Google is testing a slot at the top of its AI answers for one or two official pages, which puts a client's own website in play for questions about booking, prices and policies.

Google says it has made "continued and substantial increases" to links placed inside the wording of AI answers and will show them higher up more often, in both AI Mode and AI Overviews. It is testing a new display that shows one or two official links at the top of a response, with a line on why each is useful, for task questions where only one or two official sources exist (the example is an airline refund). It is also widening news carousels that favour a reader's Preferred Sources (use "more than 7x" since the start of the year, no method given) and testing creator names on AI Mode video results that link to a Google Search profile. The post gives no country scope for the test, says nothing about local businesses, and Google told Schwartz it is still working out which metrics to give site owners, so no click data for now.

This is an update on the research pilot announced in July, not a product or a setting. Cloudflare passes OpenAI metadata about pages, not page content: signals to skip wasted fetches, signals for "when a page is worth another visit", and harm signals in testing. A cohort of about 1% of Cloudflare's network took part, and the post says it helped OpenAI discover and index "hundreds of millions of new web pages". The post says "A signal about a page doesn't grant permission to access it." It does not say how a site joins, whether it is opt-in, or which plans qualify.

ChatGPT's only new note (15 October) is about voice and new voices. Perplexity's October entries add Nemotron 3.5 Super to its Agent API and say nothing about Sonar. Anthropic's one new post (15 October) is about independent evaluators. On the deprecations page, `claude-haiku-4-5-20251001` is still Active with "Not sooner than October 15, 2026", a date that has now passed with no notice, and `claude-sonnet-4-5-20250929` still retires 30 November. The engine's `main` is still `06233ce`, with `claude-sonnet-4-5` in `src/adapters.ts` and `claude-haiku-4-5` in `src/extract.ts` (checked today).

Method: GA4 data from the sites LLM Pulse monitors (count not given), several countries, 9 to 11 October against the same days in the two previous weeks. ChatGPT sessions fell "around 46%", more than eight in ten sites lost over 30%, and total traffic on the same sites moved about minus 2%. Perplexity and Claude referrals held steady. The author is careful: "matching dates is not proof", and "we cannot tell whether the visits are lost or just arriving untagged as Direct". No OpenAI statement.

Method: 138 projects for 130 businesses in 29 countries, tracked daily through Peec AI from January 2025 to 7 September 2026; 11,296 prompts, 394,180 answers, 2,283,923 citations across ChatGPT, Google AI Overviews and Perplexity (Gemini thinner). The business's own website made up 3.7% of citations and named competitors 16.9%. For 34.8% of businesses, their own site was not among the top 30 sources at all. Limits: an agency's client and pitch list, 104 of 138 projects in Lithuania, tables capped so the long tail is undercounted, and citations are not the same as being named.