Answer engine optimization (AEO): the complete guide

Answer engine optimization (AEO) is the practice of making a company's pages and facts easy for AI answer engines to find, cite, and recommend. The working definition in this guide, based on the research it reviews, is that AEO is search engine optimization (SEO) plus four additions: third-party corroboration, specific and verifiable facts, measurement that treats AI answers as distributions, and offers that software can read.

This guide is the hub for AEO HQ's guides on answer engines. It explains how answer engines choose sources, which tactics hold up in independent research, and how to plan and measure the work. Every factual statement links to its primary source, with evidence strength labeled where it matters. The evidence dates from September 2023 to September 2026. AEO HQ calls this model SEO+, or AI search optimization.

Scope and definitions

This guide is for people at B2B companies who are new to AEO, such as marketers, founders, and revenue teams. It covers unpaid visibility in AI answers, not advertising.

Terms used in this guide:

• Answer engine. A system that answers a question in its own words and names or links its sources. Examples include ChatGPT search, Google's AI Overviews and AI Mode, the Gemini app, Perplexity, Claude, and Microsoft Copilot.
• Retrieval. The step in which an engine fetches candidate pages from a search index while it builds an answer.
• Grounding. Basing an answer on retrieved sources rather than on the model's memory alone.
• Retrieval-augmented generation (RAG). The general design in which a language model retrieves documents and then writes its answer from them.
• Parametric knowledge. What a model learned during training and can recall without searching.
• Query fan-out. Splitting one question into several related searches.
• Entity. A distinct thing that search systems can identify, such as a company, a person, or a product.
• Structured data. Code in a page, usually JSON-LD using the schema.org vocabulary, that labels the page's facts for software. It is also called schema markup.
• Citation and recommendation. A citation is a source that an answer links or names. A recommendation is a company that an answer suggests as an option. The two are measured separately.
• Earned media. Coverage that a company does not control, such as reviews, comparisons, and editorial articles.
• Antipattern. A common practice that looks reasonable but fails or causes harm.

Related terms:

• Generative engine optimization (GEO) has the same goal. It is the term used in much of the academic research cited here. AIO and LLMO are other labels in use. AEO vs GEO (and AIO, LLMO) compares the terms, and the companion guide on generative engines reviews that research in depth.
• "AEO optimization" is a common search phrase for the same practice. The O in AEO already stands for optimization.

Guides in this hub

• What is answer engine optimization (AEO)?: a short definition.
• AEO vs SEO: what is the difference?: how the two overlap and differ.
• AEO vs GEO (and AIO, LLMO): are they the same thing?: the competing names.
• How to write content for answer engines: page structure and wording.
• Answer engine optimization examples: real, public, cited examples.
• AEO checklist: every check, with pass criteria.
• Answer engine optimization antipatterns: common mistakes and how to detect them.
• Does schema markup help AEO? What the evidence says: controlled studies of markup and citations.
• Featured snippets and People Also Ask in the AI era: Google's older answer features.
• AEO for B2B SaaS companies: applying AEO to software companies.
• AEO for HubSpot users: running and measuring AEO with HubSpot.

Why AEO matters to B2B companies

Many buyers begin with an assistant: 51% of B2B software buyers now start their research with an AI chatbot more often than with Google (opens in a new tab) (1,076 respondents, March 2026; vendor survey). On Google, people click less when an AI summary appears. In browsing data from 900 U.S. adults (March 2025), users clicked a traditional search result on 8% of visits when an AI summary appeared, compared with 15% when none appeared (opens in a new tab).

Buyers also check what an assistant tells them. In a January 2026 survey of 1,862 technology buyers, 94% of those who use AI said they fact-check its answers at least some of the time (opens in a new tab) (vendor survey). In a March–April 2026 survey of 519 U.S. B2B professionals who use AI at work, 71% said they visit a vendor's website after an AI names the vendor (opens in a new tab), and 27% said AI vendor recommendations don't reflect real pricing or contract structures (opens in a new tab) (vendor survey). These results suggest that the vendor's own page is where an AI recommendation is confirmed or dropped.

