ChatGPT Went From 2% to 61% of Sessions Running Multiple Searches. The Reads Multiplied. The Pay Did Not.
A single default-model change inside ChatGPT has rewritten how much of the web the assistant reads to answer one question.
According to Peec AI's Primary Sources report for Q3 2026, released to subscribers on 15 November, sessions in ChatGPT's default free experience that ran more than one web search rose from 2% in July to 61% in September (vendor-stated). The change followed 8 August, when GPT-5.6 became ChatGPT's default model.
For brands, that is a visibility problem. For publishers, it is a volume problem. More searches per answer means more pages fetched and read by AI agents, and the economics of each of those reads has not moved at all.
What did Peec AI find?
Peec AI is an AI search monitoring vendor. Its report combines two datasets. The first is a survey of 700 marketing decision-makers with responsibility for AI search, conducted for Peec by Sapio Research across the US (300), the UK (200) and Germany (200). The second is Peec's own analysis of platform behaviour from July to September 2026, covering ChatGPT, Claude, Gemini, Google AI Overviews and AI Mode. Perplexity was excluded for data quality reasons.
The headline platform finding is the jump in multi-search sessions after the GPT-5.6 switch: from 2% to 61% in two months (vendor-stated). Peec also reports that product pages roughly doubled their citation volume in ChatGPT after the switch, while Reddit citations fell (vendor-stated).
On the survey side, only 24% of companies say they can prove a return on AI search spend. That rises to 43% where AI search is integrated into the wider marketing strategy, against 23% where it sits apart (vendor-stated, self-reported). The most cited reasons to invest were competitors appearing in AI answers (43%) and customers using AI tools to research or buy (42%).
What is query fan-out?
Query fan-out is the technique an AI assistant uses to split one user question into several sub-queries, run each as a separate web search, and combine the results into a single answer. A question about the best running shoes for flat feet might fan out into searches on pronation, specific models, prices and reviews.
Each sub-query returns its own set of results. The assistant then dispatches Live Search Agents to fetch and read the pages it selects. A session with one search draws on one result set. A session with five searches can draw on five.
Peec measures sessions running multiple searches, not pages fetched, so the exact uplift in reads is not published. The direction is not in doubt: more searches per session means more retrieval per answer (blankspace estimate).
Why did one model switch change so much?
Peec's fan-out analysis covers ChatGPT's default free experience, which ran on GPT-5.5 until 8 August and on GPT-5.6 afterwards. Nothing about users changed in that window. The model did. The new default decided far more often that one search was not enough.
That is the uncomfortable lesson for anyone who depends on AI platforms. How much of your content gets read is now set by model releases you are not warned about, not by your audience. A third of surveyed companies (33%) name rapid platform change as a barrier to AI search work (vendor-stated).
Is Google doing the same thing?
Yes, by its own account. On 12 November Google brought Gemini 4 Hydrogen to Search globally and called it the biggest model upgrade to Search so far. Robby Stein, Google's VP of Product for Search, wrote that the model strengthens query fan-out by running simultaneous queries across related subtopics and connecting users to a broader range of sites. Google published no figures on how many more searches or pages that involves.
So the two largest answer engines moved in the same direction within the same quarter: more searches per question, more pages read per answer.
What does this mean for publishers?
More fan-out means more reads, not more readers. When an AI agent fetches a page to build an answer, it does not execute JavaScript. No analytics fire, no ad tags load and no auction runs. The page is consumed, the publisher records nothing, and the user stays in the chat window.
This is not a fringe flow. OPA Europe and Rossel found in January 2026 that one in 50 publisher visitors is already an AI bot, and that publishers receive seven times more AI bot traffic than other industries (named-party figures). The IAB Tech Lab puts publisher losses from AI bot traffic at around $2 billion a year (named-party figure). A default model that searches several times per session pushes every one of those numbers the same way.
Peec's citation shift adds a second edge. If product pages are gaining citations as forum content loses them, editorial pages in between (reviews, explainers, buying guides) are being read in more searches while competing harder for the visible link. More reads and fewer clicks is precisely the gap a referral model cannot close.
What does this mean for advertisers and brands?
For brands, fan-out is a moving target. A sub-query the brand never tracked can decide whether it appears in the final answer, and tracking only the original prompt misses most of the search activity behind it.
Peec's survey suggests most teams are not set up for that. Of the companies that raised AI search spend in the past year (89%), three in four cannot show what it returned (vendor-stated). Malte Landwehr, Peec's CPO and CMO, summed up the practical takeaway: "Fix your product data before you write another thought leadership post."
The wider point for advertisers is that organic visibility now depends on being picked across many invisible searches. That is hard to plan and impossible to buy. Display, paid search and social do not appear inside the answer at all, which is why brand visibility in LLM answers is becoming its own line in the media plan.
What should you watch next?
Three things. Whether OpenAI or Google publish retrieval volume per answer; neither has. Whether other measurement vendors replicate Peec's multi-search figure on a wider sample. And whether publishers' own server logs show a step change in AI agent requests from 8 August. That last check is free, first-party, and the fastest way to test the finding against your own traffic.
the blankspace read: Fan-out is the clearest evidence yet that the value sits in the read, not the referral. ChatGPT now goes back to the web several times to answer one question, and every trip lands on someone's page. The platforms decided to read more. Nobody decided to pay more. A publisher that can see and price each agent read is the only kind that gains when a model gets hungrier. Everyone else just serves more machine traffic for the same zero.
Key figures
| Figure | Value | Provenance |
|---|---|---|
| ChatGPT sessions running multiple searches, July 2026 | 2% | vendor-stated (Peec AI) |
| ChatGPT sessions running multiple searches, September 2026 | 61% | vendor-stated (Peec AI) |
| Date GPT-5.6 became ChatGPT's default | 8 August 2026 | vendor-stated (Peec AI) |
| Product-page citation volume in ChatGPT after the switch | roughly doubled | vendor-stated (Peec AI) |
| Survey sample | 700 decision-makers (US 300, UK 200, DE 200) | vendor-stated (Sapio Research for Peec AI) |
| Companies that can prove AI search ROI | 24% | vendor-stated, self-reported |
| ROI proof, integrated vs separate AI search strategy | 43% vs 23% | vendor-stated, self-reported |
| Companies that raised AI search spend in the past year | 89% | vendor-stated, self-reported |
| Companies naming rapid platform change as a barrier | 33% | vendor-stated, self-reported |
| Share of publisher visitors that are AI bots | 1 in 50 | named-party figure (OPA Europe / Rossel, Jan 2026) |
| AI bot traffic to publishers vs other industries | 7x | named-party figure (OPA Europe / Rossel, Jan 2026) |
| Annual publisher losses from AI bot traffic | ~$2 billion | named-party figure (IAB Tech Lab) |
| Increase in reads per ChatGPT answer | not published; direction up | blankspace estimate |
Primary source, opened and verified: Peec AI, Primary Sources, Q3 2026: https://peec.ai/reports/primary-sources-q3-2026
Also opened and verified: Google, "Google Search gets a major upgrade with Gemini 4 Hydrogen" (Robby Stein, 12 November 2026): https://blog.google/products-and-platforms/products/search/gemini-4-hydrogen-search-upgrade/
