How AI Search Engines Decide Which Sources to Cite
Amit Malvi
Founder, SearchAxe. Former Senior SEO Manager, Invideo AI
Published Updated 12 min read
Quick answer
How AI search engines decide which sources to cite comes down to three steps. The engine decides whether a question needs a web search, writes one or more queries, retrieves candidate pages from an index such as Google's, Bing's or its own, then quotes the passages that answer each part of the question most clearly. Relevance, authority, freshness, structure and third-party corroboration shape which pages survive each step.
On this page
- How AI search engines decide which sources to cite, in three stages
- When does an AI engine search the web instead of answering from memory?
- Where does each AI engine look for sources?
- Which signals decide which sources AI engines cite?
- How do the AI engines differ in the sources they cite?
- Why do cited sources change from week to week?
- What does this mean for a SaaS website?
Key takeaways
- Citations only appear when an engine searches, and engines search most for current, product-specific and comparison questions.
- Each engine retrieves from a different index, so a page missing from Bing or blocked to one crawler loses that engine entirely.
- Engines cite passages, not pages, which is why one clear paragraph can earn a citation for a page that does not rank for the original question.
- Measured citation patterns differ sharply by engine, from Wikipedia as ChatGPT's most cited domain to Reddit as Perplexity's.
- Cited sources rotate weekly, with ChatGPT search replacing 74% of its cited domains each week in SISTRIX's data, so one snapshot tells you little.
AI search engines decide which sources to cite in three steps: they decide whether to search, they retrieve candidate pages from an index, and they keep the passages that answer the question best. A page has to survive all three. If you run marketing or growth at a SaaS or AI startup and keep asking why the engines quote a competitor's page instead of yours, this explains the mechanics across ChatGPT, Perplexity, Claude, Gemini, Google AI Overviews, Google AI Mode and Copilot, using what each company documents and what the large citation studies have measured.
How AI search engines decide which sources to cite, in three stages
They search first and select second. The model turns the question into one or more search queries, an index returns candidate pages, and the model quotes the passages that answer each sub-question, citing the pages they came from. Google's developer documentation for Gemini's grounding with Google Search describes the same loop in five steps: the model decides whether a search can improve the answer, generates one or more queries, runs them, processes the results and returns a response grounded in them.
| Stage | What the engine does | What it means for your page |
|---|---|---|
| Decide and expand | Judges whether the question needs a search, then writes one or several queries for its parts | Your page can be retrieved for a sub-question it was written for, not only the original prompt |
| Retrieve | Sends the queries to an index and reads the pages that come back | A page that is blocked, unindexed or ranks poorly for those queries is never considered |
| Select and synthesise | Breaks pages into passages, keeps the clearest answers and assembles them into one response | A passage that states the answer plainly beats a better page that buries it |
Microsoft describes the selection step the same way. In its October 2025 guidance on AI search answers, Bing's Krishna Madhavan writes that assistants break content into smaller pieces, which are then ranked and assembled into answers, often drawing from multiple sources.
When does an AI engine search the web instead of answering from memory?
Engines search when the answer depends on current or changing information, and answer from training data when it does not. Anthropic's web search tool documentation lists the triggers: recent events, current prices or statistics, and information about specific organisations, people or products that might have changed. Established facts and conceptual explanations are answered directly, without a search and without citations.
That split matters for a startup. "What is retrieval-augmented generation" may be answered from memory, while "best RAG platform for a 20-person team" is a product question that usually triggers a search, and that is where your pages can be cited. The same documentation notes that simple factual queries typically use one to three searches, while comparative or multi-entity research can use ten or more, which is why comparison prompts pull in so many sources. Google's Gemini Apps help page confirms that not every response includes sources.
Find out which AI engines recommend you
A free consultation on where you appear today across AI engines and search.
Schedule a Free ConsultationWhere does each AI engine look for sources?
Each engine queries a different index, so being findable in one does not make you findable in all. Google's AI features documentation says a page must be indexed and eligible to show in Google Search with a snippet to appear as a supporting link, with no additional technical requirements. Copilot draws on Bing's index. ChatGPT, Perplexity and Claude each document their own search crawlers. Behind ChatGPT's search sits a mix of providers, including OpenAI's own index, according to Peec AI's September 2026 research.
| Engine | Retrieval source | What the documentation says |
|---|---|---|
| ChatGPT | OpenAI's search features, crawled by OAI-SearchBot | Sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers, though they can still appear as navigational links |
| Perplexity | Its own index, built by PerplexityBot | PerplexityBot surfaces and links websites in search results and is not used to train foundation models |
| Claude | Web search, with Claude-SearchBot indexing for search | Blocking Claude-SearchBot may reduce a site's visibility and accuracy in Claude's search results |
| Gemini | Google Search grounding | The model writes and runs search queries, then answers from the results |
| Google AI Overviews and Google AI Mode | Google's index, with query fan-out | Several related searches across subtopics feed one response |
| Copilot | Bing's index | Bing Webmaster Tools reports citations across Copilot and AI summaries in Bing |
The table draws on OpenAI's crawler overview, Perplexity's crawler page, Anthropic's crawler article and Bing's AI Performance announcement. A single robots.txt rule or CDN bot setting can remove you from one engine while every other engine still cites you, so access is the first thing to check.
