A quiet Saturday. A rank tracker admits a week of its "ranking drops" during Google's spam update were not real.

Morningscore, a Danish rank tracker, reports that from the evening of 24 September Google returned cut-short result pages to its automated searches. Across more than 4 million searches, about 60% came back with fewer than 40 results at the peak on 30 September to 1 October, against 1.2% to 1.5% before. In those short pages "8 of the previous top 10 sites were missing on average". Of the sites that vanished, 74% were back in the top 20 within three days. The company's verdict on itself: "Google caused the truncated results, but recording them as real ranking drops was on us." By 4 October the rate was back near 1% to 2%. Limits: one vendor, its own data, and "Google hasn't said" anything. The post does not mention AI Overviews. The start date is the same day the spam update began, and the post does not claim the two are linked.

No change on any of the three. ChatGPT's newest note is the 7 October GPT-6 entry. Perplexity's developer changelog has the same October entries and nothing on Sonar. Anthropic's table still shows `claude-sonnet-4-5-20250929` retiring 30 November and `claude-haiku-4-5-20251001` as Active with "Not sooner than October 15, 2026", which is five days away. The engine's `main` is still `06233ce` with `gpt-5.1`, `sonar` and `claude-sonnet-4-5` in `src/adapters.ts` and `claude-haiku-4-5` in `src/extract.ts` (checked today).

Sample: 9,471 English questions across 151 topics, 92,112 websites, five assistants (ChatGPT, Gemini, Perplexity, Google AI Overviews, Google AI Mode), collected January to May 2026, 86% from the US. Method: split each topic's questions in two, see which sites were cited in one half, test whether that predicted the other half. How often a site was cited on other topics was the strongest signal on every assistant, with correlations from 0.29 (Perplexity and ChatGPT) to 0.37 (Gemini). Covering a topic beat a stored 0 to 100 authority score on four of five, most clearly on Perplexity (0.25 against 0.15). ChatGPT showed almost no topic effect once reach was held level (0.043). The stated limits are unusually plain: "They show where citations cluster, not what happens when you publish more pages", and no confidence intervals are reported. The press release says January to June. The study page says the analysis covers January to May.

27 models, 27 search-diagnosis cases, 6,561 responses. With the deciding evidence present, models were right 99% of the time. With it removed, "appropriate recognition of insufficiency fell to 51.5%". Claude Haiku 4.5 scored 0.556 on the benchmark and Claude Sonnet 5 scored 0.728, while seven models scored 1.000. Limits: an agency's own benchmark, a fixed August 2026 set of responses, and a scenario pool it keeps confidential.