AINEWS
After OpenAI's August 6 model update, the share of background search queries ChatGPT runs internally using Google's site: filter — a search syntax that limits results to a single domain — jumped from under half a percent to 16–17% in a day, and
ChatGPT Search now takes orders about which sites to read. On August 6, 2026, OpenAI shipped GPT-5.6 Sol, the latest in the model line that powers ChatGPT for paying users. By August 8, a third-party dashboard saw the share of ChatGPT Search fanout queries carrying a site: operator jump from under half a percent to roughly 16–17%. The company has not described the change in any product note, and its system prompt is not public, so the only record of what shifted is the telemetry.
The site: operator is a Google search filter that restricts results to a single domain. In a normal web search it looks like site:nytimes.com climate, and the engine returns only New York Times pages. Independent AI-product observer Simon Willison argues ChatGPT's search tool now has the shape search(query, recency, domains) rather than encouraging a site: operator directly. Willison's analysis is the only public write-up of that inference, and OpenAI has neither confirmed nor denied the shape.
The signal comes from Promptwatch's ChatGPT site:operator fanout tracker. Promptwatch is a small vendor in the "GEO" space, the chatbot-era version of SEO, that monitors prompts run through its own automation. The dashboard shows a baseline of 0.3–0.5% throughout July, a brief dip to roughly 0.15% on August 3–5 consistent with staged rollout, and then a step change to 16–17% the day after OpenAI's August 6 GPT-5.6 Sol announcement. OpenAI's announcement said only that the update made the model "more reliable with facts and provide more focused answers" for Plus and Pro users. The mechanism is not in that sentence.
The number deserves three caveats. Promptwatch sells tools to publishers who want to influence chatbot answers, and it publishes dashboards like this one as content marketing. Its coverage is partial: it only sees prompts run through its own automation, so 16–17% of those prompts is not 16–17% of all ChatGPT Search queries. The vendor frames itself as a "Generative Engine Optimization" tracker, and the framing should be read as positioning, not industry standard. Treat the share as directional, not a census.
A model that uses site: more often will return more answers grounded in a smaller, more curated set of publishers. Site owners, brand teams, and anyone whose work gets summarized inside a ChatGPT answer now have a stronger reason to care which domains the model trusts on a given Tuesday. The fanout plumbing is one of the few public surfaces where that trust reshuffle shows up.
On August 18, ten days after the fanout jump, Promptwatch separately reported that ChatGPT appears to have reduced how often Reddit surfaces in those searches. More site:, less unfiltered forum output, after a model update that promised more reliable facts. The timing is suggestive, but there is no public evidence OpenAI made a deliberate call against Reddit, no specific relationship disclosed, and no contract change on the record. Correlation, not campaign.
The cadence is the part worth watching. OpenAI ships model updates on a roughly monthly clock, and each one is now a candidate reshuffle of which sites ChatGPT trusts. Only the dashboards that watch the fanout plumbing catch some of those shifts in real time; everyone else reads about them after the fact. The next GPT-5.x update is the next test.