ChatGPT writes the longest answer of any engine. It cites the least per word.

An earlier post on this blog measured citation count per response and found a more than 2x spread by engine, Gemini averaging 13.8 citations against Claude's 6.0. That post never asked how much text those citations sit inside. We pulled word count from the same 8,664 real responses and lined it up against citation count. The two numbers do not rank the engines the same way at all.

The length ranking is not the citation ranking

We pulled the response text from every real, non-error call 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, and counted words by splitting on whitespace.

ChatGPT writes the longest answer on average, Perplexity the shortest, less than half of Claude's and more than 60% shorter than ChatGPT's. But the citation-count-per-response post found a completely different order: Claude cites the fewest sources (6.0), then ChatGPT (6.5), then Perplexity (8.8), then Gemini by a wide margin (13.8). Perplexity writes the shortest answers on the engine list and still cites more sources than either of the two engines that write longer than it.

Citations per 100 words: Perplexity leads by a wide margin

Dividing each engine's average citation count by its average word count gives a citation density number, and it reorders the field again.

Inside one engine, does a longer answer mean more citations?

The numbers above compare engines to each other. A separate question is whether, within a single engine, a longer individual response tends to carry more citations than a shorter one from the same engine. We ran the correlation between word count and citation count response by response, inside each engine.

On Gemini, Claude, and ChatGPT, a longer individual response does moderately track more citations, the correlation is real though far from perfect. On Perplexity it is close to zero. A 100-word Perplexity answer and a 1,000-word one draw from roughly the same size source list, which matches the fixed 5-to-10 citation window an earlier post documented on this same dataset. Perplexity's citation count is set by something other than how much it has to say.

Category barely moves it. Engine is still the whole story.

Split by category, word count stays within about 45 words of its cross-category average for every engine, the same flat pattern citation count showed in the earlier post.

The same Perplexity-shortest, ChatGPT-longest ordering holds inside every category we track. Response length, like citation volume, is a trait of the engine you asked, not the market you are auditing.

Why this matters for a visibility strategy

A high citation count on its own can look like an engine that reads broadly. Length shows that reading isn't free: Gemini's citation lead comes with a much longer answer to read through, and ChatGPT's long answer is the one drawing on the fewest sources per word of anything it writes. Neither number tells you whether your brand gets named, that still comes down to whether your page is in the source list an engine actually retrieves. But if an audit shows your brand missing from ChatGPT specifically, you are missing from the engine giving its answer the most room and citing the least support for it, a thinner, more concentrated source list than the word count alone would suggest.

Methodology

Data drawn from the same 270 live, search-grounded audit panels used in the citation-count-per-response post (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 response text and citations field from every successful (non-error) response, both branded and organic prompts, 8,664 responses total, matching that earlier post's baseline exactly and confirming the same underlying dataset. Word count is the response text split on whitespace with no other normalization. Citation counts are the raw length of each response's citation array, with no deduplication or domain normalization applied. No development-rail or fixture data is included; all responses came from live engine calls. Data collected June-July 2026.