Brands ChatGPT rarely names from memory are rarely named even when it searches and finds pages about them, and our ChatGPT column does not record which answers came from memory.

Meta moved the start date of its updated Commercial Terms and Self-Serve Ad Terms from 30 October to 20 November. Per Nieman Lab (via the article), nothing changed except the date as of 23 October, and the US English previews still do not mention robots.txt. A Meta spokesperson said the delay is to give "more time for us to explain to our partners and stakeholders what's changing and what's not"; the UK CMA and the European Commission both say they are still looking at the terms.

ChatGPT's newest entry (26 October) is GPT-6.1 Terra rolling out in Work and Codex, nothing on search or citations. Google's newest changelog entry (26 October) is the user-generated spam guide covered in the watchlist. Perplexity's changelog is unchanged and marks no current Sonar model as deprecated. Anthropic's newest post (26 October) is its S-1 filing, nothing on search; the deprecations table is unchanged (Sonnet 4.5 retires 30 November, Haiku 4.5 still Active with no notice). Cloudflare's newest posts are about its Workers AI cache and file serving.

Method in one clause: 57 brand and product comparison questions to logged-out ChatGPT in August 2026, 3,115 brand and question pairs, answers with and without search compared, plus two smaller replications (100 prompts through the API with search switched on and off, and 200 prompts from Graphite's own tracker). The headline figure: brands substantially mentioned on at least 15% of the pages ChatGPT cited still had a median of only 11% visibility with search when ChatGPT named them in under a quarter of its no-search answers. In every group, brands named more often from memory were named more often with search. Separately, of about 957,000 background searches ChatGPT ran for 22,953 comparison questions that named no brand, 59.6% contained a brand name and 22.7% used a site: search. Graphite also puts the gap between a model's knowledge cutoff and its release at two to five months for recent models, and writes "Do not expect impact from PKO in two weeks when the knowledge cutoff gap is five months." Graphite sells answer-engine services, the brands are large consumer and software names rather than coaches, there is no control for how well-known each brand is, and the write-up does not explain how no-search answers were separated in the main set. Correlation only.