Google appears to be testing AI-written descriptions in the source cards beside AI Overviews, so the line under a cited business may no longer be that business's own words.

Brodie Clark posted on X that the source cards beside some AI Overviews now show a generated description rather than text from the cited page; he wrote that the text "is being summarised based on the information that Google thinks is most relevant to the query." Schwartz saw it on some of his own test searches but not all, and the summaries do not appear word for word on the cited pages. Google has not commented, and it is unclear whether this is a test or a rollout. Two observers, no numbers.

OpenAI added 26 markets, bringing ChatGPT ads to 89 countries, and said more than 100,000 advertisers have run campaigns (a lifetime count, not current spend). Ads show only on the Free and Go plans, are matched to the conversation rather than typed keywords, and OpenAI states "advertising does not influence the answers ChatGPT provides." Buying is self-serve in Ads Manager or through partners.

ChatGPT's newest entry is still locked chats (20 October). Google's changelog still ends at the 21 October controls page. Perplexity's newest entry renames the Agent API "preset" field to "effort" (naming only) and the changelog still marks no current Sonar model as deprecated; a September notice retires several older OpenAI model IDs on its platform on 24 October. Anthropic's newest posts (22 October) cover donating Agent Skills to a foundation and a life-sciences call. Bing's webmaster blog still ends at 13 October. Cloudflare's 22 October post is a compliance certification.

Method in one clause: three of Otterly's own blog posts got unique codes in og:title, og:description and the visible body on 6 October (a fourth left untouched), then 25 prompts a day from 6 to 18 October went to ChatGPT, Claude, Gemini, Google AI Mode, AI Overviews, Copilot and Perplexity (free logged-out versions through Otterly's platform, Claude by API with web search), 2,275 answers, plus 100 manual logged-in tests on 16 and 17 October. Across 504 questions that gave the URL and asked for the Open Graph text, no engine returned the correct og:title or og:description code. Engines passed off the body code as the Open Graph value 146 times; ChatGPT returned the body code in 36 of 36 runs and AI Mode in 33 of 36. The one exception: logged-in Claude with page access got 12 of 12 correct. Limits: one site, one run per prompt per day, US only, Perplexity data missing until 12 October, AI Overviews never triggered on prompts containing a URL, and the study measures whether the text reaches answers, not effects on citations. Otterly sells the platform used.

Method in one clause: a panel of 1,002 US owners of local businesses with 1 to 50 staff who make the marketing decisions, fieldwork dates not stated. Asked for their top two sources of new customers, owners named word of mouth (35%), social media (33%), AI tools (23%), their website (22%) and Google (22%). By industry, AI tools came in at 35% for home and garden services, 29% beauty, 28% automotive, 17% retail. The report does not say which AI tools, or how owners know. BrightLocal's own consumer data shows 61% of consumers started their most recent local search on Google.

Method in one clause: 400 fixed commercial keywords in banking, software and ecommerce, tracked monthly on US desktop from September 2025 to August 2026 with Semrush data. AI Overviews appeared on 45.75% of the keywords in September 2025 and 88% in August 2026 (97.08% weighted by search volume). "Best X for Y" searches reached 96.12%, and "reviews" searches rose from 30.14% to about 90%. No services, local or coaching keywords were in the set.

In July 2026 an AI Overview said Feature Upvote had no API a week after it launched one, citing old blog posts. The founder corrected the cited pages and submitted them for indexing by hand; the overview answered correctly within an hour, became consistent "a few more days" later, and was still right on 10 September while citing different pages. Measurement was screenshots in incognito tabs, and the author says the founder is a friend. One case, no control.