How answer engines retrieve, select, and cite sources

The five stages below simplify a process that differs by engine.

1. The engine decides whether to search

An engine can answer from parametric knowledge or search the web first. Anthropic's documentation says Claude searches when a request depends on current information, including "information about specific organizations, people, or products that might have changed" (opens in a new tab) (official documentation). A company newer than a model's training data can appear only through search. In a test of 112 startups from the 2025 Product Hunt leaderboard, a model without web access surfaced 5.4% of them at least once, and a search-augmented model surfaced 27.7% (opens in a new tab) (preprint).

2. The engine rewrites the question

Google says its AI features may use "query fan-out," issuing multiple related searches across subtopics and data sources (opens in a new tab) (official documentation). OpenAI says that when ChatGPT search uses outside search providers, it typically rewrites a prompt "into one or more targeted queries" that it sends to them (opens in a new tab) (official documentation). Answers can therefore cite pages that rank for sub-questions, not only pages that rank for the prompt as typed. In a December 2025 study, AI Overviews and Google's organic results shared only 18% of their sources (opens in a new tab) (peer-reviewed). In another study, 53% of the domains that AI Overviews consulted were outside the organic top 10 (opens in a new tab) (preprint).

3. The engine retrieves candidates from its index

This is the gate. Each engine fetches pages with a crawler, a program that downloads web pages, and stores them in an index. A page that is missing from that index, or that blocks the crawler in robots.txt (the file that tells crawlers what they may fetch), cannot be cited there. The sources in this table are all official documentation:

Two more conditions apply. Facts must be in the HTML the server sends: in data Vercel published in December 2024, none of the major AI crawlers rendered JavaScript (opens in a new tab) (hosting platform measurement). And host firewall rules must not block the crawlers that robots.txt allows.

4. The engine ranks the candidates

A reranker (a model that reorders retrieved pages by relevance) and the language model then choose which pages to use. Four findings are the most consistent:

• Position comes first. In a NeurIPS 2025 benchmark, moving a document into the top positions of the model's context gave far larger gains than any content change, and traditional SEO was more effective than the conversational-SEO methods tested (opens in a new tab) (peer-reviewed).
• Relevance beats decoration. In an ACL 2024 study, relevance to the question drove which passages models found convincing, while adding scientific references or a neutral tone had neutral to negative effects (opens in a new tab) (peer-reviewed).
• Early answers rank higher. In a full retrieve, rerank, and generate pipeline, placing the answer early in a document raised its reranking score (opens in a new tab) (peer-reviewed).
• Specific facts help; formatting alone does not. In 252,000 controlled trials across six models, being on topic, stating a price, carrying a recent rather than an old date, and being listed first raised the odds of being cited first in all six models, while structured versus dense formatting had no consistent effect (opens in a new tab) (peer-reviewed; the authors work for a marketing software company, and the test gave models the full text, so it skipped retrieval).

Brand familiarity also matters. In one study, a model was "minimally influenced by retrieved documents" and leaned heavily on what it already knew about product names (opens in a new tab) (peer-reviewed). On a first answer to a generic prompt, household brands appeared 73% of the time, mid-market brands 44%, and small brands 11% (opens in a new tab) (preprint; the author co-founded a visibility-tracking company).

5. The engine writes the answer and cites sources

Answers lean on third-party sources. In tests through the engines' developer interfaces (APIs), ChatGPT drew 93.5% of its cited sources from earned media for well-known brands and 95.1% for niche brands (opens in a new tab) (preprint). In a 2026 dataset of 149,912 citations, 2.9% pointed to the tracked brand's own domain (opens in a new tab) (preprint).

Being cited is not the same as being described correctly. In a 2023 audit of four engines, 51.5% of generated statements were fully supported by their citations, and 74.5% of citations supported the statement they were attached to (opens in a new tab) (peer-reviewed; the engines tested have since changed). In March and April 2026, 11.0% of the claims in Google AI Overviews were not supported by the pages they cited (opens in a new tab) (preprint). A company therefore needs to check what answers say about it, not only whether it appears.

