What we looked at
Every market-week from 2026-W33 to 2026-W40 (8 complete weeks) with at least 10 web-search answers: 12 market-weeks across 6 markets, 238 answers from Perplexity, ChatGPT and Claude. 28 market-weeks had fewer answers and are left out, as they are from every trend we report. No market or business is left out for how its numbers look. Refusals and empty answers aren’t counted. Computed October 11, 2026; it updates every Monday.
Why one check misleads
- The business AI named most in a market-week was named in a median 72% of that week’s answers (n = 12 market-weeks). So even the leader is missing from 28% of single answers.
- The first business an answer named was the week’s most-named business in 22% of answers (n = 238). “Ranked #1 when I asked” is about 1 in 4 odds.
- Two answers to the same question, from the same engine in the same week, named the same top 3 businesses in 6% of pairs, and in the same order in 4% (n = 242 pairs).
A single check is one draw from a wide spread. That’s why every number we report pools many answers and shows how many.
How wide the ranges are
For a typical business near the top of its market, named in about 43% of answers (the median for businesses ranked in the top 5 of their market-week, n = 60), the 95% range at different numbers of answers:
| Answers | 95% range | |
|---|---|---|
| 10 | a quick check | 17%–69% |
| 24 | a few days | 25%–61% |
| 72 | a full week | 32%–55% |
| 144 | a full week, deep sampling | 35%–51% |
Doubling the answers narrows the range by about 30%, not by half. That’s why deep sampling helps, and why a trend pools several weeks.
How this was worked out
- Every answer is a web-search answer through each AI company’s official API, stored with its date and model. A business counts once per answer.
- Ranges are 95% Wilson score intervals. The noise test is a two-proportion test at 95%, the same as on every trend. Both treat answers as independent; repeats of one question are somewhat alike, so true uncertainty is a little wider.
- Spellings of one business are grouped by a normalized name. Weeks when the AI models or our questions changed are left out of the week-to-week figures.
- No business or market is named here. The answers behind it are the same ones every report uses.
More on the method: how Algotally measures.