Our OpenAI fanout data shows no rise in site: operator use

Promptwatch reported on 10 August 2026 that the share of ChatGPT fanout queries containing the site: operator rose from 0.37% to 16.8% on 8 August, and that average fanouts per response moved from 1.08 to 1.83. Our own capture of OpenAI fanout queries over the same window records a different pattern.

TL;DR: I think the trigger is the nature of the prompt. We do two modes of prompting one quantized/nuclear in nature for visibility tracking and the other type more verbose for ad-hoc citation mining.

What we measured

551,695 fanout queries captured between 1 March and 25 August 2026 across providers. The OpenAI subset is 196,692 fanout queries from three GPT-5 versions, called through the OpenAI Responses API with web search enabled and driven by a fixed set of monitoring prompts that run daily. A fanout query is the search string the model issues while composing an answer, stored as the API returns it. A query counts as site:-scoped if the string contains site: at any position.

OpenAI site: share by month

On 8 August, the date of the reported step change, 0 of 371 fanouts contained site:. The highest single day in August is 0.86%.

Fanouts per response

Counting only responses that ran at least one search, the average number of fanouts per response has fallen through the window.

site: use tracks the model version

Splitting the OpenAI subset by model version places the change at the version boundary rather than at any date in August.

gpt-5.2 and gpt-5.4 scope roughly one fanout in eighteen to a domain. gpt-5.5, which carries all of our August traffic, scopes one in 647.

Prompt mode

The OpenAI runs divide into the two prompt modes described above. Visibility tracking uses the short fixed prompts that repeat daily. Citation mining uses the longer ad-hoc prompts. Both modes ran through the window and both are inside every figure above. Visibility tracking starts on 21 April, so March has citation mining only.

Over the full window citation mining issued 106,323 fanouts at a 6.29% site: rate, visibility tracking 90,369 fanouts at 0.44%. In every month where both modes run, the citation mining rate is 7x to 91x the visibility rate. Both modes drop together from June, and the August citation mining figure is 0.91% against the 16.8% reported for ChatGPT.

Prompt length does not separate the two groups inside a mode. Citation-mining prompts that produced a site: fanout have a median length of 249 characters against 327 for those that did not. Visibility prompts sit at 24 characters against 23.

Other providers

The same measurement applied to the other two providers in the capture.

Anthropic issued no site:-scoped fanout in 33,872 queries.

Google runs in the opposite direction to OpenAI. It issued none in March or April, started in May, peaked at 0.167% in June, then fell to 0.007% in August.

Shape of the site: queries

Across all 7,078 site: fanouts in the OpenAI subset, 6,938 strings are unique, so the same scoped query almost never repeats. The mean number of domains named in one query is 1.13, with 274 queries (3.9%) naming two or more and a maximum of 10 chained with OR. The median site: query is the 9th fanout of a response that issues 17.6 fanouts in total, and 6.8% of them are the first query the model runs. The scoped searches therefore sit in the middle of a longer sequence, after other searches have already returned.

Where the two datasets differ

Promptwatch samples the ChatGPT consumer interface. We sample the OpenAI API with web search. The prompt sets also differ: ours is a fixed monitoring panel of brand and category prompts repeated daily, theirs is drawn from platform interfaces. Surface, prompt wording, and prompt mix are all candidate causes of the gap, and our data does not separate them.

The prompt set is the part that can be tested directly. Running our own prompt panel through the ChatGPT interface on the same day as the API run would hold the prompts constant and leave the surface as the only variable. We have not run that test.