Most marketers treat a Google or ChatGPT session as a finished answer. Mark Williams-Cook treats it as a shipping crate full of signals. In a detailed Search Engine Journal breakdown, he explains how reading browser network payloads earned him a $13,337 Google bug bounty and opened a practical window into how ChatGPT actually searches.
The network tab is a free leak
Press F12 while running a search, open the Network panel and filter to Fetch/XHR, and the backend conversation becomes visible. Headers, payloads, event streams and responses show what the platform sent and what it got back. Most of it is tedious, but some of it is gold.
The Google leak that paid out
While working on AlsoAsked, Williams-Cook spotted a callback URL with a long parameter string. When he removed a parameter called _fmt:prog, Google returned structured data it clearly did not intend to share: site quality scores, inverse document frequency scores, last significant update timestamps in epoch time and server IPs.
The Google Vulnerability Reward Program panel later awarded $13,337, saying “exploitation likelihood is high” and classifying it as an abuse-related methodology with high impact.
From 4.8 million queries to a free classifier
The same digging showed Google sorts queries into eight refined query classes. Using a sample of 4.8 million English queries, Williams-Cook fine-tuned DistilBERT, a smaller BERT model, in about 50 hours on a gaming laptop. The result is QueryClassifier, a free tool that predicts how Google will classify a query. Because query class correlates with the SERP features Google shows, that is a planning signal for content teams.
ChatGPT’s freshness windows are readable
The same habit transfers to ChatGPT. Tools such as FanoutFox, built by Suganthan Mohanadasan, surface the fan-out searches behind an answer. The data shows how many days back ChatGPT reaches for different topics:
- 2 days for Nvidia stock price questions
- 7 days for Premier League weekend results
- 30 days for best AI live chat software
- 90 days for Nvidia earnings guidance
- 365 days for Reddit-based live chat support opinions
That replaces freshness guesses with objective windows, a strong input for deciding what to update and when.
Read what the model reads, not what it cites
One finding flips a popular assumption. Petra Labs’ analysis of 113,500 ChatGPT runs showed Reddit’s share of final citations fell roughly 80% in one day in mid-August 2026, shortly after GPT-5.6 became the default. But FanoutFox data shows ChatGPT still pulls Reddit heavily in the background. It simply cites the vendor’s page instead of the forum thread.
Counting links in an answer tells you what the model chose to show, not what it chose to read. That distinction changes budget decisions.
Start small, stack edges
Williams-Cook frames SEO as cumulative small advantages rather than copyable hacks. The network tab is one of the few places those edges still sit in plain sight. ChatGPT has already rebuilt parts of its search payload, so the fields change fast, but teams that inspect early can build a lead that is hard to imitate.
Source: Search Engine Journal
