Four assistants, almost no shared sources, and Google tests links inside AI Overviews that lead back into Google.

Spotted by Gagan Ghotra on X: normal-looking anchor links at the bottom of AI Overview answers that open AI Mode with follow-up questions rather than an external page. His words: "in AI Overviews answers Google is now testing showing anchor links which instead of taking to a webpage takes to AI Mode." It is a test, with no Google statement. The next day the same site carried a conflicting data point: Malte Landwehr of Peec AI posted a chart on X claiming AI Overviews went from "almost 0%" to "over 25%" containing external links. Schwartz wrote that it "doesn't 'feel' right" and disclosed a small investment in Peec. The X post could not be read directly (paywalled to our fetcher), the chart's time period and method are not stated, and we treat it as unverified vendor data.

Checked against the live repository with the token this morning: ten remote branches, none named for perplexity, sonar or agent; `main` still at `20b1172` dated 22 August; `src/adapters.ts` line 14 still `perplexity: "sonar"`, line 91 still `/chat/completions`. New since yesterday, from the changelog: the Agent API now also offers GPT-6 Sol, GPT-6 Luna, Claude Opus 5.5 and Grok 4.7. Separately, OpenAI launched GPT-6 Sol and Luna on 22 September (TechCrunch, Lucas Ropek), with Luna reaching Free and Go users on desktop. The same adapter file pins the ChatGPT cell to `gpt-5.1` on line 12.

Fifteen fixed commercial prompts run on ChatGPT, Microsoft Copilot, Google and Perplexity on 6 June 2026, producing 589 citation observations across 528 unique URLs. Verbatim: "mean pairwise Jaccard similarity was 0.0079, the median was zero, and 84.9% of engine pairs shared no cited URL." Top-five overlap "was zero in all 60 pairwise comparisons", "96.4% of observed URLs appeared in only one engine", and a single engine captured only 11.4% to 42.6% of all URLs found. Same engine, one day apart: "67.0% mean URL-set turnover." Limits: one day, 15 prompts, all about AI visibility software (the author calls it "unusually self-referential"), the Google surface is not specified, and it measures cited URLs, not which businesses were named.

34,960 unbranded prompt-engine observations from 75 client projects, 2,854 prompts, GPT and Gemini, June to September 2026. Verbatim: when neither the brand nor its domain appears in the observable retrieval path, "target mention rates are 2.8% for GPT and 3.8% for Gemini. With an own-domain citation but no branded fan-out, they rise to 49.0% and 58.4%." With both, 91.4% and 100%. Within the same prompt and engine across repeated runs, own-domain exposure is "associated with a mean mention-rate increase of 40.2 percentage points on GPT and 49.0 points on Gemini." The author states these are "observational associations, not a randomized causal effect." Industries: legal and regtech, IT, travel, beauty. No local services and no coaching.

1,600 prompts, 16 US industry reports, seven engines daily, median daily citation shares. LinkedIn "rose 39% through September, from 0.407% of all citations on 1 September to 0.564% on 20 September"; YouTube fell 42% (2.221% to 1.292%), including what Otterly calls "a 72% cut in 48 hours" on AI Mode between 13 and 15 September. Vendor data, no local or professional services category.