Researchers got ChatGPT, Perplexity and Google's AI to cite a made-up concept within a week by posting it on sites those tools already read, while posts on a little-known blog got nothing.

This updates the 17 October item. Meta's revised Commercial Terms, still effective 30 October, now let it crawl business-linked sites "in order to provide, develop, improve, research, and maintain advertising and business products". The earlier list included "train", and that word is gone. The access language (automated software, periodic retrieval of site updates) is unchanged. A Meta spokesperson told The Verge the terms "only applied to our ads and business products"; Meta told Nieman Lab the first version "wasn't clear" that robots.txt is respected. The revised clause itself still does not mention robots.txt.

ChatGPT's newest notes are still 16 October (Slides, GPT-6 Pro). Google's changelog still ends at 16 October. Perplexity's newest entry adds tool approvals for MCP calls and still names no Sonar model. Anthropic's newest post is still 18 October, Bing's webmaster blog still 13 October. Cloudflare posted three items on 19 October (inference costs, an extraction model, a NetScaler patch), none about crawlers.

Method in one clause: 17,211 citations mapped across 10 AI search tools and 6,356 domains, then marked posts from ordinary accounts on nine often-cited sites (including Medium, YouTube and Sohu), five questions per topic run daily for a week, March and April 2026. Eight of 10 tools cited the made-up "Cactus Gravity Protocol" within seven days, including all four English-web tools. Answers linking to the posts: Grok 30 of 32, Google's AI 15 of 33, Perplexity 13 of 35, ChatGPT 7 of 20. Answers repeating a planted detail: Grok 53%, ChatGPT 10%, Google 6%, Perplexity 0%. A made-up keyboard drew no citations from any English-web tool, and an opinion post only from Perplexity (11.4%). In paired tests, a post on an often-cited site produced the first citation every time, while 20 or 40 posts on a self-hosted WordPress site got none. A $14 Chinese posting package delivered 13 posts in about 28 minutes, and one Doubao citation followed within an hour (a single instance).

Limits: no real brands or businesses; the Google product tested is not named; tests ran before Google's 15 May clarification that its spam policies cover manipulating AI answers; the authors say rates may drift; small cell sizes.

Method: 60 pages about fictional companies, one fact each, in five versions; 21,600 trials across six models with pages supplied directly, 1,644 live fetches, and 90 manual app queries. Body-only facts were cited 70% of the time; adding JSON-LD or restating the fact in prose both reached 95%, with no real difference between the two. Schema-only facts through live fetches: Claude 0 of 100, GPT-5.5 0 of 100, Gemini Flash 92 of 100. In the ChatGPT and Claude apps, body facts 5 of 5, schema-only facts 0 of 10. The author's line: "Facts that existed only in schema were answered 0% of the time, in the APIs and in the apps." Limits: fictional companies, one fact per page, small app runs, only 8 of 60 pages indexed by Google.

Method: 76,450 publicly shared Gemini conversations (November 2024 to October 2026) plus live signed-out runs. 80% of cited passages start in the first half of the page and 33% in the first 10%. 37% of Gemini's search queries add a year the user did not type; 20% add a country. 42.9% of searched exchanges since May show no visible source. Limits: the Gemini app, not AI Overviews; shared conversations are not a representative sample; the "writes first, then attaches a source" model is inferred.

Method: 155 Search Console properties from one tool account, weekly, June 2025 to September 2026, queries classified by the vendor's own tool. Chat-shaped queries (bare replies like "yes please", follow-ups) rose from 0.032% to 0.22% of the median site's impressions; their click-through rate was 2.9% against 35.6% for all queries. Google does not label AI Mode queries, so this is inference; the author sells the tool.

Search's share of North American publisher traffic fell from 21% in Q3 2025 to 12% in Q3 2026, while Chartbeat says overall traffic "remains steady". News publishers only, no sample size, search engines grouped together. Direction only.