How AI Search Engines Build an Answer: Retrieval, Ranking, Synthesis

AI search engines usually build an answer in three stages: they retrieve candidate pages from a search index, rank and select the most relevant passages, then use a language model to write an answer, often with citations. Before any of that, the system decides whether to search at all. Marketers can influence each stage: be crawlable and indexed, publish clear passages that match the specific sub-questions, and be a source worth quoting.

On this page

• What happens before retrieval?
• Stage 1: How does retrieval work?
• Stage 2: How are sources ranked and selected?
• Stage 3: How is the answer synthesised?
• From MentionX data: what the citations look like
• Where does training data fit in?
• What does this mean for each stage? A summary
• Worked example: one question through all three stages
• A quick example
• How engines differ in showing their sources
• What are the most common misunderstandings?
• Diagnosing which stage failed
• Checklist
• Related
• Sources

Updated 31 October 2026. Anything marked "Our view" is MentionX's opinion, not documented engine behaviour. The worked example uses a fictional company.

What happens before retrieval?

When you type a question into ChatGPT search, Perplexity, Gemini or Google AI Mode, the system first decides whether it needs to search at all. A question like "What is photosynthesis?" can often be answered from the model's training. A question like "best CRM for a 10-person agency in 2026" needs current information, so the system searches.

Many systems also rewrite the question. Google says AI Overviews and AI Mode may use a "query fan-out" technique, issuing multiple related searches across subtopics and data sources to develop a response (Google). One question about CRMs might become searches about pricing, integrations, reviews and agency use cases.

Stage 1: How does retrieval work?

Retrieval finds candidate documents. Most AI search products rely on a web index:

This general pattern is called retrieval-augmented generation, or RAG: a model answers using documents retrieved at query time (Lewis et al.).

What you can influence: whether your pages can be crawled and indexed. If a crawler is blocked in robots.txt, or a page is noindex, it can't be retrieved.

Stage 2: How are sources ranked and selected?

Retrieval may return dozens or hundreds of documents. The system then narrows them down to the passages most useful for the answer. The exact methods aren't published, but the public guidance points to familiar signals:

• Relevance of the passage to the specific question and sub-questions
• Quality and reliability, which Google describes in its helpful-content guidance (Google)
• Freshness for time-sensitive questions
• Diversity of sources, so an answer isn't built from one site

The unit that matters is often a passage, not a whole page. A 3,000-word article might contribute one paragraph.

What you can influence: clear, self-contained passages that answer specific questions, question-style headings, accurate facts and evidence of expertise.

Stage 3: How is the answer synthesised?

The language model reads the selected passages and writes an answer. Along the way it:

1. Combines facts from several sources
2. Resolves or notes conflicts between them
3. Writes in its own words, often summarising rather than quoting
4. Attaches citations to some statements

Because generation involves sampling, the same question can produce different wording and different sources on different runs. That's why one test isn't enough to judge visibility.

What you can influence: whether your facts are easy to restate correctly. Specific numbers, definitions, and named comparisons survive summarising better than vague claims.

From MentionX data: what the citations look like

Method. We ran read-only aggregate queries over completed MentionX runs from 22 August to 29 October 2026 (UTC): 2,154 runs across 18 projects, mostly B2B software. Engines were called through their APIs with web search or grounding on, which can differ from the consumer apps. Citations are the source URLs returned with each answer. Brands are anonymised. Perplexity, Copilot and DeepSeek rows come from a single day (26 October).

Volume and spread. The runs returned 9,750 citations from 2,152 different domains. That's about 4.5 citations per run overall, or about 7.4 per run among runs that cited anything. Even the most-cited domains each made up only about 1% of all citations.

Many answers cite nothing. 39.2% of runs had no citations. The share varied by engine:

How to read it. The "decide to search" step is real and varies by engine: with search enabled, some engines still answered many prompts without returning sources. When they did cite, they drew on a long tail of sites rather than a few dominant ones. For you, that means two things. Training-based knowledge still matters, because a large share of answers show no retrieved sources. And there's no single "AI ranking" to win: the cited sources for your category are spread across many domains, so you need to map them per prompt.

Caveats. API answers, 18 projects, uneven engine coverage. Citation counts depend on how each API returns sources.

Where does training data fit in?