What AEO adds to SEO: the SEO+ model

SEO is the foundation, because retrieval is the gate and rank carries into the answer. Google states that optimizing for generative AI search "is optimizing for the search experience, and thus still SEO" (opens in a new tab) (official documentation). Bing says that "Bing and Copilot search experiences rely on the same core crawling, indexing, and ranking foundation as traditional search" (opens in a new tab) (official documentation). AEO vs SEO compares the two disciplines in detail.

Four additions sit on top of that foundation:

Third-party corroboration

Brand mentions across the web track visibility in AI answers more closely than links do. Across 75,000 brands, branded web mentions correlated at 0.66 to 0.71 with visibility in ChatGPT, AI Mode, and AI Overviews, YouTube mentions at about 0.74, and link metrics only weakly (opens in a new tab) (vendor study; correlation, not cause). For newly launched startups, discovery on Perplexity correlated with referring domains (the number of sites linking to a company) and Reddit presence, while a page-level GEO score did not predict it (opens in a new tab) (preprint). Google warns that "seeking inauthentic 'mentions' across the web isn't as helpful as it might seem" (opens in a new tab) (official documentation). Corroboration has to be earned.

Specific, verifiable facts

In the controlled trials described above, price raised the odds of being cited first in all six models, and specifications, comparisons, confident wording, and consistent claims did so in at least four (opens in a new tab) (peer-reviewed). In another lab study, a fictional brand beat a well-known one half the time when it had a 0.075-star rating edge, 1.6 times the reviews, or a 7.3% lower price (opens in a new tab) (preprint). Buyers then check those facts on the vendor's site, as the surveys above show.

Measurement as distributions

When the same prompt was repeated, ChatGPT and Google's AI returned the same list of brands less than once in 100 runs, and Claude only slightly more often (2,961 runs, November–December 2025) (opens in a new tab) (industry study; one co-investigator works for a tracking vendor). Visibility is therefore a rate across many runs, not a rank.

Machine-readable offers

A machine-readable offer is one that software can read and act on: its facts are in the HTML the server sends, structured data repeats them exactly, and an agent can complete checkout up to the buyer's approval of payment.

• Perplexity says it is "more likely to recommend merchants who provide deeper product details such as availability, reviews, pricing, and specifications" (opens in a new tab) (official documentation).
• In a practitioner test of five AI systems, none found a price that appeared only in JSON-LD on the test page (opens in a new tab) (a single test page). Structured data should repeat the visible text, not replace it.
• Buying inside an assistant is limited. OpenAI stopped offering in-chat Instant Checkout to new merchants on March 24, 2026; purchases that start in ChatGPT are completed on merchants' own sites (opens in a new tab). Google's Universal Commerce Protocol (UCP) checkout, which serves AI Mode in Google Search and the Gemini app (opens in a new tab), lists services such as lessons and online classes, subscriptions, and bundled setup services as ineligible (opens in a new tab). ChatGPT's cloud browser asks the user to confirm before a payment (opens in a new tab) (all official documentation).
• Stripe describes how a business can let AI agents book and pay for its services (opens in a new tab), but its agent payments need the customer to approve each spend request and are available to U.S. and Canadian consumers (opens in a new tab) (official documentation).

For a B2B service, this means a plain-text offer page and a checkout an agent can reach. The guide to how AI agents find and buy services covers the protocols.

An AEO strategy in eight steps

These steps are recommendations. Each notes the strength of the evidence behind it.

