Studies
How the AI engines choose the names they recommend, and what you can do about it.
Asked the same question on the same day, two AI engines share 7.3% of their cited sources on average, and ChatGPT at most 3.5% with each of the other five. Measured on 2,847 answers, August 15 to September 21, 2026.
Continue reading
Between August 10 and 17, 2026, Perplexity stopped citing YouTube videos: from 61.5% of its sourced answers to 0.7%, on the same questions, in six languages. On the model Google now serves to the public, youtube.com became the source Gemini cites most. What changed, what held, and where a video still reaches AI answers.
Continue reading
On 3 August 2026 we put to AI engines eighteen questions business owners ask about them, three times each: 324 answers. Asked about a single place to submit your information, they almost all say there is none. All of them send you to the web: your own website first, then a short list of places run by others, which changes from one AI to the next.
Continue reading
Search optimisation teaches the numbered title, AI optimisation teaches the question title. Neither holds: among the pages an engine had in front of it, a number at the head of the title measures an odds ratio of 1.02 and a question mark 1.10. What decides is the match between the shape of the title and the shape of the question, and that match flips the sign: a superlative title is cited 24% of the time on a "how" question and 57% on a ranking question. Measured on 4,791 answers from six engines, 6 July to 2 September 2026.
Continue reading
A month ago, the gap between the AI engine that names the most sources and the one that names the fewest ran from 1 to 5.6. It now runs from 1 to 5.7, and no engine changed rank. Underneath, Perplexity gained four sources per answer and Gemini stopped citing in one answer out of six. Measured on 3,705 answers between 2 and 31 August 2026.
Continue reading
The same twelve buying questions, put to ChatGPT, Claude and Perplexity on the same day, do not return the same market. Salesflare, Nutshell, OnePageCRM, Insightly, NetHunt and EngageBay take 4 to 10% of the answers on Claude and Perplexity, and 0% on ChatGPT. The cited sources say why: ChatGPT reads the large vendors' own sites, Claude reads the independents' blogs, Perplexity reads Reddit. Measured on 108 answers on 2 September 2026.
Continue reading
Second wave, 3,849 answers archived between 1 and 29 August 2026 across six AI engines. reddit.com moves from 102nd to 11th on ChatGPT, takes the top spot on Perplexity, holds first on Gemini and stays at zero on Claude. What moved, by how much, and the licensing and court record dated alongside it.
Continue reading
Asked the same question on the same day, two AI engines share 6.4% of their cited sources on average, and no pair of the six reaches 13%. Measured on 2,883 answers, July 27 to August 14, 2026.
Continue reading
Twelve entities sorted by how much independent coverage exists about them, put to six AI engines on 6 August 2026 with a neutral question and an adversarial one: 144 answers. Every thickly covered entity drew criticism unprompted. On thinly covered names the criticism that comes back belongs to somebody else, sourced and stable.
Continue reading
Five recommendation questions put on 2 August 2026 to the AI engines that search the web: 225 answers, 1,813 cited sources, 264 domains. The professional's own site holds 61.9% of them, the directory 30.4%, the press 6.0% and the social network 1.7%. The breakdown by trade, and the domain classification method.
Continue reading
An answer from one AI engine names 13 sources. An answer from another names 2. Between the two sits a factor of 5.6, and two numbers explain all of it: AI Coverage, how often an engine cites anyone at all, and AI Density, how many it names when it does. Measured on 1,949 answers between 6 July and 1 August 2026.
Continue reading
Across 204 answers from six AI engines to 20 category-discovery questions in leisure travel, 1,943 cited sources partition mechanically: 19.5% are the site of a professional the answer ranks, 80.4% are third parties. The 397 third-party domains, named and counted.
Continue reading
Across 1,427 archived answers, reddit.com tops Gemini's citations, ranks second on Perplexity, 102nd on ChatGPT and never appears on Claude. The measured map, and the licensing, robots.txt and court record behind it.
Continue reading
923 archived answers, four engines, 2,049 domains cited: the web AI answers stand on is far more open than assumed, and each engine trusts a different profile of sites. Measured in July 2026.
Continue reading
Generative engines learn who you are from co-occurrence, not from hyperlinks: an unlinked corroboration already moves their answers. What the 2024-2026 studies, the engines' documented pipelines and our own surveys show, and what the link still buys, door by door.
Continue reading
Across 883 answers archived between July 10 and 24, 2026, youtube.com is the most cited source on Gemini and on Perplexity, and absent from 1,655 citations on the other engines. What part of a video the engines read, and what that makes worth doing.
Continue reading
For two years, AI assistants leaned on classic search indexes, so ranking well meant being visible in AI answers. That proxy is breaking: each assistant now assembles its own retrieval pipeline. Where AI answers come from in 2026, and what it changes.
Continue reading