Even when a system searches, the model's training shapes how it interprets questions and which brands it already associates with a category. Training data comes from large crawls of the public web and other sources, collected well before the answer is generated. Separate crawlers often handle training versus search; OpenAI, for example, documents GPTBot for training and OAI-SearchBot for search (OpenAI). See Training data vs live retrieval.

What does this mean for each stage? A summary

Worked example: one question through all three stages

Illustrative example. Ledgerline is a fictional AP automation company, and everything below is invented.

A finance manager asks an AI assistant: "What's the best way for a 300-person company to automate invoice approvals?"

Decide to search. The question asks for current options, so the assistant searches.

Fan-out. It runs several queries, for example "invoice approval automation mid-sized company", "AP automation software comparison 2026" and "invoice approval workflow best practices".

Retrieval. The results include two software roundups, a review site's category page, a forum thread, a competitor's guide to approval workflows and Ledgerline's homepage. Ledgerline's detailed guide on approval workflows isn't retrieved. It isn't in the index, because a noindex tag was left on the blog's guide section after a redesign.

Ranking. From the homepage, nothing matches "approval workflows" closely. The competitor's guide has a section headed "How to set up invoice approval rules", with a numbered list. That passage is selected.

Synthesis. The answer explains approval workflows using the competitor's steps, cites the competitor's guide and both roundups, and recommends two tools. Ledgerline is mentioned once, from a roundup, without a reason.

What Ledgerline fixes. Remove the stray noindex and confirm indexing. Give the guide a section that answers the question in its first two sentences, with numbered steps. Make sure both roundups describe Ledgerline's approval features accurately. None of these steps involves "writing for AI". Each fixes a specific stage that failed.

A quick example

Illustrative. A user asks Perplexity, "Which project management tools work best for construction firms?" The system runs several searches, retrieves review sites, vendor pages and forum threads, and selects passages that mention construction-specific features. A vendor whose page has a clear section titled "Project management for construction firms", with a feature table, is far more likely to be quoted than one with only a generic homepage claim.

How engines differ in showing their sources

The stages are similar, but what you can see differs:

• ChatGPT search shows inline citations and a Sources list. OpenAI says it may rewrite your question into targeted queries and send follow-ups (OpenAI Help).
• Google AI Overviews and AI Mode show supporting links. Google says both may use query fan-out (Google).
• Perplexity shows numbered sources with almost every answer, which made it the engine with the fewest uncited runs in our data.
• Copilot has a Sources view and can show the web query it ran (Microsoft).

Our view: use the engine that shows the most as your diagnostic tool, even if your buyers use another. Seeing the queries and sources for one engine tells you a lot about what the others are probably retrieving, because they search the same web.

What are the most common misunderstandings?

• "AI search uses a different internet." Mostly it uses the same web, through search indexes.
• "You need special files to be included." Google says no new machine-readable files or special schema are required for AI features (Google).
• "The top-ranking page always wins." Ranking helps retrieval, but passages from several pages are combined.
• "One test shows the truth." Answers vary; measure over repeated runs.

Diagnosing which stage failed

When you're missing from an answer, work backwards through the stages instead of guessing:

1. Was the page retrievable? Check robots.txt for the engine's search crawler, check your logs for 200 responses, and confirm the page is indexed in Google and Bing. If not, stop here and fix access.
2. Was it retrieved? Look at the cited sources across several runs. If your page never appears but competitors' similar pages do, compare them with the sub-queries the engine is likely running.
3. Was the right passage there? If your page is cited but your brand isn't named or recommended, the cited passage probably doesn't say who you're for. Rewrite it.
4. Did synthesis distort it? If the answer cites your page but gets a fact wrong, check whether the fact is stated clearly and consistently, with a date, and whether an older third-party page says something different.

Our view: most teams start at step 3 and rewrite content, when the real problem is at step 1. Checking access first takes an hour and rules out the most common cause.

Checklist

• Search crawlers allowed in robots.txt
• Priority pages indexed
• Each priority question answered in a clear passage
• Question-style headings and tables
• Facts specific and consistent across your site
• Visibility measured over repeated runs

Related

• What is generative engine optimization?
• Why AI answers differ every time

Sources

• Google Search Central, AI features and your website
• Google Search Central, Creating helpful, reliable, people-first content
• OpenAI, Overview of OpenAI crawlers
• Perplexity, Perplexity crawlers
• Microsoft Bing, Webmaster guidelines
• Lewis et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks