Of the 248 domains AI engines cited for the first time on Aug 31, 40 (16.1%) were cited in both of the next two weeks and 129 (52.0%) in neither. Missing week one was not final: 32 of 161 came back in week two

Method, stated before the numbers

The instrument is the one behind every weekly survival post. Twelve pinned questions a small business might ask about its own AI visibility go to ChatGPT, Perplexity, Gemini and Claude through their APIs, three samples each, 144 answers per run, on Mondays. A domain is cited in a run if it appears in at least one citation array an engine returned that run, with the hostname stripped of www. An entrant for run N is a domain cited in run N and not in run N-1. Each entrant is then followed through runs N+1 and N+2 and lands in one of four groups: cited in both, cited in week one only, cited in week two only, cited in neither.

The computation is `content/news/entrant_second_week_cut.mjs`. It reads only the eight stored weekly cut files and writes `content/news/entrant-second-week-cut-2026-09-18.json`. 21 self-checks pass, 0 fail. They include matching every cohort's week-one count to the survival cut we published on 09-14, and the four groups summing to the cohort in all five cohorts. The artifact holds counts only.

The Aug 31 cohort at two weeks

87 entrants were cited in week one and 72 in week two, so the weekly rate fell only from 35.1% to 29.0%. Only 40 domains are in both counts. Fewer than half of the week-one group repeated, 40 of 87 (46.0%), and 32 of the 72 domains cited in week two had been absent the week before. Incumbents of the same run are the comparison: 201 of 309 (65.0%) were cited in both weeks and 39 of 309 (12.6%) in neither.

For anyone tracking their own domain, the practical number is the third row. A domain that appeared once and then missed a week was cited again the week after in 32 of 161 (19.9%) cases. One missed week does not mean the citation is gone. One cited week does not mean it has settled either, since 47 of the 87 cited in week one dropped out in week two.

Five cohorts, the same shape

The both-weeks share runs from 13.4% to 19.9% across the five cohorts, and the neither share from 52.0% to 61.1%. In every cohort a week-one citation made a week-two citation more likely. The smallest gap between the last two columns is 26.1 points, and pooled it is 33.0 points (49.7% against 16.7%). Roughly one entrant in six is still being cited after two weeks without a gap.

By the engine that cited the entrant first

ChatGPT's entrants had the best first week of the four, 38 of 77 (49.4%). In week two that fell to 25 of 77 (32.5%), and 20 of the 77 were cited in both weeks. Gemini introduced the most entrants (114) and kept the fewest: 9 of 114 (7.9%) in both weeks. Gemini's entrants were also the most likely to return after a miss, 20 of 86 (23.3%), which fits an engine whose citations rotate week to week. A domain cited by two engines on Aug 31 appears in two rows, so the rows sum to more than 248. The Perplexity and Claude rows hold 28 and 35 entrants, and three domains move either row by ten points.

By how many answers cited the entrant in its first week

171 of the 248 entrants were cited in a single answer on Aug 31. 105 of those 171 (61.4%) were never cited again in the two weeks, against 30.8% and 31.6% for the two multi-answer bands. The single-answer band did not decay between the weeks: 42 cited in week one, 44 in week two. The multi-answer bands did. Their week-one advantage (61.5% and 55.3%) shrank to 38.5% and 34.2% by week two. Being cited in several answers in the first week predicts the next week well and the week after less well.

Some entrants are returning domains

An entrant is defined against the previous run only, so a domain cited in July, absent on Aug 24 and cited again on Aug 31 counts as an entrant. 110 of the 248 had been cited in some earlier run and 138 were new to the whole corpus. The returning group came back after a missed week more often, 20 of 70 (28.6%) against 12 of 91 (13.2%) for first-timers. Their both-weeks shares were close, 17.3% and 15.2%. A domain with a history of being cited on and off keeps doing that.

The pre-registered read

The 2026-09-07 cohort of 218 entrants gets its second week on the 2026-09-21 run. Three statements, written before that run fires.

1. 1(A) Between 10% and 22% of the 218 are cited on both 09-14 and 09-21. The five cohorts so far read 13.4 / 14.2 / 13.6 / 19.9 / 16.1.
2. 2(B) Entrants cited on 09-14 are cited on 09-21 at a rate at least 20 points above entrants that missed 09-14. The smallest gap so far is 26.1 points.
3. 3(C) At least 10% of the entrants that missed 09-14 are cited on 09-21. The five cohorts so far read 17.8 / 15.6 / 14.9 / 15.8 / 19.9.

Any of the three outside its statement is a falsification and gets published as one. The run fires on 2026-09-21 around 13:07 UTC and the read publishes that week, next to the weekly survival read it extends.

Artifact

`content/news/entrant-second-week-cut-2026-09-18.json` holds the headline cohort, its splits by first-citing engine, by first-week answer count, by engine breadth and by whether the domain had been cited in an earlier run, the incumbents of the same run over the same two weeks, all five two-week cohorts, the pooled row and 21 self-checks. It is produced by `content/news/entrant_second_week_cut.mjs` from the stored weekly cuts `growth/citations/cited-domain-cut-2026-07-28.json` through `cited-domain-cut-2026-09-14.json`. This post is generated from the artifact by `content/news/entrant_second_week_post.mjs`.