When OpenAI switched default models on March 4, the number of websites cited per response dropped by a fifth, and never recovered. But the citation drop is only part of the story.
We also reverse-engineered ChatGPT’s internal browsing tools, ran a honeypot experiment, reconstructed its system prompt, and released a new version of our ChatGPT Search Capture plugin.
What happened
On March 4, ChatGPT switched its default model from GPT-4o/5.2 to GPT-5.3 Instant. The result: the average number of unique domains cited per response dropped from 19 to 15, a decline of more than 20%.
Unique URLs per response followed the same trajectory, falling from 24 to 19. We tracked 400 daily prompts over 14 weeks, using monitoring data provided by Meteoria.
Why we care
ChatGPT has 900 million weekly active users. The citation surface in each response hasn’t changed, but fewer websites are sharing it. Same pie, fewer slices.
This likely reflects a structural shift toward higher-authority sources, but it also means fewer winners overall. Sites that don’t make the cut are losing visibility that was previously within reach.
We named this phenomenon after the “Bigfoot update” (identified by Dr. Peter J. Meyers of Moz in 2012), when Google would sometimes let a single domain occupy the entire first page of results.
ChatGPT now retrieves fewer domains per response, but the URL-to-domain ratio has remained stable at 1.26. Crawl depth per domain hasn’t changed. What has changed is how many distinct websites get a seat at the table.
GPT-5.4 Thinking amplifies the concentration further. The model uses “site:” operators to restrict searches to trusted domains and distributes its queries across often more than 10 “fan-out queries” per response, each targeting a specific source.
Independent log analysis by Jérôme Salomon (Oncrawl) confirms the trend. ChatGPT-User bot crawl volume has settled at a lower level since the switch to 5.3. Some pages simply aren’t being crawled anymore.
The cause goes beyond model updates: more than 90% of ChatGPT’s weekly users are on the free plan, and the default experience triggers fewer web searches, uses fewer queries, and produces fewer citations.
How ChatGPT Search actually works
Our study also includes a full reverse engineering of ChatGPT’s internal search system, called web.run. Before 5.3, the model sent compact text commands separated by pipes (fast|query|recency). After 5.3, it sends structured JSON objects with typed parameters.
This isn’t just a format change. It reflects a different architecture in how the model formulates and distributes its web operations.
The web tool now supports 12 operations, up from 4 (plus a separate widget system called genui). These include:
search_queryopenfindclickscreenshotproduct_query- Specialized widgets for sports, finance, weather, and more.
GPT-5.4 can chain 5 to more than 10…
Source link
Disclaimer
We strive to uphold the highest ethical standards in all of our reporting and coverage. We blogs.grocliq.com want to be transparent with our readers about any potential conflicts of interest that may arise in our work. It’s possible that some of the investors we feature may have connections to other businesses, including competitors or companies we write about. However, we want to assure our readers that this will not have any impact on the integrity or impartiality of our reporting. We are committed to delivering accurate, unbiased news and information to our audience, and we will continue to uphold our ethics and principles in all of our work. Thank you for your trust and support.
Website Upgradation is going on for any glitch kindly connect at [email protected]