1. List the questions buyers ask, in their words. Collect 20 to 50 questions a buyer might put to an assistant: definitions, comparisons, prices, and "who offers this" questions. In lab trials, query terms that were present rather than missing raised citation odds (opens in a new tab), and replacing plain words with technical terms caused the largest retrieval drops (opens in a new tab) (both peer-reviewed). Evidence: strong.
2. Make every important page retrievable. Get core pages indexed in Google and Bing. Allow the crawlers in the table above, check that firewall rules do not block them, and serve the facts in server-rendered HTML. IndexNow is a protocol that tells search engines when a URL changes; Bing asks site owners to submit changed URLs through it and keep sitemap dates accurate (opens in a new tab). Evidence: strong (official documentation).
3. Make the company and its people unambiguous. Use one name and one description on the site, profiles, directories, and structured data. Give each author a profile page, and connect their other profiles with sameAs markup, which points to the same person's other profiles. Language models often merge information about different entities that share a name (opens in a new tab) (peer-reviewed), and Google asks for an author URL that uniquely identifies the author, and accepts sameAs links, to tell authors apart (opens in a new tab) (official documentation). Evidence: moderate for the risk; weak for the effect on recommendations.
4. Publish the facts buyers check. State price, scope, deliverables, turnaround, and terms as plain text. Keep them identical on every page and profile. Evidence: moderate.
5. Write one page per distinct question, with the answer first. Open with a two- or three-sentence answer that stands on its own. Do not generate a page for every phrasing: Google says that 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 policy (opens in a new tab) (official documentation). How to write content for answer engines explains the method, and the collection of AEO examples shows it in use. Evidence: moderate.
6. Earn third-party corroboration. Get described accurately on independent sources: third-party review platforms, lists that publish their method, trade press, podcasts, and communities where the company takes part openly. Do not buy placements or post from fake accounts. Evidence: moderate (correlational).
7. Make the offer readable and purchasable by software. Keep the purchase path free of login walls and bot challenges. Stripe asks agents to report what stopped a purchase, with tags that include CAPTCHAs, anti-bot scripts, firewall blocks, and required logins (opens in a new tab) (official documentation). Evidence: strong for the platform facts; unproven for recommendations.
8. Measure, change one thing, and compare with a control. In the only controlled field study found, ChatGPT referrals to pages that were not changed grew 3.5 times over the same period (opens in a new tab) (preprint). A before-and-after comparison without a control would have credited that growth to the changes. Evidence: strong for the method.

AEO best practices

These page-level practices are recommendations.

• Match the buyer's wording in the title, H1, URL, and first paragraph. In a lab pipeline, optimizing titles, meta descriptions, headings, and schema fields raised the retrieval hit rate by 22% (opens in a new tab) (peer-reviewed; the retriever matched keywords).
• Answer first. The one controlled field study combined question-form titles and standalone two- to three-sentence summaries with other changes, and measured 1.82 times more ChatGPT referrals (95% confidence interval 1.31 to 2.54), but a placebo test did not rule out chance (p = 0.16) (opens in a new tab) (preprint; the authors work for the site studied). Treat that result as suggestive.
• Use headings, lists, and tables. Microsoft says assistants such as Copilot parse pages into "smaller, structured pieces" and advises against hiding key answers in tabs or expandable menus, or leaving them only in PDFs or images (opens in a new tab) (official guidance). Structure helps engines find content; the lab trials above found no consistent effect on which source a model cites once it has the full text.
• State claims with confidence, and keep them consistent. Confident rather than hedged wording, and consistent rather than contradictory claims, raised citation odds (opens in a new tab). A promotional tone does not help: the original GEO study found no significant improvement from a persuasive, authoritative tone (opens in a new tab) (both peer-reviewed).
• Keep content current and dates honest. A 2026 date beat a 2019 date in all six models tested, but dated and undated pages showed no consistent difference (opens in a new tab) (peer-reviewed). Change the visible date only when the content changes.
• Treat structured data as hygiene. In a matched study of 1,885 pages that added schema, AI citations did not rise: AI Overviews citations fell 4.6%, and the changes in AI Mode and ChatGPT were indistinguishable from zero (opens in a new tab) (vendor study; every page was already heavily cited). Google says that no special schema.org markup, AI text files, or other machine-readable files are needed for its AI features (opens in a new tab) (official documentation). Mark up only what the page visibly shows.

