Key takeaways
- ChatGPT's browser interface stopped exposing the
search_model_queriesfield around the GPT-5.3/5.4 rollout, and any GEO dashboard scraping that field for fan-out data lost its source overnight - The data wasn't deleted by OpenAI, it was hidden from the UI. It's still retrievable through the Responses API with web search tools enabled
- Field visibility has proven inconsistent across model versions since then. It reappeared under GPT-5.6 in some cases, then behavior shifted again with the August 2026 site: operator surge
- Promptwatch's own fanout tracking data shows average queries per response collapsing from 2.15 to roughly 1.0 between December 2025 and April 2026, then nearly doubling overnight on August 8, 2026
- No major GEO vendor announced removing fan-out tracking as a feature. The disappearance is a platform-side data source breaking, not a product decision, and it's a warning about building strategy on scraped internals
What actually happened
If you noticed fan-out query data quietly disappear from your GEO dashboard sometime in early 2026, you weren't imagining it. It wasn't a bug in your tool. It was a change on OpenAI's end that rippled through every product built on the same underlying trick.
For over a year, extracting ChatGPT's fan-out queries was almost embarrassingly simple. Open Chrome DevTools, go to the Network tab, filter for the conversation endpoint, and there it was: a search_model_queries field sitting right in the JSON response, listing every sub-query ChatGPT generated to answer your prompt. GEO tools, browser extensions, and a fair number of homegrown scripts built entire tracking pipelines on top of that one field.
Then GPT-5.3 shipped. The field got thinner. By GPT-5.4, it was gone from the browser payload entirely. Tools like the "ChatGPT Conversation Analyzer" extension started returning empty Queries columns. Anyone relying on that method for competitive intelligence lost visibility into one of the only windows they had into how ChatGPT actually retrieves information.
Chris Long at Nectiv was one of the first to flag it publicly and ship a workaround. Jerome Salomon confirmed something important shortly after: the data wasn't gone, it was just hidden from the UI. It still lived in OpenAI's API responses. That distinction matters a lot for anyone deciding whether this was a dead end or just a detour.
Why fan-out data was worth tracking in the first place
Query fan-out is the mechanism behind almost every modern AI search product. When you ask ChatGPT, Perplexity, or Google's AI Mode something that needs current information, the model doesn't just search for your literal question. It breaks the prompt into several parallel sub-queries, sends each one out, gathers the results, and stitches together a synthesized answer.
SEOcrawl's explainer on the mechanic puts it plainly: all of these systems decompose your prompt before they answer, so a page that only addresses the literal question is fighting at a disadvantage. If you don't know what the sub-queries look like, you're optimizing blind. You can't tell whether ChatGPT is searching for "best CRM for startups" or for "best CRM for startups pricing 2026 site:g2.com," and those two searches favor completely different pages.
That's why fan-out visibility mattered enough for entire dashboard features to be built around it. It showed which domains and page types the model targets, what modifiers it tends to add (years, "best," "vs," "pricing"), and how aggressively it validates claims against third-party sites using the site: operator. Lose that, and you lose a meaningful chunk of your ability to reverse-engineer what's actually driving citations.
The numbers behind the disappearance
Promptwatch's fan-out tracking data gives a concrete timeline of what happened to ChatGPT's search behavior through 2026, and it's messier than a clean "feature removed" story.
Average fan-out queries per response fell from 2.15 in early December 2025 to 1.84 by early March 2026. Then there was a genuine data gap. When fan-outs reappeared in April, they landed at exactly 1.0 query per response, described by Promptwatch as a markedly leaner search pattern than what came before. Query length collapsed too, dropping from roughly 117 characters per query down to about 53 once things stabilized in April, less than half the original length.
Then, on August 8, 2026, everything reversed. ChatGPT Search's use of the site: operator jumped from about 0.4% to roughly 17% of all fan-out queries overnight, a nearly 46x increase, and average queries per response nearly doubled from about 1.08 to 1.83 in a single day. You can see the full daily breakdown in Promptwatch's data on ChatGPT's site-operator fan-out surge. There was even a brief dip to 0.15% share between August 3 and 5, which looks a lot like a staged rollout or an internal experiment before the wider launch.
