AEO works by making your pages the easiest source for an answer engine to retrieve, extract and cite. The engine interprets the question, decides whether to search, runs one or more queries, reads the pages that come back, selects passages that answer each part, writes a response and attaches links. AEO targets the retrieval, extraction and citation steps with crawlable, direct, specific and corroborated passages.
This is part two of our series on how search and AI visibility work. Part one covered how pages get into a search index. This part covers what happens when an engine reads those pages and writes an answer. We have deeper pieces on individual engines, including ChatGPT search, Perplexity and Claude, and a marketer's introduction to retrieval and RAG. This explainer stays on the shared mechanism and the page pattern that follows from it.
The short version
- Interpret The model reads the question and conversation context to work out the subject, intent and constraints.
- Decide It decides whether its trained knowledge is enough or whether it needs to search the web.
- Retrieve It writes one or more search queries and gets candidate pages back from an index.
- Extract It reads the returned text and keeps the passages that answer each part of the question.
- Compose It writes a response that combines those passages with its own knowledge.
- Cite It attaches links to the sources that support specific statements in the answer.
How an answer engine turns a question into an answer, step by step
ChatGPT, Claude, Perplexity, Gemini, Copilot and Google's AI features differ in their indexes and interfaces, but the developer documentation for the tools behind them describes the same loop. The details below come from that documentation.
Step 1: Interpret the question
The model reads the prompt together with any earlier turns in the conversation. B2B questions are usually loaded with constraints: a company size, an industry, an existing system, a compliance requirement. "Best contract management software for a 300-person fintech already on Salesforce" contains a category, a segment, a regulated industry and an integration. Each constraint becomes a condition the answer has to satisfy, and each one is a reason a generic page gets passed over.
Step 2: Decide whether to search
Engines do not search for every prompt. Anthropic's web search tool documentation says Claude searches when a request depends on information that is current or changing, including details about specific organizations, people or products that might have changed, and answers directly for stable knowledge such as established facts or concepts. OpenAI's web search guide likewise says the model can choose whether to search based on the prompt.
For B2B this has a practical consequence. Vendor, pricing and comparison questions usually trigger a search, so live retrieval decides who is cited. Definitional questions ("what is a data clean room") may be answered from training data with no search at all, which is a different problem covered in how GEO works.
Step 3: Write queries and fan out
When the model searches, it writes its own queries rather than pasting in the prompt, and it often writes several. Google calls this query fan-out: concurrent related queries generated to fetch additional results. Anthropic's documentation notes the search step can repeat several times in one request, and that comparative or multi-entity research can use ten or more searches. OpenAI describes agentic search, where a reasoning model analyzes results and decides whether to keep searching, and deep research, which can draw on hundreds of sources.
Fan-out is why a page can be cited for a question it does not rank for. The engine never ran your head term; it ran a sub-question your page answers well. Our guide to query fan-out optimization shows how to map those sub-questions.
Step 4: Retrieve candidate pages
Each query returns candidates from a search index, and each engine reaches the web through its own crawlers. Google's AI features use Google's index. Microsoft Copilot uses Bing. Perplexity says PerplexityBot exists to surface and link websites in its search results. OpenAI says OAI-SearchBot surfaces websites in ChatGPT's search features. If the relevant crawler is blocked, or your firewall challenges it, your page is not a candidate. This is the step where classic SEO and crawler access do most of the work.
Step 5: Read and extract passages
The engine now reads the candidate text and keeps what is relevant. Anthropic documents a version of this directly: with dynamic filtering, Claude writes and runs code that filters search results before they reach its context window, so only relevant content is kept. Google's ranking systems work at a similar grain; its passage ranking system identifies individual sections of a page to judge relevance.
The practical reading: the engine is not evaluating your page as a whole. It is looking for a stretch of text that answers one part of the question. Text that exists only in images, PDFs behind forms, tabs that load on click, or client-side JavaScript often never reaches this step.
Step 6: Compose the answer and attach citations
The model writes a response from the passages it kept plus its own knowledge, and links statements to sources. Anthropic's documentation says citations are always enabled for web search, and each citation carries the URL, the page title and up to 150 characters of cited text. OpenAI's responses attach url_citation annotations with the URL, title and the location of the cited text, and a separate sources field lists every URL the model consulted, while inline citations show only the most relevant ones.