How to measure AEO

Measure AI visibility as a rate across repeated runs, per engine, with an error range. Do not report a single answer as a rank. AEO metrics and KPIs defines each metric. The research reviewed for this guide found no rigorous study of how long a new company takes to appear in AI answers, so this guide gives no timeline.

A prompt panel is a fixed set of prompts run on a schedule. These six practices for running one are recommendations:

1. Fix the panel. Use the buyer questions from step 1, with three to five phrasings each, on the engines your buyers use. A survey of 45 studies recommends 7 to 8 repetitions per prompt, 3 to 5 paraphrases, several named engines, and multiple time windows (opens in a new tab) (preprint).
2. Repeat each prompt. A standard error is the typical size of the error in an estimate. In one panel, one run per prompt gave a standard error of 0.370 for a brand's detection rate, seven runs gave 0.081, and eight gave 0.062 (opens in a new tab). The same panel found that cited-source sets overlapped only 32% to 43% between runs, and about 65% of sources turned over from one day to the next (opens in a new tab) (preprint; the first author is affiliated with a vendor in this field).
3. Add prompts before adding runs. Answers to the same prompt are correlated. Repeating a prompt reduces only the variance within that prompt, and clustered standard errors can be more than three times larger than naive ones (opens in a new tab) (preprint).
4. Test the consumer interfaces. API results shared only 12.0% (ChatGPT) and 14.8% (Gemini) of cited domains with the consumer interfaces (opens in a new tab) (preprint). API runs are rough proxies.
5. Use intervals that suit small samples. Normal-approximation intervals are too narrow below a few hundred data points, where nominal 95% intervals covered 92.5% of cases at 100, and the authors recommend Wilson or Bayesian intervals instead (opens in a new tab) (peer-reviewed).
6. Report four rates per engine: the share of runs that name the company, the share that cite its pages, its share of voice (its mentions as a share of all mentions of a set of named competitors), and the share of answers that describe it correctly.

First-party reports fill in the rest (all official documentation):

• Google Search Console's generative AI performance report counts impressions in AI Overviews and AI Mode (opens in a new tab). It reports impressions only and has been available worldwide since August 31, 2026.
• Bing Webmaster Tools' AI Performance report shows citations, cited pages, and grounding queries for Copilot and Bing's AI summaries (opens in a new tab). Grounding queries are the key phrases the AI used when it retrieved the cited content. The report does not show clicks.
• In GA4, the default AI Assistant channel excludes Google's AI Overviews and AI Mode, which count as Organic Search (opens in a new tab). ChatGPT adds utm_source=chatgpt.com to referral URLs (opens in a new tab).

AI referral traffic is small and hard to attribute. Across 973 e-commerce sites from August 2024 to July 2025, ChatGPT referrals were under 0.2% of traffic (opens in a new tab), and they converted below organic search, paid search, email, and affiliate traffic, and above paid social only (opens in a new tab) (peer-reviewed; the data came from the first author's employer). First-touch attribution credits a sale to the first channel recorded for the buyer. In one agency's records, first-touch attribution credited AI with only 28 of the 189 leads that named an AI tool when asked how they heard of the agency (opens in a new tab) (a single company). Asking buyers where they heard of you captures influence that referrer data misses.

What replicates and what does not

The table grades common AEO tactics. "Strong" means official documentation or several independent studies agree. "Moderate" means consistent evidence that is mostly lab-based, correlational, or thin. "Contested" means credible studies disagree.

Why the "GEO +40%" result does not generalize

The most-quoted number in this field comes from the original 2024 GEO paper. In that study, adding quotations raised a source's share of the generated answer by about 41% and adding statistics by about 31%, in a simulated engine where the page was already retrieved (opens in a new tab) (peer-reviewed). The result is real, but three conditions limit it:

1. The engine was a simulation. It answered with GPT-3.5 from the full text of the top five Google results (opens in a new tab), not through a production answer engine.
2. Retrieval was assumed. The page was already among the sources, and the authors did not test effects on retrieval or ranking (opens in a new tab). When a later study ran a full retrieve, rerank, and generate pipeline, rewriting page text to optimize it for generative engines hurt retrieval (opens in a new tab) (peer-reviewed).
3. The metric was share of words, not choice. The study measured how much of an answer's text came from a source. When a 2025 benchmark measured citation rank across four models, it found statistically significant gains in 3 of 54 cases, and adding statistics lowered rank in 19 of 24 (opens in a new tab) (peer-reviewed).

