The engines disagree with each other, sharply
Why a single screenshot is worse than no measurement.
Published 2026-08-05 · last reviewed 2026-09-14 · 1 sources
Corrected
The figures in the first paragraph cited a write-up we withdrew on 13 September because its sample was too thin once Claude's answers were re-scoped. Those figures stand on their own ranges, so the source now points to our published measurement data. The website counts in the second paragraph are corrected to completed searches: across 624 answers from 15 to 29 August, a median of 15 websites for Perplexity, 6.5 for Claude and 3 for ChatGPT. They previously read 790 answers and medians of 16, 7 and 3.
Corrected
Claude's figures are re-scoped to its completed searches. Under the collection setting then in force Claude was allowed one search per answer, and 166 of its 261 answers report the search stopping early while still carrying a source list. Its count here was 9 of 57; on completed answers it is 3 of 21.
Corrected
This finding called the comparison site “the most-cited independent site”, but a public forum was cited in one more answer (48 to 47); it now says “the most-cited comparison site”. It also explained the gap as assistants that search the web against assistants answering from memory, when all three searched the web for these answers. That explanation is replaced with what we measured.
Measured on the same questions, on the same days, with web search on, the engines do not behave like one market. In our own four-day measurement of an Australian health category, Perplexity quoted the most-cited comparison site in 33 of 60 answers (55%, 90% band 44-65%). Claude quoted it in 3 of 21 completed answers (14%, band 6-31%). ChatGPT, in 5 of 60 (8%, band 4-16%). Perplexity's band does not touch either of the others: that separation is real. Claude's and ChatGPT's bands overlap, so by our own rule those two are too close to call on this sample, and we do not rank them.
A screenshot from either engine would have been badly misleading, in opposite directions. All three were searching the web for these answers, so this is not search against memory. They read different amounts of it: across 624 answers we collected from 15 to 29 August, a Perplexity answer drew on a median of 15 websites, Claude's on 6.5 (completed searches only) and ChatGPT's on 3.
So a visibility number without its engines and its sample size attached is not a measurement. It is an anecdote with a percentage sign.
SOURCES
On the same frozen questions over four days, Perplexity cited the most-quoted comparison site in 33 of 60 answers, Claude in 3 of 21 completed answers and ChatGPT in 5 of 60.
GEOMG, own daily measurement (Australian weight-loss telehealth, 22-25 Aug 2026) · 2026-08-25 · 177 graded answers
Income Lab (Pepform Pty Ltd). "The engines disagree with each other, sharply". https://incomelab.me/findings/engines-disagree. Accessed 2026-09-22. Licensed CC BY 4.0.
READ NEXT