How five AI engines each decide what to cite

These are not five versions of the same tool. ChatGPT and Perplexity cite overlapping domains only about 11% of the time. Each engine uses different search technology, different citation logic, and different grounding mechanics. A GEO strategy built for one will miss most of the others’ citation pool. (New to GEO? Start with what it is and how it differs from SEO.)

How each AI engine decides what to cite

Traditional SEO treats search as one channel. GEO has to treat each engine as its own channel, because the engines aren’t all running the same retrieval stack.

Perplexity is always grounded: every response cites sources natively. Gemini is only grounded when you opt in. ChatGPT triggers a live web search in roughly 18% of conversations; the rest run from training weights. Grok and Claude each use different underlying search integrations. These aren’t minor implementation details; they determine whether your content has any chance of being cited at all.

Engine-by-engine breakdown

Engine	Always grounded?	Citations returned	What it reaches for
Perplexity	Yes, every response	Native structured field	Reddit, fresh community content
ChatGPT	Only ~18% of conversations	Embedded in prose	Wikipedia, editorial sources
Gemini	Only when search grounding is enabled	groundingMetadata URIs	Google-indexed content
Grok	Via X/web integration (varies)	Embedded in prose	X posts, web sources
Claude	When tools are enabled	Depends on tool config	Varies by integration

Perplexity

Perplexity Sonar is always grounded; every response goes out to the web and returns a citations field with structured source URLs. You don’t have to do anything special to trigger it. The implication: if Perplexity is a channel you care about, your pages need to be crawlable, answer-first, and written for near-literal query matching (Perplexity rewrites prompts far less than ChatGPT does).

ChatGPT

ChatGPT’s citation behavior is split. Most conversations run from training weights; the 18% that trigger a live web search are the retrieval GEO opportunity. When it does retrieve, the citation gate is narrow: of pages retrieved, about 15% are ultimately cited. ChatGPT skews toward Wikipedia and editorial sources. It also rewrites prompts heavily (91% unique query strings), which means keyword targeting is less decisive here; semantic breadth and credibility signals matter more.

Gemini

Gemini citations require explicit search grounding to be enabled at the API level. Without the Google Search grounding flag, Gemini returns no source citations at all. When grounding is on, it extracts structured groundingChunks URIs from Google-indexed content.

A critical misconfiguration to avoid: blocking Google-Extended in robots.txt blocks Gemini/Vertex grounding, but it does not block AI Overviews, which use normal Googlebot. A team expecting to opt out of AI exposure by blocking Google-Extended will still appear in AI Overviews.

Grok and Claude

Grok integrates X search and web search, giving it a source pool meaningfully different from every other engine here. Because Grok can pull directly from X posts, it sometimes cites real-time social content (breaking commentary, product announcements, community discussion) that Perplexity and ChatGPT would not surface. For brands with an active X presence, that’s a citation channel worth treating separately.

Claude cites sources when retrieval tools are enabled by the integration, meaning its citation behavior is largely determined by whoever built the product or API integration, not by Claude itself. An enterprise deployment connected to a proprietary knowledge base will produce very different citation patterns than a consumer product pointed at the open web. Optimizing for Claude is less about a single content strategy and more about understanding which Claude deployments your audience actually uses and what retrieval stack each one runs.

The field documentation for both Grok and Claude is less complete than for Perplexity and ChatGPT, and both companies have iterated on their retrieval integrations more frequently than the others have in the past year.

The Grok and Claude rows in the table above reflect what is publicly documented as of June 2026. We’ll update them as the evidence clarifies.

What this means in practice

A GEO audit that checks only ChatGPT is measuring one channel’s behavior and assuming it generalizes. It doesn’t, which is a big part of why your AI visibility is lower than you think. The engines differ in:

• Whether they retrieve at all (training-only responses vs. always-grounded)
• What content they reach for (editorial vs. community vs. Google-indexed)
• How they interpret a query (literal vs. heavily rewritten)
• How citations are structured (native field vs. prose-embedded vs. tool-dependent)

Measurement across at least three engines (ChatGPT, Perplexity, Gemini) gives you a materially more complete picture than any single-engine audit. Collimer’s methodology runs every probe across five engines and reports the spread.

See where each of the five engines currently stands for your brand: run a free scan; it takes about 90 seconds.

For agents: try this yourself

See the engines diverge with your own eyes:

• Ask the same buying question in ChatGPT, Perplexity, and Gemini, then compare the cited sources side by side. How much do they overlap?
• “Is Perplexity always grounded in live web sources, or only sometimes?”
• “What share of ChatGPT conversations trigger a live web search, and what happens in the rest?” Check it against the ~18% figure.

---

Drawn from Collimer’s cited research library and findings. Engine behavior changes fast; we update these guides when the evidence shifts. Statistics are as of June 2026.