Google put numbers on how slowly it notices a change: about 30 days to re-read a page it already knows.

On the last day of Google's three-day Search Central event in Barcelona, Gary Illyes showed a table of how long Google's steps usually take. The figures, as carried by all three write-ups: a new URL is discovered in about 20 hours; a known URL is refreshed about every 30 days; a sitemap is processed in about 24 hours; a robots.txt change is picked up in about 24 hours; indexing end to end takes about 1.5 hours once a page is fetched; a title or snippet change shows in 1 to 2 days. Each row has a slow case, and for discovery and refresh the slow case is "weeks to never". Recovery after a core update is given as 3 to 6 months. Search Engine Journal states the limits: "Reported ranges are reference points, not deadlines", with no sample size, no measurement period and no definition of "typical". The source is an agency's recap of slides, not a Google document. The table is about Google Search. It does not mention AI Overviews, though the same recap records Illyes saying "AI results are built on SEO infrastructure".

The recap records Illyes saying "Google filters out 40 billion spam pages a day" and "Scaled content is now becoming more of a problem than link spam". It records Duy Nguyen of Google's search quality team saying "There have been more spam updates recently because there is so much more new content" and "The point is the quality, not who wrote it. Low quality content can be human or AI generated." The timing table gives a spam update as "1 to 2 weeks" in the usual case and "months (batch refreshes)" in the slow one. None of this names the September 2026 update. Google's incident for that update still shows one post, the 24 September launch, on day twelve of a rollout Google said "may take up to two weeks".

The table lists `claude-haiku-4-5-20251001` as Active with a tentative retirement date of "Not sooner than October 15, 2026". It is not deprecated, and Anthropic promises at least 60 days' notice, so nothing fails this month. It is worth a line because the same pattern played out last week: Claude Sonnet 4.5 passed its "not sooner than" date on 29 September and was formally deprecated on 30 September. In the engine, `src/extract.ts` line 129 sets `NER_MODEL = "claude-haiku-4-5"`. That is the step that pulls the names of other businesses out of each answer. Sonnet 4.5 is unchanged: deprecated, retirement 30 November 2026, replacement `claude-sonnet-5-5`. The engine's `main` is still `06233ce`.

Unchanged. The migration page still says Sonar requests "are being reformulated as Agent API requests, rolling out gradually by model" and names no model. No October changelog entry.

Otterly counted 3.95 million citations on German prompts across 16 industries and seven assistants, from 16 August to 28 September 2026. Verbatim: "64.9% of German AI Search citations point to a domain that carries a local region or language signal". Association sites take 8.9% of German citations against 5.3% in its US data, and Reddit 0.60% against 1.79%. The local share runs from 75.2% on Perplexity to 57.7% on Claude. This is vendor data, it counts cited pages and not named businesses, and the period sits partly inside Google's spam update.