That same period lines up with a separate, related drop: average citations per ChatGPT response fell from about 6.4 sources the week before the GPT-5.3 rollout to 4.7-4.9 by late March, a roughly 27% decline with no recovery a month later, according to Promptwatch's citation drop analysis. The drop hit GPT-5.3, GPT-5.4, and GPT-5-Mini at the same time, which tells you it was platform-wide, not something specific to one model checkpoint. Promptwatch's framing on this is worth repeating here: citation behavior in AI search is a platform-controlled variable that can change overnight, which makes single-snapshot audits unreliable and continuous monitoring necessary.
For context on how unstable this baseline has been across the board, Promptwatch's data on average sources per response shows Perplexity holding almost exactly ten sources per answer day after day, while Microsoft Copilot has swung from under 2 to nearly 17 sources per response within a matter of weeks. ChatGPT's instability isn't an outlier, it's part of a pattern of platforms still actively re-architecting how they cite sources.
Third-party studies fill in more detail (with caveats)
Writesonic ran repeated studies tracking fan-out volume across ChatGPT model versions, and the swings are dramatic. GPT-5.3 Instant averaged about 1.0 query per prompt. GPT-5.4 Thinking jumped to 8.5 average queries per prompt, an 8.5x increase, alongside 109.4 average web results read and 14.8 average citations per response. Of GPT-5.4's queries, 37% used the site: operator to validate claims against specific domains.
Then GPT-5.5 walked much of that back. Site: operator usage collapsed to around 10-13%, down from GPT-5.4's 40%+ peak, a shift Writesonic called "the death of the site: operator" at the time. That headline didn't age well. GPT-5.6 "Sol," which replaced the old Thinking toggle with a Medium/High effort picker in mid-July 2026, blew past every previous record: 59.3% site: usage at Medium effort, 71.2% at High effort. In one example from the study, GPT-5.6 ran twelve separate site:-scoped queries against quickbooks.intuit.com for a single accounting-software prompt.
The same study found something that matters for content strategy beyond just fan-out counts: first-party (brand-owned) citation share rose from 47.2% to 58.1% at Medium effort, and the median age of cited pages nearly halved, from 88 to 52 days. AI search is increasingly favoring fresher content and rewarding brands that get cited directly on their own domains rather than through third-party mentions.
One caveat worth flagging honestly: Writesonic's methodology pulls the same conversation-endpoint payload that periodically breaks, the exact mechanism this whole article is about. The researchers say as much themselves, describing any single-run study, including their own, as a map rather than a measurement to the third decimal. Treat these figures as directional, not gospel.
What still works if you want fan-out visibility
The browser-scraping method is fragile by design, since it depends on a field OpenAI has no obligation to expose. But a few approaches still hold up.
The most reliable option is OpenAI's Responses API with web search tools enabled. Call https://api.openai.com/v1/responses with a web_search tool block, and the response includes a web_search_call item with an action.queries array containing the full fan-out set. Watch out for a common mistake here: there's also a singular action.query field that only holds the first query, which will quietly undercount your data if you grab the wrong field. Citations show up separately, in the final message's content[].annotations[] array, tagged with type: 'url_citation'.
A few practical notes if you go this route. Omit the user_location parameter and the API defaults to a US-based searcher, which will skew your fan-out results if you're tracking a non-US market. Cost is trivial, each prompt is a single API call, and running 20 or 30 target prompts costs pennies, though web-search-enabled calls run slower, often 10-30 seconds, and cost more than standard calls. The tradeoff versus the old DevTools method is real: the API is schedulable and structured, so you can pipe it into something like n8n or a spreadsheet and catch drift automatically instead of manually checking one conversation at a time.
A weaker but still useful fallback is CDN-level tracking. It won't show you the exact sub-queries, but server logs can show which pages got cited, when, and, via UTM parameters like utm_source=chatgpt.com, that the traffic actually came from ChatGPT. It's a consolation prize, not a replacement.
One more twist worth knowing: field visibility isn't permanently gone, it's inconsistent across model versions. Independent tracking from tools like GPTSpy found the search_model_queries field reappeared populated in the browser network tab under GPT-5.6. The practical lesson is to check your own network tab before assuming the data has vanished for good. It comes and goes.
