Fan-out: how ChatGPT searches before it answers

Key numbers since measurement start

Measuring days

23

Runs with search

445

Total queries

1748

Queries per run

3.9

with freshness window

67 %

most common window

other

with site:

7 %

with domain filter

39 %

with location scope

18 %

days with / without call text

23 / 0

Timeline

The fan-out events

• retractedxFan-out: new filter domain apple.com (never seen before)1/33 prompts · instant1/33 · ievidence: measuring day 11 Sept 2026
• retractedxFan-out: new filter domain asics.com (never seen before)1/33 prompts · instant1/33 · ievidence: measuring day 11 Sept 2026
• retractedxFan-out: new filter domain brooksrunning.com (never seen before)1/33 prompts · instant1/33 · ievidence: measuring day 11 Sept 2026
• retractedxFan-out: new filter domain idealo.de (never seen before)1/33 prompts · instant1/33 · ievidence: measuring day 11 Sept 2026
• retractedxFan-out: new filter domain nike.com (never seen before)1/33 prompts · instant1/33 · ievidence: measuring day 11 Sept 2026
• retractedxFan-out: new filter domain runnersworld.com (never seen before)1/33 prompts · instant1/33 · ievidence: measuring day 11 Sept 2026
• retractedxFan-out: new filter domain samsung.com (never seen before)1/33 prompts · instant1/33 · ievidence: measuring day 11 Sept 2026
• retractedxFan-out: new filter domain zdfheute.de (never seen before)1/33 prompts · instant1/33 · ievidence: measuring day 11 Sept 2026

Fan-out profile per prompt category

Across all measuring days, searching runs only; categories in prompt-set order. Prompt echoes and actions are not counted.

Seeds → results → cited

Per category and run on average: queries → result domains → loaded via open → citations in the answer. Plus the share of reddit, wikipedia and YouTube among result domains.

Domain filters and site: targets over time

Domains ChatGPT queries deliberately — via |domain filter or the site: operator. First and last sighting, number of measuring days.

Reading the events

Twenty-three measuring days, 445 searching runs, 1,748 queries. Before answering, ChatGPT issues 3.9 queries on average, two thirds of them with a freshness window, one in five with a location scope, one in fourteen with a site: operator. That is normal operation, and it was remarkably quiet for three weeks. The modes business, fast, search, image, slow and product appeared every single day, the most common window was 30 days, and queries per run drifted from 4.4 down to about 3.7 and stayed there.

Against that quiet background four things stand out. From 26 to 31 August, and again on 5 and 6 September, two modes appeared that never showed up otherwise: genui_run and genui_search. They coincide exactly with the rollout of map widgets and location prompts that the tracker reported as stream events on the same days. ChatGPT adapted its fan-out to the new surface during that phase and rolled it back.

The site: operator is the noisiest signal in the series. On 23 August there were 14 site: queries, on 8 and 9 September none, on 11 September three again. Anyone concluding from a week of observation that ChatGPT dropped or ramped up site: is reading noise. The operator shows up for price checks, opening hours and news, and whether it appears on a given day depends on which question triggers which lookup round.

The freshness window sat steadily at 30 days, with three outliers downward: a median of 18.5 days on 1 and 3 September, 7 days on 11 September. That last value comes from the morning run of the day the call texts disappeared; whether it was a precursor or coincidence can no longer be checked.

Because on 11 September, between 09:15 and 15:20 CEST, OpenAI removed the web.run call texts from everything that reaches the browser. In the live stream, in both retrieval endpoints, retroactively for weeks-old conversations, on two accounts. What remains are bare query strings without type, window, location or filter. The morning run of 11 September is the last complete capture; a probe run at 17:42 the first without. Everything this page says about freshness, modes and site: is therefore a closed dataset, 20 August to 11 September 2026. On the same day Metehan Yeşilyurt had published his piece on chunk ranking. Whether the two are connected we do not know. Both are dated.

What this means for site owners

The fan-out decides which pages are candidates at all. Everything after it, the opening, the citing, the map, the carousel, works with that candidate set. Three observations from the 23 days are directly usable for your own site.

First, the window. Two thirds of the queries carry a freshness window, mostly 30 days. A page whose last visible date is older than that does not lose on quality, it never enters the filter. A dated, updated article is not a ranking factor, it is the ticket.

Second, the lookup round. For local and product questions ChatGPT runs a second round searching by name: five restaurants, four shoe models, with site: queries against the business's own domain for opening hours or prices. Whoever is not found in that round, because the domain does not match the name or the page lacks the information asked for, drops out of the map even though round one had them. The tracker holds cases where ChatGPT read nine pages of one domain and then did not show the business.

Third, the domain filters. In 39 percent of the runs ChatGPT restricts a query to specific domains from the start: reuters.com, tagesschau.de, idealo.de, nike.com, openai.com. That list is short and changes slowly. If you are not on it you compete in open search; if you are, you get queried directly for matching questions. The table above shows which domains have been on it since when.

Since 11 September none of this can be observed from outside. That does not change how ChatGPT searches. It changes who can see it.

Limits of the measurement

One account, 26 prompts, one run per day, German questions from 14 categories. This is a probe, not a sample. The fan-out of a single prompt varies day to day, which is why the timeline shows shares and medians across all searching runs, and why an event is only reported once it affects 70 or 90 percent of runs. What happens to a single question lives in the drilldown, not in the changelog.

The measuring account is a Plus account located in Munich. ChatGPT passes that location into business queries, and the distances in local answers are computed from it. A different location, plan or A/B branch may look different. On 11 September the measuring account and a second Plus account were affected identically; that argues against a small test branch but proves nothing about the whole population.

The mode and freshness values come from the text of the web.run calls as ChatGPT delivered it until 11 September. Whether the server used those values exactly as written we cannot verify; we document what the client received. The mode search is the echo of the user's question in the function form search("…") and, since schema 17, does not count as a fan-out query. Language attribution is a heuristic. The published snapshots up to 11 September were rebuilt from the raw archive with the current extractor; the events of previous days were left untouched.

And the most obvious limit: this page describes a state that has not been measurable since 11 September. Should OpenAI deliver the call texts again, the tracker will report it as an event and the series will continue.