On 8 August ChatGPT stopped only searching the open web and started sending one in six of its background searches straight at a named site. A separate study of 129 million citations found trade publications out-cited mainstream media by 213%.

When ChatGPT answers a question it writes its own background searches first. Promptwatch tracks what those searches look like. On 8 August 2026 the share of them scoped to a single domain with the `site:` operator went from 0.37% to 16.8% in one day, roughly a 46-fold jump, after sitting between 0.3% and 0.5% for weeks. Over the same day the average number of background searches per answer rose from about 1.08 to about 1.83. A dip to 0.15% on 3 to 5 August suggests a staged rollout. The change landing inside a single day points at a model or system-prompt release on OpenAI's side, not drifting behaviour.

Caveats that have to travel. Promptwatch describes its collection as "millions of prompt responses, citations, and click data from the actual user interfaces of major AI platforms" over a 26 billion data point base, but publishes no sample size, no date range and no method for this specific measurement, and states no limitations on this page. OpenAI has confirmed nothing. Promptwatch does not say which domains ChatGPT scopes to, which is the number that would actually settle the question.

Google's own dashboard records the August 2026 spam update as starting 18 August and running 2 days, 16 hours. That closes the item left open on 19 August. For comparison on the same page, the March 2026 spam update ran 19 hours 30 minutes and the June one ran 2 days 1 hour, so this was the longest of the year without being extraordinary.

The entry now reads: "A logging error caused a decrease in impressions on the Generative AI performance report in Search for data from August 13 - August 17, 2026." On 18 August the same entry was open-ended, saying only "starting on August 13". A second, separate entry covers a Discover performance report error on 13 August alone. Neither is marked resolved, but the Search one is now bounded.

Otterly.AI supplied 129.3 million citations captured during June 2026 across seven platforms (ChatGPT, Google AI Overviews, Google AI Mode, Perplexity, Microsoft Copilot, Gemini, Claude), covering 19,812,878 cited URL records. Press Ranger supplied 91 confirmed licensing agreements between AI companies and news publishers with public evidence through 28 July 2026, mapped to 314 publisher domains. Citations were normalised per cited page so that a big publisher does not win by publishing more. A separate industry pass ran 16 US industry prompt sets collected 8 to 14 August 2026.

- Pages from OpenAI-licensed publishers earned 10.2 citations per page on ChatGPT versus 6.9 for unlicensed, a 48% premium. Publishers signed with OpenAI and nobody else were 112% ahead on ChatGPT. - OpenAI was the only licensor with a home-platform advantage. Google-licensed publishers got a 0% premium on Google's surfaces. Perplexity's licensed publishers landed at roughly parity. Microsoft looked like +85% until the OpenAI overlap was removed, at which point Microsoft-only licensees sat 56% below unlicensed. - Trade and niche publications drew 213% more citations than mainstream media across the 16 industries, and dominated 15 of them. Gemini showed the strongest trade-media preference at 82.2%, Google AI Mode 69.5%. Finance and insurance was the single exception, where mainstream led at 52.2%. - What licensed publishers got cited FOR was overwhelmingly commercial evergreen content, 58.8% of the total. Best-of lists alone were 38.1%. Dated news articles were 5.1%. Press releases and wire copy were under 1% for both groups. - News and media content was only 7.2% of all 129.3 million citations. Claude cited news most at 14.0%, ChatGPT 8.3%.

Caveats. The authors say this outright: "Descriptive data: association, not causation" and "a one-month snapshot". Five media groups account for 69% of all licensed-publisher citations, so licensing and sheer corporate scale are tangled together. Attribution accuracy is stated at 90.8%. The citation data is June 2026 and the industry data is August, so the two halves are not the same window. And Otterly.AI sells AI visibility monitoring, which makes this a competitor publishing research, though the method is more disclosed than the category norm.