Do new AI-cited sources survive a week? 84 of the 279 domains that entered the cited set on Aug 17 were cited again on Aug 24 (30.1%); domains already cited two weeks running returned 75.0%

The read

Every Monday at about 13:07 UTC our audit engine asks the same 12 buyer questions about AI visibility to ChatGPT, Perplexity, Gemini and Claude, three times each, and stores every citation the engines return. Five runs are now on the ledger: 07-28, 08-03, 08-10, 08-17 and 08-24. Reading them as a sequence lets us ask a question the single-run cuts cannot: when a domain is cited for the first time, is it still there a week later?

Mostly, no. 231 domains entered the cited set on 08-03 and 68 came back on 08-10 (29.4%). 247 entered on 08-10 and 68 came back on 08-17 (27.5%). 279 entered on 08-17 and 84 came back today (30.1%). Incumbents, meaning domains cited in two consecutive runs, returned at 77.9%, 76.5% and 75.0% on the same three transitions. The corpus is stable in size (520 to 571 distinct domains per run) while roughly half of its membership turns over every week, and the turnover is concentrated in the newest members. Today's run has 559 distinct domains: 256 entrants and 303 incumbents, with 268 of last week's 571 (46.9%) gone.

By engine: Perplexity and Claude keep their new sources, ChatGPT and Gemini drop them

Splitting the 279 entrants of 08-17 by the engine that cited them: sources ChatGPT introduced returned 22 of 95 times (23.2%), 20 of those through ChatGPT itself. Gemini introduced the most, 130, and kept 37 (28.5%). Perplexity introduced 29 and kept 14 (48.3%); Claude introduced 31 and kept 14 (45.2%). That is the same ordering the within-run consistency cut found two weeks ago, where Perplexity held 63.8% of its citations across three identical asks and ChatGPT 7.4%: the engines that are consistent within a run are also the ones whose new sources are still there a week later. A domain can be counted under more than one engine, which is why the four engine rows sum to 87 survivors against 84 distinct.

Being cited more than once in week one is what predicts week two

Of the 211 entrants that appeared in exactly one answer on 08-17, 45 (21.3%) returned. Of the 38 that appeared in two answers, 23 (60.5%) returned; of the 30 cited in three or more, 16 (53.3%). Engine breadth points the same way on a small base: 81 of 273 single-engine entrants returned (29.7%) against 3 of the 6 two-engine entrants (50.0%, n=6). A single citation is closer to a coin the engine flipped once than to a position the domain holds.

Which entrants survived, and which sources never leave

Eleven of the 84 survivors widened from one engine to two in their second week, among them cruelx.com (1 answer on 08-17, 5 today, Gemini and ChatGPT), arobis.ai (3 to 4, Gemini and Perplexity), digitalsuccess.us (3 to 4, Claude and Gemini) and wix.com (1 to 4). At the other end, 165 domains have been cited in every one of the five runs. Today they are led by reddit.com (25 answers), zapier.com (23), trysight.ai (21), g2.com (20), useomnia.com (19), ayzeo.com (17), semrush.com (16), amplitude.com (16) and linkedin.com (16). The corpus has a stable core of roughly 165 domains and a weekly churn layer of 230 to 280 that mostly does not come back.

Two first-party notes for completeness. Of the 11 entrant domains we wrote to about their citation last week, 2 were cited again today, 18.2% on a sample too small to separate from the 30.1% base rate. And our own domain, askedabout.com, was cited in 8 of 144 answers today by ChatGPT and Perplexity, against 7 last week; it entered the cited set on 07-28 and has been an incumbent since.

Method

The instrument is the weekly self-audit we run on our own product: 12 pinned buyer questions about AI visibility (panel version 2026-06-27), asked to ChatGPT, Perplexity, Gemini and Claude through the same production configuration as the paid audit, three independent samples per question per engine, 144 answers per run. Citations are read from each engine's returned citation array, never parsed from answer text; Gemini's grounding sources arrive as bare hostnames and are counted as such. A domain counts once per answer it appears in, however many URLs from it that answer links, and is normalised to its hostname minus www. An entrant for run N is a domain present in run N and absent from run N-1; an incumbent is present in both; survival means present in run N+1, on any of the 12 questions. Today's run has 130 of 144 answers carrying citations (ChatGPT 36 of 36, Perplexity 36 of 36, Gemini 33 of 36, Claude 25 of 36), the same 130 as last week. The cut script and every per-run figure are in the artifact, which reconciles exactly with the published weekly corpus cuts (08-17 entrants 279, 08-24 churn 268).

Limits

• One sample per transition. Each survival rate is a single week-to-week comparison; three transitions in a 27.5–30.1% band is the evidence, not any one of them.
• Entrant is relative to the previous run only. A domain cited on 07-28, absent on 08-03 and back on 08-10 counts as an 08-10 entrant. We have not yet separated first-ever appearances from returns.
• Domain level, any question. A domain that was cited on one question on 08-17 and on a different question on 08-24 counts as surviving. The stricter same-question, same-URL survival is the 08-19 question-span cut's territory.
• One vertical. The 12 questions are buyer questions about AI-visibility tools and audits; the corpus is what the engines cite for that vertical, not the web.
• The denominator moves. Answers carrying citations ranged from 124 to 131 per run, so a run with fewer cited answers mechanically produces fewer survivors.

Pre-committed read

The claim as stated: a newly cited domain is cited again the following week 25–35% of the time and a domain cited two weeks running 70–80% of the time. Read on 2026-08-31 on the 08-24 to 08-31 transition (256 entrants, 303 incumbents). Entrant survival outside 20–40%, or incumbent survival outside 65–85%, falsifies the gap as stated and we say so in this post; inside both bands, the fourth transition is published beside these three.

If you want to know whether the engines cite your business at all, run the free 60-second check. The $79 audit samples 25 questions three times each across all four engines, so a citation that appeared once is not reported as a position you hold.