That last distinction matters for measurement. Being read is not the same as being cited, and being cited is not the same as being named as a recommended vendor. How many slots exist also varies by engine: Orbit Media's 2026 study of 72 B2B prompts and 13,184 citations averaged 4.5 sources per answer in ChatGPT, 3.6 in Claude, 8.1 in Gemini and 19.2 in Perplexity.
Where the words in an answer come from
An answer blends two kinds of knowledge. The first is what the model learned in training, stored in its parameters. The second is text retrieved at answer time. The original retrieval-augmented generation paper (Lewis et al., 2020) framed the benefit of retrieval as giving models provenance for their outputs and a way to update their knowledge without retraining. Answer engines are that idea at web scale.
| Source of the answer | How your content gets in | How current it is | Can it link to you | What AEO can influence |
|---|---|---|---|---|
| Training data | Public pages collected by training crawlers before a cutoff | Months or more behind | No direct link | Consistent, widely repeated facts about your company |
| Live retrieval | Search crawlers index pages; the engine fetches results per question | As fresh as the index | Yes, as a citation | Crawler access, passage quality, specificity, freshness |
| User-triggered fetch | A user pastes your URL or asks about it; an agent fetches it | Live | Yes | Raw HTML that reads well without scripts |
| Structured knowledge | Knowledge graphs built from many sources | Varies | Sometimes | Entity consistency and sameAs links |
What makes a passage extractable and citable
No engine publishes a scoring formula for passages. What follows is inferred from the mechanism above and from Google's published guidance, and it is consistent with what we see when benchmarking B2B prompts.
- It is in the HTML. Most AI crawlers read raw HTML and do not run JavaScript. If View Source does not contain the sentence, assume the engine cannot quote it.
- It stands on its own. The passage names its subject. "Acme's AP automation connects to NetSuite through a certified SuiteApp" can be lifted out. "It connects natively" cannot.
- It answers first. The answer comes in the first one or two sentences under a heading that matches how the question is asked, and the caveats follow.
- It is specific. Numbers, names, time frames and conditions. Google asks for non-commodity content, and a generic sentence gives an engine no reason to prefer your version over twenty others.
- It agrees with other sources. When your page, your G2 profile and an analyst note all state the same integration list, the claim is easier to trust. When they conflict, engines hedge or pick the third party.
- It is attributable. A named author, an update date and a source for any figure give the engine something to hang a citation on.
- It is allowed to be used. Google's robots meta documentation says
nosnippetprevents content from being used as a direct input for AI Overviews and AI Mode, andmax-snippetlimits how much can be used.
One caution. Google says there is no requirement to break content into tiny pieces and no need to write in a special way for AI. The pattern below is good editorial structure that happens to suit machines. It is not a trick, and pages written only for extraction read badly to the humans who make the purchase.
The answer-block pattern
An answer block is a reusable unit for any question a buyer asks. Build it in this order:
- A heading phrased as the question, in the buyer's words: "How long does SOC 2 Type II take for a fintech startup?"
- A direct answer of roughly 40 to 60 words that names the subject, gives the answer and states the main condition.
- The "it depends" layer: the two or three factors that change the answer, each in a sentence.
- Evidence in a scannable form: a table, a short list of steps or a set of figures with their source.
- Proof: a named customer example, a screenshot or a link to documentation.
- Ownership: author, reviewer where relevant, and the date the answer was last checked.
<h2>How long does AP automation take to implement with NetSuite?</h2>
<p>Most mid-market teams go live with [Product] on NetSuite in four
to six weeks. The range depends on the number of subsidiaries, whether
purchase orders need three-way matching, and how many approval rules
move over from email.</p>
<table>...phase, duration, who is involved...</table>
<p>Example: [Customer], 3 subsidiaries, live in five weeks.</p>
<p>Reviewed by [Name], Implementation Lead. Updated September 2026.</p>The figures in the skeleton are placeholders. Use your own delivery data, and only publish a range your team actually hits.
How competing for an answer differs from competing for a ranking
| Dimension | Classic search result | Answer engine citation |
|---|---|---|
| Unit that competes | A page | A passage, usually a few sentences |
| Number of slots | About ten organic links per page of results | A handful of citations, varying by engine |
| Query that matters | The query the user typed | The queries the model wrote, often several |
| What wins | Relevance, quality, links, page experience | Retrievability first, then directness, specificity and corroboration |
| Success signal | Position and clicks | Mention, citation, description accuracy, then clicks |
| Main failure mode | Not ranking | Retrieved but not quoted, or quoted without being named |
| Measurement | Search Console, rank tracking | Prompt benchmarks, Bing AI Performance, Search Console, GA4 |
How to measure AEO
Measure in layers, from closest to the engine to closest to revenue.
