer engine can only quote text its crawler can fetch and read, so crawler access and plain HTML text matter far more than schema markup or an llms.txt file, which current evidence shows do little for citations.

AI crawlers. Each AI company runs separate crawlers for search and for training, and you control them separately in your robots.txt file. OpenAI says sites that block OAI-SearchBot won't be shown in ChatGPT search answers, while GPTBot collects content for training, and "each setting is independent of the others." Changes take about 24 hours to apply to search (OpenAI). Perplexity uses PerplexityBot to surface pages in its answers and says it doesn't use it for training (Perplexity). Anthropic runs ClaudeBot for training, Claude-SearchBot for search, and Claude-User for pages fetched when someone asks Claude a question (Anthropic). You can block training and still be quoted, as long as the search crawlers get through. Check your CDN as well as robots.txt: Ahrefs notes that CDNs such as Cloudflare block AI platforms by default (Ahrefs).

JavaScript. Vercel and the consultancy MERJ watched crawler traffic across Vercel's network and found that none of the major AI crawlers they measured run JavaScript, including OpenAI's, Anthropic's ClaudeBot, and PerplexityBot. Gemini is the exception, because it uses Googlebot's infrastructure (Vercel). If your page's main text only appears after scripts run in the browser, most AI crawlers see an empty page. Microsoft adds that answers hidden in tabs or expandable menus may also be skipped (Microsoft Advertising).

Schema markup. Ahrefs tracked 1,885 pages that added JSON-LD structured data between August 2025 and March 2026 and compared them with about 4,000 similar pages that didn't. Over the following 30 days, citations changed by plus 2.2 percent in ChatGPT and plus 2.4 percent in AI Mode, both indistinguishable from no change, and by minus 4.6 percent in AI Overviews. Ahrefs concluded that it "can't tell whether the schema did a tiny bit of good or nothing at all." It also reports a test by searchVIU in which ChatGPT, Claude, Perplexity, Gemini, and AI Mode all ignored JSON-LD when fetching a page live (Ahrefs). Google says structured data isn't required for its AI features but is still worth keeping for rich results (Google Search Central). Accurate schema is good housekeeping. It won't stand in for a clear answer in the text.

llms.txt. An llms.txt file is a plain Markdown list, placed on your site, that points AI tools to your most useful pages. Its proposal describes it as a guide for AI agents that need information while helping a user, especially coding tools reading documentation, and doesn't claim answer engines use it to choose citations (llmstxt.org). Ahrefs checked the 137,210 sites in its web analytics product in May 2026 and found that 97 percent of the valid llms.txt files received no requests at all, from bots or people. Slackbot fetched the files more often than PerplexityBot did (Ahrefs). SE Ranking studied nearly 300,000 domains and found no link between having the file and how often AI answers cited the domain (SE Ranking). Google says its search doesn't use these files (Google Search Central). The file is cheap to publish and may matter more as agents do more browsing, but it isn't what gets a page quoted today. The guide llms.txt: What It Is and How to Create One for Your Website Coming soon covers the file in detail.

The table below sets the common AEO tactics side by side with the strongest evidence for each, so you can see which ones rest on measurement and which rest on opinion.

You know AEO is working when the same buyer questions, asked again in the same engines a few weeks after a change, start quoting your page. Because AI answers shift on their own, you need repeated checks and small batches of changes to tell your work apart from background noise.

The noise is large. Amsive, a search agency, relays a Profound analysis of 80,000 prompts in which the cited sources changed from one month to the next at rates of 59.3 percent in Google AI Overviews, 54.1 percent in ChatGPT, 53.4 percent in Microsoft Copilot, and 40.5 percent in Perplexity (Amsive). In Lilach Bullock's 90-day test, Perplexity changed which source it cited for 4 of her 9 cited pages without any change on her part (Lilach Bullock). Makarenko's healthcare client saw weekly visits from ChatGPT and Perplexity fall from about 600 to 650 in April to about 70 to 120 by mid-June, after a Google core update rather than any change to the pages (Makarenko on Medium).

A sound measurement routine handles that noise in a few ways:

• Fix the question list. Choose the buyer questions you care about and ask the same ones every time, so one round compares fairly with the next.
• Record a baseline first. Note which pages each engine quotes today, yours and your competitors', before you change anything.
• Change a few pages at a time. If you rewrite everything at once, you can't tell which change earned a citation.
• Wait before you recheck. Engines need time to re-read a page, so a check the next day shows little.
• Use the free reports you already have. Google Search Console has a Generative AI performance report for AI Overviews and AI Mode (Google Search Central).

