Of the 218 domains that entered our AI-cited set on Sep 7, 75 (34.4%) were cited again on Sep 14. Incumbents held at 72.2% (241 of 334). Sixth weekly transition inside the pre-registered bands, and 8 of 167 swept entrant domains publish a named person in the US, UK or Canada

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 2026-09-14 run completed 144 of 144 answers and returned citations in 127, citing 615 distinct domains (breadth: 516 cited by one engine, 64 by two, 27 by three, 8 by all four).

The computation is the committed script `content/news/entrant_survival_cut.mjs`, unchanged since 2026-08-31. It reads only the eight stored weekly cut files and writes `content/news/entrant-survival-cut-2026-09-14.json`. Its self-checks include reproducing the three transitions published on 08-24 exactly (29.4 / 27.5 / 30.1 for entrants, 77.9 / 76.5 / 75.0 for incumbents) and matching this run's churn, survival and entrant counts to the stored cut; 10 of 10 pass. Anyone with the cut files gets the same tables.

The six transitions

The 2026-09-14 run itself: 236 of the 552 domains cited on 09-07 dropped out (42.8%, the highest churn of the eight runs), 316 stayed, 299 entered (292 cited by one engine, 5 by two, 2 by three), and 28 incumbents climbed from one engine to two or more. 299 entrants is the largest entering cohort of the series, and 615 the largest cited set. Those 299 entrants are next week's test population. askedabout.com was cited in 6 of the 144 answers, 5 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 reversed

Last week the ordering was ChatGPT, Perplexity, Claude, Gemini from most to least durable. This week it is Claude, Perplexity, Gemini, ChatGPT. ChatGPT-introduced entrants fell from 49.4% to 29.6% and Claude-introduced entrants rose from 37.1% to 52.0%; Gemini, last for two weeks, moved to third at 37.8% and again introduced the most entrants (98 of 218). The Claude cell is 25 entrants and the Perplexity cell 28, so a swing of a few domains moves those rates by ten points. "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 still separates the survivors

Entrants cited in a single answer on 09-07 came back 23.2% of the time (33 of 142). Two answers: 52.9% (18 of 34). Three or more: 57.1% (24 of 42). Last week the same buckets read 24.6% / 61.5% / 55.3%. The single-answer figure is within two points of last week and the split that holds every week is single-answer against the rest. Breadth reads the same way: 64 of 204 one-engine entrants came back (31.4%) against 11 of 14 two-engine entrants (78.6%); last week 33.9% and 83.3%.

Survivors that grew, and the domains that have never left

11 of the 75 survivors were cited by more engines this week than the week they entered. 123 domains have been cited in all eight runs since 2026-07-28; last week 139 had been in all seven, so 16 of those missed this run. The top of that list is the same kind as every week: community sites, review aggregators and vendor pages. This post publishes counts only; the domain lists are in the artifact.

Our own contacted entrants, stated

We wrote to 3 of the 09-07 entrants with their own row. 0 of the 3 were cited again on 09-14. All 3 had been cited in one to three answers, so the base rates for that group run from 23.2% to 57.1%; 0 of 3 is inside what those rates produce and too small to say anything else. We report it because the number would be reported if it had gone the other way.

A second read on the same cut: how many new entrants can be reached

Every week since 08-24 we have taken the entrants plus the incumbents that climbed to two or more engines, dropped every domain already fetched or contacted in a prior week, and fetched the rest at eight conventional contact paths (the homepage, /contact, /contact-us, /about, /about-us, /team, /privacy, /privacy-policy). An address counts only on the outlet's own registrable domain. A domain passes if a full name is published beside or for that address on the domain and the domain publishes a US, UK or Canadian location. Role inboxes (hello@, info@, press@ and the like) never pass. The rule stated on 08-24 was that two consecutive weekly reads under 3% close this population as a source of people to write to.

This week: 327 candidate domains (299 entrants, 28 climbers), 156 already fetched or contacted in a prior week, 4 excluded as platform or enterprise hosts, 167 fetched. Of those, 32 published at least one person-shaped address, 67 only role inboxes, 66 no address at all, 2 did not answer. The 32 person-shaped domains were checked by hand: 8 pass. The 24 that did not: 10 carried a role inbox that the extractor had mis-read as a person, 7 publish a location outside the US, UK and Canada, 4 publish a first name with no surname, 2 sell an AI-visibility product themselves and are excluded by our own rule, 1 had the address only inside an image filename. The read is 8 of 167, 4.79%. Under the 08-24 rule, 2 of the four reads were under 3% (08-24 and 09-07), neither followed by a second consecutive read under 3%, so the population stays open. Roughly one in twenty newly cited domains is a person you can write to by name at a US, UK or Canadian address; the rest are role inboxes, anonymous pages, or outside those countries.

The pre-registered read: three statements, three results

1. 1(A) Of the 218 domains that entered on 09-07, between 20% and 40% are cited again on 09-14: 34.4%, inside the band.
2. 2(B) Of the 334 incumbents, between 65% and 85% are cited again: 72.2%, inside the band.
3. 3(C) Single-answer entrants survive at a lower rate than both multi-answer buckets: 23.2% against 52.9% and 57.1%, as stated.

The read held. The bands are re-registered for the 2026-09-21 run rather than narrowed after a sixth pass: (A) of the 299 domains that entered on 09-14, between 20% and 40% are cited again on 09-21; (B) of the 316 incumbents, between 65% and 85% are; (C) single-answer entrants survive at a lower rate than both multi-answer buckets. A fourth statement is added for the contact read: (D) of the 09-14 entrants and climbers fetched on 09-21, between 2% and 6% publish a named person at a US, UK or Canadian address, the span of the four reads so far. Any of the four outside its statement is a falsification and will be published as one. The run fires 2026-09-21 around 13:07 UTC; the read publishes that day or the next.

Artifact

`content/news/entrant-survival-cut-2026-09-14.json` holds the eight run summaries, the six transitions, the per-engine, per-breadth and per-answer-count survival tables, the 75 survivors with their 09-07 and 09-14 engines, the contacted-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-2026-09-14.json`, which are themselves produced by `growth/citations/cited_domain_cut.mjs` from the engines' returned citation arrays. The contact sweep is `growth/citations/entrant_sweep.mjs` writing `address-availability-cut-entrants-2026-09-14.json` (every fetch, address and class recorded per domain); the hand adjudication with a reason per rejected domain is in `growth/citations/pitch-batch-2026-09-14.json`. This post is generated from those files by `content/news/entrant_survival_week6_post.mjs`. Written without em dashes.