Perplexity cites 5 to 10 sources every time. Gemini ranges from zero to fifty.

Earlier posts on this blog looked at which domains get cited across many responses, and found a small, repeating set of gatekeeper pages. None of them asked a simpler question: how many source URLs does a single AI response actually carry? That number turns out to be one of the most consistent per-engine fingerprints in the whole dataset, and it has almost nothing to do with category or query.

The average hides a more than 2x spread by engine

We pulled the citation count from every real, non-error response across our 270 live audit panels (project management, customer support, and HR software; ChatGPT with web search, Gemini with grounding, Perplexity Sonar, and Claude with web search), 8,664 responses total, both branded and organic prompts. Overall, a response carries an average of 8.9 citations and 96.6% carry at least one. Split by engine, that average is not one number, it is four very different ones.

Gemini's average citation count is more than double Claude's. That is not a rounding effect from a handful of outlier responses either, the median tells the same story: half of Gemini's responses carry 12 or more citations, half of Claude's carry 6 or fewer. Whatever grounding step each engine runs before it writes an answer, it retrieves a structurally different volume of source material.

Perplexity's citation count barely moves. Gemini's swings wildly.

The range column above is the more interesting number. Across 2,358 real Perplexity responses, not one fell outside a 5-to-10 citation window. Perplexity never cited zero sources and never cited more than 10, in any of the 270 panels, any category, any prompt type. Gemini did the opposite: its 2,326 responses ranged from a genuine zero all the way to 50 citations in a single answer, a spread 10 times wider than Perplexity's despite drawing from the same prompt set.

Category barely moves the number. Engine is the whole story.

If citation volume were mostly about how much is written about a category online, you would expect a real spread across project management, customer support, and HR software too. It is close to flat.

The full spread across three categories, 8.0 to 9.2, is smaller than the gap between Claude and Perplexity alone. Citation volume is a property of the engine you asked, not the market you are in.

Why this matters for a visibility strategy

Citation count is not a proxy for whether your brand gets named, a high-citation-count Gemini answer is not more likely to mention you than a tight 6-source Claude answer built from the same prompt. What it does tell you is how much competitive room exists in a single answer on each engine. A Claude response built from 6 sources has less room for your brand's page to be one of the ones read before the answer is written. A Gemini response built from 20 or 30 sources has more slots, but also more competition for each one. If an audit shows your brand missing from Claude specifically, the fix has less margin for error there than it does on an engine pulling from a wider source pool.

Methodology

Data drawn from 270 live, search-grounded audit panels (project management, customer support, and HR software brands), each run across four AI engines: ChatGPT with web search, Gemini with grounding, Perplexity Sonar, and Claude with web search. We pulled the citedSources field from every successful (non-error) response, both branded and organic prompts, 8,664 responses total, this session. Citation counts are the raw length of each response's citation array as returned by the engine, with no deduplication or domain normalization applied (unlike CitedWell's production top-cited-source ranking, which does normalize and exclude Gemini's opaque vertexaisearch.cloud.google.com redirect links when identifying which domains matter, covered in an earlier post). No development-rail or fixture data is included; all responses came from live engine calls. Data collected June-July 2026.