Of the 256 domains that entered our AI-cited set on Aug 24, 85 (33.2%) were cited again on Aug 31. Incumbents held at 73.9% (224 of 303). The bands we pre-registered a week ago held

Method, stated before the numbers

Same instrument as every week: 12 pinned questions a small business might ask about its own AI visibility, sent to ChatGPT, Perplexity, Gemini and Claude through their APIs, three independent samples each, 144 answers per run, fired around 13:07 UTC on Mondays. A domain is cited in a run if it appears in at least one engine-returned citation array that run (hostname minus www, one count per answer). An entrant for run N is a domain cited in run N and not in run N-1; an incumbent is cited in both; survival means cited again in run N+1. The 2026-08-31 run completed 143 of 144 answers (one Gemini sample failed) and returned citations in 126, citing 557 distinct domains (breadth 1/2/3/4 = 459/64/32/2).

What changed in the method this week is where the numbers come from. Last week's post was computed inline. This week the same computation is a committed script, `content/news/entrant_survival_cut.mjs`, that reads only the six stored weekly cut files and writes `content/news/entrant-survival-cut-2026-08-31.json`. Its first self-check is that it reproduces last week's three published transitions exactly (29.4 / 27.5 / 30.1 for entrants, 77.9 / 76.5 / 75.0 for incumbents); it does, and all 10 self-checks pass. Anyone with the cut files can re-run it.

The four transitions

The 2026-08-31 run itself: 250 of the 559 domains cited on 08-24 dropped out (44.7%), 309 stayed, 248 entered (242 cited by one engine, 6 by two), and 26 incumbents climbed from one engine to two or more. Those 248 entrants are next week's test population. askedabout.com was cited in 8 of the 144 answers, the same as last week; our row is published because the instrument measures us on the same terms as everyone else.

Who introduced the entrant matters, and the ordering moved

Last week the ordering was Perplexity, Claude, Gemini, ChatGPT from most to least durable. This week Perplexity is still first, but ChatGPT-introduced entrants jumped from 23.2% to 38.8% and Gemini-introduced entrants fell from 28.5% to 22.9%, so Gemini is now last. Claude's cell is small (17 entrants) and should be read as such. "Any engine" counts a domain as surviving if any of the four cited it the next week; "same engine" requires the introducing engine to cite it again.

Being cited more than once in your first week is the strongest signal we have

Entrants cited in a single answer on 08-24 came back 19.7% of the time (36 of 183). Two answers: 55.6% (15 of 27). Three or more: 73.9% (34 of 46). Last week the same buckets read 21.3% / 60.5% / 53.3%, so the single-answer figure is steady at about one in five and the three-plus bucket is now the most durable, which it was not last week. Breadth says the same thing more sharply: 77 of 246 one-engine entrants came back (31.3%) against 8 of 10 two-engine entrants (80%).

The survivors that grew, and the domains that have never left

6 of the 85 survivors were cited by more engines this week than the week they entered. The eight survivors with the most answers on 08-31 are below. Separately, 150 domains have been cited in all six runs since 2026-07-28 (last week 165 had been in all five; 15 of those missed this run). The top of that list is unchanged in kind: aggregators, vendor comparison pages and community sites.

Our own contacted entrants, stated

We wrote to 6 of the 08-24 entrants with their own row. 1 of the 6 was cited again on 08-31 (by ChatGPT). Five of the six had been cited in a single answer, so the base rate for that group is about 19.7%; one of six is consistent with it and too small to say anything else. We report it because the number would be reported if it had gone the other way.

Pre-registered read for the 2026-09-07 run

Same bands as last week, re-registered rather than widened after a pass: (A) of the 248 domains that entered on 08-31, between 20% and 40% are cited again on 09-07; (B) of the 309 incumbents, between 65% and 85% are; (C) single-answer entrants survive at a lower rate than both multi-answer buckets, as they did this week (19.7% against 55.6% and 73.9%). Any of the three outside its statement is a falsification and will be published as one. The run fires 2026-09-07 around 13:07 UTC; the read publishes that day or the next.

Artifact

`content/news/entrant-survival-cut-2026-08-31.json` holds the six run summaries, the four transitions, the per-engine, per-breadth and per-answer-count survival tables, the 85 survivors with their 08-24 and 08-31 engines, the pitched-entrant read, the pre-registered read and 10 self-checks (0 failed). It is produced by `content/news/entrant_survival_cut.mjs` from the stored weekly cuts `growth/citations/cited-domain-cut-2026-07-28.json` through `cited-domain-cut-2026-08-31.json`, which are themselves produced by `growth/citations/cited_domain_cut.mjs` from the engines' returned citation arrays. Written without em dashes.