ChatGPT Search’s underlying retrieval just changed: ChatGPT is now doing large-scale, targeted site scraping that upends GEO strategy
A single answer is no longer just a broad sweep of the web. ChatGPT may now discover candidate sources first, then lock onto specific sites to dig deeper. As the retrieval order shifts, GEO expands from page-level ranking into a three-layer contest: domain inclusion, page matching, and citation conversion.
- After August 8, Promptwatch saw the share of site: queries jump from a daily average of 0.368% to 16.78%, a roughly 45.5x increase.
- The most likely new flow is a two-stage process: first discover candidate sources across the web, then lock onto specific domains to pull docs, rules, and updates. OpenAI hasn’t confirmed the internal implementation.
- GEO thus becomes a two-layer fight: domains first get selected, then individual pages compete for citations. Sites should prioritize building crawlable, verifiable first-party content systems.
ChatGPT Search is changing how it searches the web in a way that warrants a fresh look from every site owner.
The old mental model of AI search was simple: the model grabs a few relevant results from across the web and assembles them into an answer. But continuous monitoring of the actual ChatGPT interface by Promptwatch shows that after August 8, ChatGPT began generating backend queries with site:domain.com at scale. The system now often locks onto a specific website first, then digs deeper within that domain for information.
The "targeted domain mining" in the headline refers to this domain-restricted retrieval pattern. It doesn't bypass site permissions, and it doesn't mean ChatGPT crawls an entire website. The change is in the order of operations: the system may decide which sites to visit first, and only then decide which page to cite within those sites.
This shifts the competitive logic of GEO. Sites used to compete primarily on page-level ranking. Now they may first need to make it into ChatGPT's candidate domain list before their pages even get a shot at the next round.
What actually changed in ChatGPT Search
Promptwatch is a GEO platform that tracks visibility in generative search. It runs client-defined questions repeatedly across real interfaces like ChatGPT, Claude, and Gemini, recording the answers, citations, and backend search queries it can observe.
Its public 16-day dataset shows a striking break in the trend. Computing unweighted daily averages separately for the 13 days before August 8 and the 3 days after:
- Queries containing
site:as a share of all fanout queries jumped from 0.368% to 16.78%, roughly 45.5 times the previous level; - Search responses that included at least one
site:query rose from 0.345% to 12.1%; - The average fanout per search response increased from 1.06 to 1.837, up about 73%.
The denominator here is the easiest thing to misread. The "roughly one in six" figure refers to the expanded backend search queries, not one in six of all ChatGPT answers. The share of search responses that actually contain a domain-restricted query is closer to one in eight.
The first raw chart shows the shape of the change: it hovered near zero for weeks, then spiked almost vertically to 16–17% on August 8. A step change like this, completed in a single day, doesn't look like the gradual result of many sites optimizing. It looks like a coordinated adjustment on ChatGPT's side, either in search strategy or in how queries are displayed.
The second raw chart shows that the increase on that same day wasn't just about the site: string. The total number of search queries triggered per response also rose from about 1.08 to 1.83. That suggests domain-targeted search is layered on top of the existing search process, not replacing ordinary search outright.