Which signals decide which sources AI engines cite?
Six signals show up consistently in the documentation and the studies: relevance to the queries the engine runs, authority, freshness, structure, entity consistency and corroboration from third-party pages. None of the engines publishes weights, so treat the figures below as correlations measured from outside.
Relevance to the original question and its fan-outs
Google says Google AI Overviews and Google AI Mode may issue "multiple related searches across subtopics and data sources" to build a response. Surfer's September 2026 study of 173,902 URLs found a Spearman correlation of 0.77 between the number of fan-out queries a page ranked for and its Google AI Overviews citations. Ranking for the sub-questions a buyer has not typed yet counts. We cover that gap in why ranking first on Google does not guarantee AI citations.
Authority from mentions and links
Authority shows up in two forms. Ahrefs' May 2025 study of 75,000 brands found branded web mentions correlated 0.664 with Google AI Overviews brand visibility, against 0.218 for backlink count. SE Ranking's November 2025 study found referring domains were the strongest predictor of ChatGPT URL citations. Mentions drive whether you are named, links help whether your URL is retrieved, as we explain in do backlinks still matter for AI search visibility.
Freshness
AI assistants cite newer pages than Google's organic results. Ahrefs' July 2025 freshness study of nearly 17 million citations found AI assistants cited pages 25.7% fresher than organic results, averaging 1,064 days old, with ChatGPT citing the newest pages at 958 days. Anthropic's search results expose a page_age field for each result, so the model sees when a page was last updated. Freshness still loses to substance: the same Ahrefs study found the average cited page was close to three years old.
Structure that makes a passage easy to lift
Engines quote short spans. In Anthropic's documentation each web citation carries a cited_text field of up to 150 characters, roughly one sentence. Microsoft's guidance says assistants can often lift question and answer pairs word for word, and warns against hiding answers in tabs, PDFs or images. A sentence that answers the question on its own is the unit that gets cited. The page-level detail is in how to structure a web page so AI engines can quote it.
Entity consistency
Engines trust facts that agree with each other. Bing's AI Performance announcement tells publishers to align text, images and video so they "consistently represent the same entities, products, or concepts", and Google's May 2025 guidance says all content in structured data should also be visible on the page. If your homepage, docs, G2 profile and LinkedIn page describe your product three different ways, an engine has three conflicting answers to choose from.
Corroboration from third-party pages
Your own site competes with pages about you. Profound's June 2025 analysis of 680 million citations found Wikipedia was 7.8% of all ChatGPT citations and Reddit was 6.6% of Perplexity citations and 2.2% of Google AI Overviews citations. Moz's February 2026 study of nearly 40,000 queries found the top four domains took 10% of Google AI Mode citations, with YouTube the second most cited source. A claim repeated on reviews, roundups and community threads gives the engine more than one page to retrieve it from.
How do the AI engines differ in the sources they cite?
They differ enough that a single strategy tuned to one engine will miss the others. Even Google's two AI surfaces disagree: Ahrefs' December 2025 comparison of 730,000 response pairs found only 13.7% of cited URLs matched between Google AI Overviews and Google AI Mode, though the answers were 86% semantically similar.
| Engine | Measured pattern | Source |
|---|---|---|
| ChatGPT | Wikipedia is its most cited domain at 7.8% of citations; it cites the freshest pages of the assistants tested, averaging 958 days old | Profound, June 2025; Ahrefs, July 2025 |
| Perplexity | Reddit is its most cited domain at 6.6% of citations | Profound, June 2025 |
| Claude | We have not seen a large published study of Claude's citation patterns | Anthropic documentation only |
| Gemini | Answers are grounded in Google Search results, and not every response includes sources | Google documentation |
| Google AI Overviews | The most stable sources week to week, with about 11 domains cited, 8 staying put, and no source changing over 17 weeks for 53% of prompts | SISTRIX, May 2026 |
| Google AI Mode | About 56% of sources change each week, with YouTube the second most cited source | SISTRIX, May 2026; Moz, February 2026 |
| Copilot | Draws on Bing's index, with citations reported in Bing Webmaster Tools | Microsoft Bing, February 2026 |
How often each engine's citations overlap with Google and Bing rankings is covered in why ranking first on Google does not guarantee AI citations. Because Copilot draws on Bing's index, submit your site to Bing Webmaster Tools even if Bing sends you little organic traffic.
