ChatGPT labels each search with its own freshness window in days

An SEO can read how fresh ChatGPT expects a page to be for a given topic, by looking at a number ChatGPT writes next to each search it runs. Mark Williams-Cook describes the finding in a new Search Engine Journal piece, alongside a free Chrome extension and a Google bug bounty that both came out of the same approach.

Williams-Cook, the co-founder of AlsoAsked, explains in the piece how he and his team watch Google and ChatGPT by opening the browser's network tab, filtering to Fetch and XHR, and reading what travels back from the backend. He writes that the network data for a ChatGPT search carries the search query followed by a vertical pipe and a number. The number is the time window, in days, that ChatGPT wants results from.

Williams-Cook gives measured examples. A question about whether Nvidia is overvalued gets a window of two days. Premier League weekend results get seven. "Best AI live chat software" gets 30. Nvidia earnings guidance gets 90. A Reddit lookup for live chat support opinions gets 365, pinned to reddit.com. The claim is that these numbers come straight from the system, so for a given topic a content team can see how recent the information actually has to be to qualify.

The $13,337 Google find

Williams-Cook also recounts how his team came across a Google callback URL carrying a parameter named _fmt=prog. Removing the parameter made Google return structured data: site quality scores, inverse document frequency scores, an epoch timestamp for the last major page update, and the server IP. Google's Vulnerability Reward Program paid out $13,337, calling the exploitation likelihood "high" and the finding "an abuse-related methodology with high impact". From that data he built a classifier of Google's eight refined query semantic classes, fine-tuned on 4.8 million queries, and put it online as a free tool called QueryClassifier.

FanoutFox, and a 365-day Reddit window

The free extension in the piece is FanoutFox, built by Suganthan Mohanadasan. Williams-Cook writes that it reads a reader's own ChatGPT session and shows the fan-out queries, the sources fetched, who got cited, and who was only mentioned in the model's parametric answer with no retrieval at all. He gives an example where asking ChatGPT to recommend a self-help book returns an answer with no web search behind it, which tells a brand that the response is coming from training data rather than from any page it could update next week.

Williams-Cook also cites Tomek Rudzki's research for Peec AI, which found that ChatGPT's result_source field carries four values, Labrador, Bright, Oxylabs and SERP, and that Labrador is a family of vertical indexes covering the general web, PDFs, YouTube, news, arXiv, Wikipedia, local, finance, legal, medical, shopping and images. He points to Petra Labs' analysis of 113,500 ChatGPT runs showing Reddit's share of final citations falling about 80 percent in one day, around when GPT-5.6 became the default model. Suganthan's data shows Reddit is still read in the fan-out searches at the same rate; what changed is how often ChatGPT chooses to quote what it read there.

For a team doing GEO work, the practical read is that the browser's network tab is a cheaper source of truth about ChatGPT than another industry study. Open DevTools, filter to Fetch and XHR, and read the search payload the model sent. The freshness window number tells a content planner whether a topic needs a page refreshed every two days or whether a page from a year ago qualifies.

Source

Search Engine Journal: Wiretapping the SERP: Reverse Engineering Network Data, by Mark Williams-Cook.