Google's fan-out was never visible to begin with
It's worth separating ChatGPT's situation from Google's, because they're not the same problem. Google's AI Mode and AI Overviews use an equivalent technique, formally described in Google's patent as "query variant generation," but Google has never exposed those sub-queries in any public interface or devtools panel. SEOs are not losing access to something they used to have here, they never had it.
That's why tools like iPullRank's Qforia and WordLift's fan-out tool exist as simulators rather than extractors. They try to predict what Google's fan-out queries might look like based on patterns and prompt structure, not pull the real thing. According to Google's own Robby Stein, AI-powered search experiences including query fan-out now serve roughly 1.5 billion users a month, so the stakes of not having this visibility are high, even if the workaround is fundamentally a guess rather than a measurement.
Why no GEO vendor actually announced removing the feature
Here's the distinction that matters for evaluating your current tools: no major GEO dashboard vendor, not Profound, not AthenaHQ, not Semrush, publicly announced pulling fan-out tracking as a product decision. What happened is that any tool built on browser scraping or the conversation-payload trick inherited the same breakage when OpenAI changed its interface. This wasn't a vendor choosing to cut a feature, it was a platform-side data source disappearing out from under everyone at once.
That's the real lesson here, and it's bigger than fan-out queries specifically. If your GEO strategy, or your vendor's product, depends on scraping an undocumented field from a competitor's website, you're one interface change away from losing the data entirely. It happened with zero warning, and it will happen again with something else.
What this means for how you track AI visibility
The fan-out saga is a decent argument for relying on tools that build their tracking on top of more stable foundations, official APIs, crawler logs, and structured citation data, rather than scraping internal fields out of a browser's network requests.
| Approach | Reliability | What you get | Risk |
|---|---|---|---|
| Browser DevTools scraping | Low | Exact fan-out queries as seen in the web UI | Breaks whenever OpenAI changes the payload, no warning |
| OpenAI Responses API | Medium-high | Fan-out queries and citations, close approximation of UI behavior | May not perfectly match what the actual web interface returns |
| CDN / server log tracking | Medium | Confirms which pages got cited and when, not the queries themselves | Doesn't reveal the sub-queries at all |
| Continuous third-party monitoring platform | High | Ongoing trend data across models, normalized over time | Depends on vendor's own data pipeline staying current |
This is the gap that a continuous monitoring platform like Promptwatch is built to close. Rather than treating fan-out and citation data as a one-time scrape, it tracks prompt volumes, citation trends, and crawler activity across ChatGPT, Gemini, Perplexity, Claude, AI Overviews, and AI Mode over time, so a sudden shift like the August 2026 site: operator surge shows up as a trend line instead of a mystery. It also goes further than tracking alone, pairing that visibility data with content gap analysis and automated content generation so teams can act on what the data shows instead of just watching it change.
A few practical takeaways
If you're building GEO strategy around fan-out data, a few things are worth internalizing after watching this play out over 2026.
First, single-snapshot audits are close to useless for platform behavior that changes this often. A study run in March told you almost nothing true about August. Continuous tracking, even imperfect continuous tracking, beats a perfect one-time analysis.
Second, don't overfit to a specific query pattern. The LinkedIn post from Ishaan Shakunt on this topic makes a fair point: adding "2026" to your title because fan-out data told you ChatGPT likes recency modifiers might be one of the worst uses of this data if you're doing it mechanically rather than because it genuinely helps the reader. Use fan-out patterns to understand intent and structure, not as a checklist to game.
Third, if you're evaluating GEO tools for your own tracking, ask vendors directly how they source their fan-out and citation data. If the answer is "we scrape the browser," ask what their plan is for the next time OpenAI changes the payload, because it will happen again. You can compare a wider set of monitoring and optimization platforms in the GEO software directory at bestgeosoftware.com if you want to see how different vendors have handled this kind of platform instability.
The field visibility problem isn't solved. It's just better understood now, and there are workarounds that don't depend on OpenAI leaving a debug field exposed in a browser tab.