A comparison prompt makes two engines write 40 to 62 percent more. Two others barely notice.
An earlier post on this blog measured word count by engine and found category made almost no difference: the same engine ordering held inside project management, HR, and customer support software alike. That left prompt intent unchecked. We split the same real responses by whether the prompt asked for a recommendation, a head-to-head comparison, or a best-of list, and found a split answer: two engines treat intent as close to irrelevant, and two engines write dramatically more when the prompt is a comparison.
Comparison prompts inflate two engines and not the other two
We pulled 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), joined each response's promptId back to that panel's prompt list to get its intent, and counted words by splitting the response text on whitespace. 8,661 of 8,664 real responses matched a recorded intent. The prompt generator only produces three intents in this dataset: recommendation, comparison, and best-of. Review, alternative, and research prompts do not appear in the live panel set, so this post covers those three only.
| Engine | Best-of: avg words (n) | Comparison: avg words (n) | Recommendation: avg words (n) | Spread |
|---|---|---|---|---|
| Claude | 469.3 (110) | 463.2 (962) | 467.0 (1,608) | 1.3% |
| ChatGPT | 683.2 (53) | 674.2 (480) | 704.0 (767) | 4.4% |
| Gemini | 552.4 (95) | 771.5 (829) | 599.9 (1,399) | 39.7% |
| Perplexity | 191.5 (94) | 310.3 (835) | 221.0 (1,429) | 62.1% |
Claude's three intent averages sit within 1.3% of each other. ChatGPT's sit within 4.4%. Neither engine's response length moves in any consistent way with what kind of question was asked. Gemini and Perplexity are a different story, and both move the same direction: a comparison prompt draws a response 39.7% longer than a best-of prompt on Gemini (771.5 vs 552.4 words), and 62.1% longer on Perplexity (310.3 vs 191.5 words). Against the recommendation baseline the same two engines run 28.6% and 40.4% longer on comparison prompts respectively.
Holds inside every category, so it is intent, not category
We re-split Gemini's and Perplexity's word counts by category to check whether one crowded vertical was driving the average. It is not. Comparison responses run the longest of the three intents in all three categories on both engines, and best-of the shortest in all three:
| Engine / category | Best-of avg (n) | Comparison avg (n) | Recommendation avg (n) |
|---|---|---|---|
| Gemini, project management | 554.6 (60) | 745.9 (575) | 596.9 (968) |
| Gemini, customer support | 523.5 (20) | 843.0 (170) | 610.8 (263) |
| Gemini, HR software | 582.1 (15) | 802.1 (84) | 600.2 (168) |
| Perplexity, project management | 186.2 (65) | 302.6 (602) | 218.1 (1,016) |
| Perplexity, customer support | 233.1 (14) | 324.0 (124) | 246.2 (196) |
| Perplexity, HR software | 175.6 (15) | 336.8 (109) | 211.9 (217) |
The ordering never flips. That rules out the alternative explanation that comparison prompts happen to cluster in whichever category naturally produces longer answers. This is a prompt-intent effect on Gemini and Perplexity specifically, not a category effect riding along with intent.
Not just a restatement of the latency finding
An earlier post found branded, comparison-style prompts run slower than organic prompts on three of four engines (Claude, Gemini, Perplexity), and traced it to the engine having to reconcile two named entities instead of pulling one ranked list. It would be a reasonable guess that the length effect here is the same story told twice: slower because it is writing more. Gemini and Perplexity fit that guess. Claude does not. Claude's comparison prompts ran 10.5% slower in the latency post but its comparison-prompt word count in this dataset is flat, even fractionally shorter than its best-of average. Whatever costs Claude the extra time on a comparison prompt, it is not spending it on writing more words. The two effects travel together on two engines and separate on a third, which is itself informative: response length and response latency are correlated but not interchangeable measurements of "how hard did the engine work on this prompt."
Consistent with the known baseline
Blending all three intents back together per engine (Claude 465.7, Gemini 659.2, ChatGPT 692.1, Perplexity 251.4, all across 8,661 responses) lands within a fraction of a percent of the response-length-by-engine post's published figures (466, 658, 692, 251), confirming this is the same dataset read a third way, not a different sample.
Find out whether the AI engines your buyers use write a quick verdict or a long, hedged comparison when your brand comes up against a named competitor.
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 that panel's panel.json prompt list to recover its recorded intent (recommendation, comparison, or best-of; the live panel generator does not produce review, alternative, or research intents), and counted words by splitting the response text on whitespace. 8,661 of 8,664 real responses carried a matching intent record. No development-rail or fixture data is included; all responses came from live engine calls. Data collected June-August 2026.