A critical survey of 45 studies rejects "GEO increases visibility by 40%" as a general claim and finds no technique with a stable, longitudinal, cross-platform causal effect on organic discoverability (opens in a new tab) (preprint). Gains also shrink as competitors copy a tactic: in the 2025 benchmark, the gain per adopter fell toward zero as adoption rose (opens in a new tab). The practical reading is that genuine, query-relevant specifics help, while numbers and quotations added for their own sake do not.

Antipatterns

The full list, with a test to detect each one, is in answer engine optimization antipatterns. The most common are:

• Hidden text or instructions aimed at AI. Google's spam policies cover attempts to manipulate generative AI responses (opens in a new tab).
• A page for every variation of a question. Google treats this as a violation of its scaled content abuse policy when the aim is to manipulate rankings or AI responses (opens in a new tab).
• Rewriting the whole site with a language model to suit engines. It hurt retrieval in a realistic pipeline (opens in a new tab).
• Fake dates, fake reviews, or posts from fake accounts. They misstate the facts that engines and buyers check.
• Promising "guaranteed citations" or "+40% visibility". Google lists "AEO" and "GEO" tools among the offers to be wary of, and says third-party tools "can't guarantee performance" (opens in a new tab).
• Reporting one answer as a rank. ChatGPT and Google's AI returned the same brand list less than once in 100 runs (opens in a new tab).

AEO checklist

The AEO checklist gives each check with how to verify it, the pass criteria, and the source. The short version:

• Core pages are indexed in Google and Bing.
• robots.txt, firewall, and bot-protection rules allow Googlebot, Bingbot, OAI-SearchBot, Claude-SearchBot, and PerplexityBot.
• Search Console's Search generative AI setting includes the site.
• Prices, scope, deliverables, and timelines appear in server-rendered HTML.
• Facts match across the site, structured data, and third-party profiles.
• Each priority page opens with a standalone answer in the buyer's words.
• Each author has a profile page, linked to other profiles with sameAs.
• Visible dates change only when the content changes.
• No page has hidden text, instructions aimed at AI, or templated variants.
• A fixed prompt panel runs on a schedule and reports rates with intervals for each engine.

Frequently asked questions

Is AEO the same as GEO?

In practice, yes. Both aim to get a company cited and recommended in AI-generated answers. GEO is the term used in much of the academic research.

Does AEO replace SEO?

No. Answer engines retrieve from search indexes, so a page must first be crawlable, indexed, and relevant. Google describes optimizing for its AI features as "still SEO" (opens in a new tab). AEO adds corroboration, specific facts, measurement, and machine-readable offers to that base.

Do strong Google rankings still influence AI citations in 2026?

Yes, but less directly than in classic search, as stage 2 above shows. Where AI citations do overlap with search results, they skew toward the top-ranked result (opens in a new tab) (preprint). Because engines fan a question out into several searches, ranking for those sub-questions also counts.

Can an agency guarantee my brand appears in ChatGPT?

No. Answers vary too much from run to run for anyone to promise a placement. Google warns that third-party tools "can't guarantee performance" (opens in a new tab). A credible provider reports rates with error ranges and does not promise placements.

Why does ChatGPT recommend my company to me but not to my customers?

Answers vary by run and by user. In one study, signals about who the user was significantly changed the recommendations chatbots gave (opens in a new tab) (peer-reviewed). Test in clean, logged-out sessions, repeat each prompt, and compare rates rather than single answers.

Next steps

AEO HQ sells this work as fixed-price services with published prices: a $499 automated audit, a $2,500 audit, a $4,995 blueprint, and an $8,995 package that adds technical implementation to the blueprint. See what each AEO service includes.

Change log

• September 27, 2026: First published.