A comparison prompt pulls 22 to 42 percent more citations on two engines. Two others barely move.
Citation count has already shown up as an engine trait, a branded-vs-organic effect, and a mention-outcome correlate in earlier posts, but never checked against the finer-grained intent field: recommendation, comparison, or best-of. A separate post found that comparison prompts also make two engines write meaningfully longer answers. We joined the same intent field to citation count to see whether it is the same story or a different one.
Comparison prompts pull more citations than recommendation prompts, and it holds in every category, on two engines
Across 270 live audit panels and 8,664 real non-error responses, we joined each response's promptId to its panel.json prompt record to read the intent field, then averaged citation count per engine per intent.
| Engine | Recommendation avg (n) | Comparison avg (n) | Difference |
|---|---|---|---|
| Gemini | 12.11 (1,399) | 16.84 (832) | +39.1% |
| Claude | 5.45 (1,608) | 6.90 (962) | +26.5% |
| ChatGPT | 6.42 (767) | 6.39 (480) | -0.4% |
| Perplexity | 8.78 (1,429) | 8.82 (835) | +0.5% |
Gemini's comparison responses average 39.1 percent more citations than its recommendation responses. Claude's are 26.5 percent higher. ChatGPT and Perplexity stay within half a percentage point of flat, no meaningful movement either direction.
Splitting by category (PM, HR, CS software) checks whether one vertical is driving the two engine-level numbers.
| Engine | PM | HR | CS |
|---|---|---|---|
| Gemini | +42.0% | +37.9% | +30.5% |
| Claude | +26.5% | +22.2% | +31.3% |
| ChatGPT | +10.1% | -3.8% | -5.3% |
| Perplexity | -1.1% | +5.7% | +3.9% |
All six Gemini/Claude category cells land positive, 22.2 to 42.0 percent, confirming the engine-level averages are not one category's artifact. ChatGPT and Perplexity stay inside a roughly plus-or-minus 10 percent band with no consistent direction, matching their flat blended averages rather than a category effect canceling itself out.
Best-of prompts are a smaller slice of the dataset (53 to 110 responses per engine, versus hundreds to over a thousand for recommendation and comparison) so we report them without a category split. Best-of citation counts run higher than recommendation on Claude (+22.0%) and ChatGPT (+24.6%), roughly flat on Perplexity (+1.7%), and slightly lower on Gemini (-6.2%). With this few responses per engine we are not confident enough in the best-of numbers to build a claim on them; comparison versus recommendation, backed by hundreds of responses per cell, is the reliable finding here.
This is a different engine pairing than the same question asked of response length
A separate post checked the same intent field against word count instead of citation count. There, Gemini and Perplexity were the two engines that swung hard on intent, comparison prompts ran 39.7 percent longer than best-of on Gemini and 62.1 percent longer on Perplexity, while Claude and ChatGPT stayed close to flat (1.3 and 4.4 percent spread).
Here, on citation count, the pairing is Claude and Gemini, not Gemini and Perplexity. Only Gemini shows the effect on both measures. Perplexity swings hard on length by intent but not at all on citations, consistent with a pattern already established twice before: Perplexity's citation count sits inside a fixed 5-to-10 window on nearly every response regardless of what else changes, so intent has nowhere to push it. Claude shows the opposite split, a real citation effect with almost no length effect, meaning its comparison responses are not longer, just more heavily sourced per word. ChatGPT stays flat on both.
- Gemini: intent moves both length and citation count. The two effects likely share a cause, a comparison prompt makes it do more retrieval and write more about what it retrieved.
- Claude: intent moves citation count but not length. The extra sourcing is not turning into extra words.
- Perplexity: intent moves length but not citation count, capped by its fixed citation window regardless of how much it writes.
- ChatGPT: flat on both. Intent does not change how ChatGPT approaches a prompt in either dimension measured so far.
Four engines, three ways of not moving together. Word count and citation count are correlated within some engines and independent within others, so neither one is a reliable proxy for the other across the board.
Find out which of the four engines your buyers actually use, and how differently each one treats a comparison question versus a recommendation question in your category.
Get an AI Visibility Audit, $490Methodology
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 re-read every real (non-error) response from results.jsonl fresh this session, joined each response's promptId to its panel.json prompt record to read the intent field (recommendation, comparison, or best-of, the only three intents the live panel generator produces), and used the raw length of the response's citations array as the citation count. Every promptId matched a recorded intent, no fallback needed. Total response counts by engine (Gemini 2,326, Perplexity 2,358, Claude 2,680, ChatGPT 1,300) and blended per-engine citation averages (Claude 6.02, Gemini 13.77, ChatGPT 6.48, Perplexity 8.80) match every prior citation-count post's baseline exactly. No development-rail or fixture data is included; all responses came from live engine calls. Data collected June-September 2026.