Claude's newer model asks the user for more detail in six of ten answers, and that is the model our Claude column has to move to by 30 November.

No platform change today that touches search or citations. ChatGPT's newest note is still 9 October. Google's changelog newest is still 8 October. Perplexity's October entries are unchanged and say nothing about Sonar. Anthropic's two 14 October posts are about the US midterms and a Singapore office. On the deprecations page, `claude-haiku-4-5-20251001` is still Active with "Not sooner than October 15, 2026", which is today, and `claude-sonnet-4-5-20250929` still retires 30 November with `claude-sonnet-5-5` as the named replacement. 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: 711 ordinary questions, each run 20 times on Sonnet 5 and on Sonnet 5.5 through Anthropic's API, single turn, web search on, 28,440 answers in all. Sonnet 5.5 searched in 77.0% of answers against 74.8%, but ran 1.17 searches per searched answer against 1.83. It asked the reader for more details (budget, use case, location) in 60.6% of answers against 12.5%. It cited about the same number of links (about 12), and the most-cited pages were younger: a median of 107 days old against 164. In Profound's words, "Sonnet 5.5 asks the reader for more details about their situation". Limits: vendor data, no significance tests, one question sample not split by industry, single turn only.

New since yesterday: a measurement with dates, and a mechanism. Of the ChatGPT answers Profound collected that had citations, about 98% included at least one tagged link on 8 October, 75% on 9 October and 66% on 14 October. Answers using the new "Intelligent UI" layout went from 2% to about 35% over the same days. Links shown directly in the answer were often untagged; links inside a product or information panel sometimes still carried the tag. Profound: "People are still being recommended products, clicking through to websites, and making purchases." and "we haven't seen official communication confirming a fix or timing." Limits: no total sample size given, one week of data, still no OpenAI statement.

An agency that sells GEO audits and a free analytics tool ran the same 250 shopping prompts before and after 10 October. The brand was named in 40% of answers before and 43% after. In a live test of nine ChatGPT answers with 47 outbound links, none carried a referrer or a tag. One store's ChatGPT arrivals fell 66% (9 to 13 October against 1 to 7 October) and a second store's fell 57%, while Gemini arrivals on the first store rose 16%. Separately, "Shop at" buttons appeared in about one in five shopping answers from 10 October, and 56% of them sent shoppers to big-box retailers. Limits: product brands only, two stores, four days of button data, which the author calls provisional.