Why do cited sources change from week to week?
How stable an answer's sources are depends on the engine. SISTRIX's May 2026 citation drift study tracked 82,619 prompts for 17 weeks. ChatGPT search replaced 74% of its cited domains each week, and its median prompt had no domain that stayed all 17 weeks. Google AI Overviews cited about 11 domains on average with 8 staying put. Google AI Mode replaced 56% each week, yet 86.5% of its prompts had a stable core of one to five domains, while 89% of the other cited domains rotated each week.
Two things follow. A single check of one prompt tells you little, so measure the same prompt set repeatedly. And the goal is to become part of the core, which in the SISTRIX data meant a brand's own domain for its brand queries (present all 17 weeks for 43% of brand prompts), documentation and institutional sources in ChatGPT search and evergreen product pages in Google AI Mode, rather than news articles, which had a long-term retention rate of 1.4%.
What does this mean for a SaaS website?
It means working on all three stages, in order, for the prompts that decide your category. Check retrieval and selection before spending on authority, because a page that is blocked or buries its answer gains nothing from more links. This is the sequence an AI search optimization programme should follow.
Getting a SaaS site through retrieval, selection and corroboration
- Confirm every engine can retrieve youAllow Googlebot, Bingbot, OAI-SearchBot and PerplexityBot in robots.txt and at the CDN, and check Claude-SearchBot too. Verify indexing in Google Search Console and Bing Webmaster Tools.
- Cover the sub-questions engines search forFor a category prompt, engines look up pricing, integrations, alternatives and use cases. Give each one a page or a clearly headed section.
- Answer in the first sentencePut the plain answer, with a number or named example, directly under each heading so a 150-character span can stand alone.
- Make your facts match everywhereUse the same product description, category name and pricing model on your site, docs, review profiles and LinkedIn.
- Earn mentions on the pages engines already citeCollect the third-party URLs cited for your prompts and work on being included in them.
- Refresh facts that change and measure monthlyUpdate pricing, feature and comparison pages when they change, and track the same prompts across engines against a fixed baseline.
For the publishing side of this, with a pre-publish checklist, see how to get your content cited by AI search engines.
Frequently asked questions
How does an AI engine choose which passage to quote from a page?
It splits retrieved pages into smaller pieces, ranks them for authority and relevance, and assembles the best into one answer, often drawing on several sources, according to Microsoft's October 2025 guidance. In Anthropic's documentation, each Claude web citation carries a cited_text span of up to 150 characters, roughly one sentence. A paragraph that states the answer in its first sentence gives the engine a span it can lift whole.
Why do ChatGPT and Perplexity cite different sources for the same question?
They search different indexes and favour different source types. Peec AI's September 2026 research found ChatGPT combines OpenAI's own index with several outside providers, while Perplexity runs its own crawler and index. Profound's June 2025 analysis found Wikipedia was ChatGPT's most cited domain at 7.8% of citations, while Reddit led Perplexity at 6.6%. Treat each engine as a separate channel.
Do AI engines cite sources when they answer from training data?
Usually not. Citations come from the web search step. Anthropic's documentation says Claude answers directly, without searching, for stable knowledge such as established facts and concepts, and searches for current or product-specific information. Google's Gemini help pages say not every response includes sources. Product and comparison questions are the ones most likely to trigger a search and cite pages.
How often do the sources in AI answers change?
Often. SISTRIX tracked 82,619 prompts for 17 weeks and found ChatGPT search replaced 74% of its cited domains each week. Google AI Overviews were the most stable, while Google AI Mode replaced 56%. ChatGPT search's median prompt had no domain that stayed all 17 weeks, while in Google AI Mode most answers keep a core of one to five domains and rotate the rest, so track the same prompts repeatedly instead of trusting one snapshot.
Does blocking one AI crawler remove you from every engine?
No. Each engine uses its own crawler, so a robots.txt rule only affects the engine it names. OpenAI says sites that opt out of OAI-SearchBot will not be shown in ChatGPT search answers, though they can still appear as navigational links. Anthropic says disabling Claude-SearchBot may reduce a site's visibility in Claude's search results. Google AI Overviews and Google AI Mode follow your robots.txt rules for Googlebot.