This loop is also what makes AEO a good job for an AI marketer. The work is verifiable: either the engines quote your page for a question or they don't, and each result tells you which page to work on next. Kite is an AI marketer. It researches your business, builds a growth strategy, and ships the work that brings you new customers, and you can work with it in your browser or in Slack. For AEO, Kite proposes the buyer questions worth tracking from what it knows about your products, customers, and competitors, and you can add your own. It asks those questions in ChatGPT, Gemini, Perplexity, and Claude and records whether you're mentioned, which competitors appear, and which pages each answer quotes (Kite).

From there, Kite reads the pages that win the quote and compares them with yours. It runs a 24-check scan of whether AI crawlers can reach and read your site and fixes anything blocking a page before it rewrites, because a page crawlers can't load earns no quotes. Then it rewrites the pages that already cover a question but lose the quote: headings become buyer questions, the first sentence answers in full and names your product, and tables or steps go where the question calls for them. Kite states only facts it can support from your website and its research, and when a page needs one it doesn't have, such as a setup time you've never published, it asks you. It asks for your approval before publishing, then waits two to four weeks, asks the same questions again, and changes a small number of pages per round so each result traces back to a specific rewrite (Kite).

Start by finding out where you stand. Kite's free AI visibility checker asks ChatGPT, Gemini, Perplexity, and Claude the questions your customers ask, and in about 45 seconds it shows whether they recommend your business and who they recommend instead. It's free and needs no account.

From there, the work follows the order of this guide:

1. Pick the buyer questions that matter most and note which pages each engine quotes for them today.
2. Find your pages that already cover one of those questions but lose the quote to a competitor. They're usually the fastest to change, because only the shape of the page is holding them back.
3. Make sure the search crawlers from OpenAI, Perplexity, and Anthropic can reach those pages and that the main text loads without JavaScript.
4. Rewrite each page so its headings ask the buyer's questions and the first sentence under each one answers in full, with specific facts and named sources.
5. Ask the same questions again a few weeks after the changes go live, and use what you learn to choose the next pages.

To see how Kite does this on your own pages, read Make your pages easy for AI answers to quote. For new content written to be quoted from the start, see How to structure a blog post so ChatGPT and other AI assistants quote it. For the wider job of getting your brand named in AI answers, including pitching the outside sites the engines already cite, see Improve your visibility in AI answers and How to get cited by AI answer engines Coming soon .

• Google users are less likely to click on links when an AI summary appears in the results, Pew Research Center
• Generative engine optimization, Wikipedia
• GEO: Generative Engine Optimization, Pranjal Aggarwal and colleagues, arXiv and KDD 2024
• What Is AI SEO Called? GEO, AEO, LLMO and Every Other Name, Outrigger
• What is answer engine optimization (AEO), and how does it change SEO?, HubSpot
• AEO vs SEO: core differences & how to win visibility in both, Semrush
• AI search: What marketers call it (SEO), per a survey, Search Engine Land
• Optimizing your website for generative AI features on Google Search, Google Search Central
• Google's New AI Search Guide Calls AEO And GEO 'Still SEO', Search Engine Journal
• Optimizing Your Content for Inclusion in AI Search Answers, Microsoft Advertising
• Searching the web with ChatGPT, OpenAI Help Center
• What is Answer Engine Optimization (AEO)?, Ahrefs
• Web search tool, Anthropic
• ChatGPT heavily favors the top of content when selecting citations, Search Engine Land
• Studies Reveal AI Citation Clues, Practical Ecommerce
• AI Search Optimization Case Study: From Zero Visibility to 1,631 Citations Across ChatGPT, Perplexity, AI Overview, and Gemini, Roman Makarenko on Medium
• Getting Cited by ChatGPT and Perplexity, New Techno Times
• An AI Agent Content Pipeline Case Study: Six Months In, Onplana
• How to Get Cited on Perplexity in 2026, Lilach Bullock
• Why Ranking on Page One Isn't Enough, AirOps
• Google doesn't want you to create bite-sized chunks of your content, Search Engine Land
• Moving from a Google-shaped Web to an Agent-shaped Web: A Refutation of Misinformation about Chunking, iPullRank
• Structural Feature Engineering for Generative Engine Optimization: How Content Structure Shapes Citation Behavior, Junwei Yu and colleagues, arXiv
• OpenAI Crawlers, OpenAI
• Perplexity Crawlers, Perplexity
• Does Anthropic crawl data from the web, and how can site owners block the crawler?, Anthropic
• The rise of the AI crawler, Vercel
• We Tracked 1,885 Pages Adding Schema. AI Citations Barely Moved., Ahrefs
• llms.txt, llmstxt.org
• We Analyzed 137K Sites: 97% of llms.txt Files Never Get Read, Ahrefs
• Does LLMs.txt impact your AI visibility and citations? No, according to research, SE Ranking
• Answer Engine Optimization (AEO): Your Complete Guide to AI Search Visibility, Amsive
• How much does AEO cost? A breakdown by approach, HubSpot