How AI assistants choose sources: ChatGPT, Perplexity & Claude

Ask ChatGPT, Perplexity, Claude, Microsoft Copilot, or Google's AI Mode a question and, for most queries, the assistant runs a web search before it answers. It rewrites your question into several search queries, pulls candidate pages from a search index, reads the passages, and quotes the ones that answer most directly. The visible credit — a linked footnote, a named source — is an AI citation, and engines grant it passage by passage, not page by page.

The architecture behind this has a name: retrieval-augmented generation, or RAG. The system retrieves relevant documents from an index first, then writes its answer from those documents, citing them as sources. RAG is why a page published on Tuesday can be quoted by Friday — the assistant fetches it live, and nobody has to retrain the model.

A brand reaches an AI answer through two doors. Training-data visibility is the model mentioning you from patterns memorized during training: slow to build, impossible to edit, and mostly a record of what the web said about you before the training cutoff. Retrieval visibility is an assistant citing a live page it fetched at question time: controllable within weeks through content structure, freshness, and presence in the index each engine searches. Everything actionable in this guide works through that second door.

The audience justifies the effort. ChatGPT reached 800 million weekly active users by OpenAI's DevDay in October 2025, up from 500 million that March, and Gartner predicted in 2024 that traditional search engine volume would fall 25% by 2026 as chatbots and virtual agents absorb the queries.

No two assistants read the same web. Each one cites from a different index, retrieves differently, and sends different crawlers to your site — so visibility in one engine says little about visibility in the next. The engine-by-engine plumbing, from primary sources:

ChatGPT searches through Bing and third-party providers, per OpenAI's help documentation — then re-ranks hard. RivalHound compared ChatGPT's citations with Bing's top results and found only 26.7% alignment, so a Bing ranking gets you into the candidate pool without deciding the outcome. OpenAI's layer favors reference material: Wikipedia alone holds 47.9% of ChatGPT's top-10 citation share, per Profound's analysis of 680 million citations collected between August 2024 and June 2025.

Perplexity owns its plumbing. Its index tracks more than 200 billion unique URLs, and it re-searches the live web on every query, which makes it the fastest major engine to reflect a new page. It also leans hardest on community discussion: Reddit accounts for 46.7% of Perplexity's top-10 citation share in the same Profound dataset. Fresh content and a presence in real conversations move Perplexity; a static brochure site does not.

Claude runs on Brave Search. Brave's appearance on Anthropic's subprocessor list in March 2025 confirmed the partnership, and RivalHound measured an 86.7% overlap between Claude's cited results and Brave's top non-sponsored results — lighter re-ranking than ChatGPT shows. Claude visibility is Brave visibility; check your Brave rankings directly.

Microsoft Copilot distills your prompt into keyword queries and sends them to the Bing service, a process Microsoft's documentation calls grounding. Classic Bing SEO — indexed in Bing Webmaster Tools, ranking for the keywords inside the prompt — carries most of the weight here.

Google AI Mode and AI Overviews cite from Google's own index. The official documentation states there are "no additional requirements" to appear: a page needs to be indexed and eligible to show in Search with a snippet. Both surfaces use query fan-out — one question split into parallel subtopic searches — which is why an AI Overview cites a wider set of pages than the classic results below it. Google's surfaces have enough rules of their own that we cover them separately in our AI Overviews guide; this article stays on the cross-engine mechanics.

Read the table as a routing map. The same page can be cited by all five engines, but it arrives through five different doors — and which doors stand open depends on your robots.txt, your index coverage, and your rankings in indexes most owners never check.

Barely. Ahrefs matched AI assistants' citations against Google's results for the same prompts and found only 12% of cited URLs rank in Google's top 10. The other 88% come from somewhere else — deeper in the results, different indexes, or pages Google never showed for that query at all.

Two mechanics explain the gap. Different indexes first: as the table above shows, four of the five major assistants search Bing, Brave, or a proprietary crawl, so a Google-only strategy leaves most doors untested. Passage-level selection second: engines quote the paragraph that answers, not the domain that ranks. A precise, self-contained answer on a modest site beats a top-ranked page that spends four hundred words warming up.

Rankings still matter as the admission ticket — an engine can only re-rank what its index returns — but they matter per index. Google rankings feed Google's AI surfaces. Bing rankings feed ChatGPT and Copilot. Brave rankings feed Claude. Checking all three takes an afternoon: Google Search Console, Bing Webmaster Tools, and a handful of Brave searches for the queries you care about.

Evidence and structure, per the study that named the discipline. In 2024, researchers from Princeton, Georgia Tech, and the Allen Institute published "GEO: Generative Engine Optimization" at KDD, testing multiple content tactics across 10,000 queries. The clear winners: adding statistics, quotations, and citations from credible sources, which boosted a site's visibility in generative-engine responses by up to 40%. Keyword stuffing, the oldest reflex in SEO, gained nothing. Engines reward pages that read like evidence — and a page that already cites its sources is the easiest page to cite in turn.

