Perplexity answers in 3.6 seconds. Claude takes 23.3.

Every post on this blog so far has asked what an AI engine says, which brands it names, which sources it cites. This one asks how long it takes to say anything at all. Every response in our audit data carries a latencyMs field the engines return, and it turns out to be one of the sharpest per-engine fingerprints in the whole dataset, sharper than citation count, sharper than mention rate.

A 6x gap between the fastest and slowest engine

We pulled the latency field 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, the same baseline count behind our citation-count posts. Split by engine, the wait time for a single grounded answer is not close.

The median Claude response takes 23.3 seconds to come back. The median Perplexity response takes 3.6. That is a 6.4x gap, and it is not a mean skewed by a few slow outliers, the middle 50% of each engine's responses (P25 to P75) tells the same story: Claude's middle half lands between 20.8 and 25.6 seconds every time, Perplexity's between 3.1 and 4.6.

The spread is a real distribution, not a handful of timeouts

Before trusting the averages we checked the shape of each engine's distribution directly, bucketed into 5-second windows. If a slow average were really a fast engine plus a handful of stuck retries, we would see a spike near zero and a long thin tail. That is not what shows up.

Each engine has its own tight home range, and almost every response in the dataset lands inside it. Claude's 43.6-second max and Perplexity's 41.6-second max are the rare true outliers, one response each out of thousands, not the shape of the data. The 6x gap between engines is a structural difference in how long each one spends retrieving and grounding an answer before it writes a word, not noise.

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

The same question we ask of every metric on this blog: is this an engine trait or a category trait? Averaged across all four engines, the three categories we track land within 1.3 seconds of each other, 12.7s for project management software, 13.4s for HR software, 14.0s for customer support software. That is a far tighter spread than the 18.6-second gap between Claude and Perplexity alone, and the engine ordering holds inside every category without exception.

What latency does not tell you

We checked whether slower responses correlate with anything that matters to a visibility strategy, and they do not. A slow Claude answer is not more thorough or more likely to name your brand than a fast Perplexity one; mention rate and citation count are governed by retrieval and category dynamics covered in earlier posts, not by how long the engine took to write. What latency does explain is operational: the four engines in a single audit panel do not finish at the same time, and Claude is consistently the long pole. If you have ever wondered why a full multi-engine audit takes longer to complete than any single engine call would suggest, this is why. It also means engines are not interchangeable as products even before you get to what they say. A tool built on Perplexity's API will feel close to instant to an end user. The same tool built on Claude's web search will not, regardless of answer quality.

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 latencyMs field from every successful (non-error) response, both branded and organic prompts, 8,664 responses total, this session, matching the same non-error baseline used in our citation-count posts. Responses that errored (rate limits, provider 5xx, or a zero-latency abort) were excluded, since a failed call does not measure how long a real answer takes. Distribution buckets were built in 5-second windows to confirm the averages were not skewed by a small number of stuck retries. No development-rail or fixture data is included; all responses came from live engine calls. Data collected June-July 2026.