- Visibility. Build a fixed set of 50 to 100 buyer prompts and run them across engines on a schedule, with repeat runs because answers vary between runs. Record mention rate, citation rate and share of voice. Our guide to benchmarking AI visibility has the formulas, and answer volatility explains why single runs mislead.
- Exposure. Bing Webmaster Tools AI Performance reports total citations, average cited pages, a sample of grounding queries and page-level citation activity across Copilot and Bing AI summaries, though Bing notes it does not show ranking or a page's role in an answer. Google counts AI Overview and AI Mode traffic in the Search Console Performance report and points site owners to its generative AI performance report.
- Traffic. Google's GA4 default channel group documentation now lists an AI Assistant channel for visits from sources such as ChatGPT, Gemini, Copilot and others, excluding Google's AI Overviews and AI Mode.
- Pipeline. Self-reported attribution and CRM fields connect all of it to revenue, covered in how citations become pipeline.
Keep third-party tools in perspective. Google's own guide says no third-party tool has access to its internal ranking or AI systems. Trackers sample outputs; they do not see inside the engine.
What you control: the AEO checklist
- Search-facing crawlers (Googlebot, Bingbot, OAI-SearchBot, Claude-SearchBot, PerplexityBot) can fetch every public commercial page.
- Core copy, tables and FAQs are present in the raw HTML.
- Each priority page opens with a direct answer to the question the page exists for.
- Every buyer question cluster has one answer block: pricing model, implementation time, integrations, security, alternatives.
- Passages name your company and product instead of "we" and "our solution" where the sentence needs to stand alone.
- Claims carry numbers, conditions and dates, with sources for any statistic.
- Your facts match your review profiles, partner listings and LinkedIn page.
- Author and update dates are visible on guides and comparison pages.
- No
nosnippetor aggressivemax-snippeton pages you want quoted. - A prompt benchmark runs monthly and results reach the team that edits the pages.
Illustrative example: an implementation-time question
Illustrative example. The vendor, prompts and results below are invented to show the mechanism. They are not a client result.
A mid-market accounts payable automation vendor tracks the prompt "How long does it take to implement AP automation with NetSuite?" across four engines. In the first benchmark, none cite the vendor. Two cite a competitor's help center article that states a week range, and one cites a review site listicle.
- Retrieval check. The vendor's implementation page is indexed, and logs show OAI-SearchBot and PerplexityBot fetching it, so access is not the problem.
- Extraction check. The page says "fast, guided onboarding" and puts the timeline in a downloadable PDF. There is no passage that answers the question.
- Fix. The team adds an answer block with a four-to-six-week range from its own delivery records, the three factors that stretch it, a phase table and a named customer example, and updates the matching G2 profile text so the facts agree.
- Re-measure. The prompt set is rerun monthly with three runs per engine. The team records whether the page is cited, whether the range is quoted accurately, and whether the vendor is named or only linked.
The lesson is about sequence. The fix was not more content or new markup. It was one specific, self-contained passage that existed nowhere before, published where crawlers could already reach it.
Glossary
- Answer block: a question heading, a direct answer, conditions, evidence and proof, built to be read by people and quoted by engines.
- Answer engine: a system that responds to a question with a written answer and sources, rather than a list of links.
- Citation: a link attached to part of an answer, pointing to the source that supports it.
- Citation rate: the share of tracked prompts in which your domain is cited.
- Grounding: tying a model's answer to retrieved sources so it reflects current, checkable information.
- Mention rate: the share of tracked prompts in which your brand is named, linked or not.
- Passage: a short section of a page that can be understood and quoted on its own.
- Query fan-out: the model generating several related searches from one question.
- RAG: retrieval-augmented generation, where a model retrieves documents and uses them to write its answer.
Next in the series
AEO is about the passage. How GEO works steps back to the brand: how models form a view of your company from training data and retrieval, and why some vendors get recommended by name. For a complete self-audit and benchmark template, get the B2B AEO and GEO playbook, or see our AEO, GEO and SEO service.