Format shapes the odds too. Wix's AI Search Lab classified 1,056,727 citations across ChatGPT, Google AI Mode, and Perplexity by page type: listicles earned 21.9% of all citations, articles 16.7%, and product pages 13.7% — three formats carrying more than half of everything cited. The shared trait is extractability. Numbered items, question-phrased headings, and self-contained sections give the model clean passages to lift; a page that buries its answer in narrative gives it nothing to quote.

On your own pages, three habits do most of the work. Open every section with the answer: a short, direct response under each question-phrased heading, elaboration after. Attribute every number inline — named source, year, link. And keep pages honestly current: Perplexity re-crawls the live web at query time, so freshness is a live variable there, and a visible updated date signals maintenance to every engine. None of this replaces being indexed and rankable in the first place, which remains the classic work of SEO — this discipline sits on top of that foundation, not beside it.

Two names circulate for the work. Generative engine optimization (GEO), the Princeton coinage, means structuring content so AI systems select it as a source. Answer engine optimization (AEO) means formatting content so a machine can lift a direct, self-contained answer from the page. They describe one goal from two angles — become quotable, then get quoted — and both reward the same habits above.

No special file earns citations. Google's documentation on AI features answers the question directly:

A page needs to be indexed and eligible to show in Search with a snippet — the same bar as ordinary Google results. Standard structured data keeps its normal job: Organization and LocalBusiness markup help machines confirm who and where you are. But no AI-specific markup, meta tag, or llms.txt file is required by any major engine, and none has confirmed that llms.txt affects citations. The page a human reads is the page the model quotes; spend the effort there.

The file that does gate the channel is robots.txt, and it cuts the other way. Blocking GPTBot, OAI-SearchBot, PerplexityBot, or ClaudeBot removes your pages from the indexes those assistants cite from — invisibility by choice. If you want the channel, allow the crawlers listed in the table above and confirm your firewall or CDN is not refusing them silently. Blocking is a legitimate call for paywalled or proprietary content you never want summarized; for a business courting customers, it closes off the channel this article describes. One Google nuance worth repeating: Google-Extended governs Gemini model training only, while AI Overviews eligibility rides on ordinary Googlebot indexing.

Two measured facts pull in opposite directions. Fact one: AI answers absorb clicks. Ahrefs analyzed 300,000 keywords and found the presence of an AI Overview correlates with a 34.5% lower click-through rate for the top-ranking page — a penalty its follow-up analysis later measured at 58%. Pew Research Center's study of 900 U.S. adults' browsing found users clicked a traditional result in 8% of searches that showed an AI summary, versus 15% without one — and clicked a link inside the summary itself in 1% of visits.

Fact two: the clicks that survive are worth more. Semrush's AI search traffic study measured the average visitor arriving from an AI search tool at 4.4 times the conversion value of the average organic search visitor. A plausible reason is that the assistant has done much of the comparing before the click, so the visitor tends to arrive closer to decided.

The battleground is predictable, too. Pew found only 8% of one- or two-word searches trigger an AI summary, against 53% of searches of ten words or more. Short navigational queries still behave the way they did a decade ago; conversational, question-shaped queries — the ones that sound like a person asking for advice — are where the assistant answers first and your citation is the whole game.

Citation share, not rankings, is the KPI — and a monthly loop with no specialized software covers most of it:

1. Run a prompt panel. Write 10–20 buying-intent prompts a customer would ask ("best [service] in [city]", "[your product] vs [competitor]") and run them monthly across ChatGPT, Perplexity, Claude, and Google AI Mode. Log every mention, citation, and competitor that appears. Consistency beats scale: same prompts, same engines, every month.
2. Segment AI referrals in GA4. Group traffic from chatgpt.com, perplexity.ai, and copilot.microsoft.com into one channel, and watch conversion rate alongside volume — the 4.4x pattern only shows up if you measure value, not sessions.
3. Read your server logs. GPTBot, OAI-SearchBot, PerplexityBot, and ClaudeBot in the logs mean the engines are ingesting your pages; their absence means a block somewhere upstream. Crawl activity is the earliest signal you get.

The loop tells you where to act. Mentioned by Perplexity but absent from ChatGPT points at reference-style authority; crawled but never cited points at structure. This is the same loop we run inside our AI visibility service, and our free SEO report includes an AI Visibility Score that checks crawler access, answer-first structure, and entity clarity for any URL. We built both after watching businesses — from Central Coast service companies to AI startups — discover they were invisible to engines their customers had already switched to.

The engines will keep changing names and shuffling indexes. The mechanics hold: an assistant needs an index you are in, a retrieval step you survive, and a passage worth quoting. Write for the passage.