ChatGPT Fan-outs: What the Freshness Window Does to Your Page

When ChatGPT searches the web for you, it first writes down what it wants to search for. These notes can be read along and for a few weeks they have looked different. Instead of a JSON structure with a query inside, there is now one line per search with four fields. One of them is a number and two people have taken a closer look at this number: Suganthan Mohanadasan described it on 21.08.2026 and David Konitzny of Peec AI measured on 01.09.2026 what it does.

In my news radar of 04.09.2026 this was the item with the fewest sources in the entire run: Konitzny’s analysis and one post at SEO Südwest. Everything else that day was already everywhere. That is why I am writing about this topic and not about the model update that everyone had that day.

In the analyses the number is called a freshness window and the name itself is already an interpretation, because OpenAI has explained it nowhere. What exists is network traffic that two people read independently of each other. Plus a dataset of around 250,000 chats against which the interpretation could be tested. It holds up with one limitation, which is the most important one for your work: The window is a wish addressed to the search system and not a boundary.

This post walks through the syntax, shows the numbers with how they were collected and ends with the question of what a page can do to fall inside such a window. Google’s variant of the same mechanism is the query fan-out in AI Overviews and AI Mode, I have separate posts on that and they are linked here. This one is only about ChatGPT.

Suganthan Mohanadasan has been reading along for months what ChatGPT sends to its tool when searching. On 16.08.2026 a search call still looked like this for him:

{"system1_search_query":[{"q":"site:intercom.com Fin AI Agent pricing 2026"}]}

On 20.08.2026 he asked the same question in the same account and the call now looked like this:

fast|Intercom Fin AI agent pricing 2026 live chat support|30|intercom.com
fast|Gorgias AI Agent pricing 2026 customer support|30|gorgias.com
fast|Zendesk AI agents pricing 2026|30|zendesk.com
fast|best AI live chat support Intercom Gorgias Zendesk Ada Tidio Crisp reddit|365|reddit.com
length|long

His summary: “Somewhere in those 4 days, OpenAI replaced the JSON with a compact query language of its own.” The change therefore sits between those two captures and not on any announced date. One line is one search and the fields are separated by vertical bars: the call type first, then the query, then a number. If ChatGPT means a specific website, the domain follows. Everything after the query is optional, in Suganthan’s words: “The fields after the query are optional.” The block ends with a line length|long, length|medium or length|short. Suganthan reads it as an instruction, in his words, “for how much text to bring back per result”.

Two things about this are the real news for SEOs. First, the number and the domain slot had no place in the old format, they are new and not merely renamed. Second, a field that used to be there is missing: search_queries. Tools such as FanoutFox read the fan-out queries from this metadata. In Suganthan’s words the field is “gone from the payload entirely”. Anyone observing fan-outs through a tool should check whether the tool still sees anything at all since mid-August.

Two things in this own capture go beyond Suganthan’s description. First, neither line carries a number. There is fast, a vertical bar, the query, and then nothing. That matches Suganthan’s sentence “The fields after the query are optional” and Konitzny’s measurement, according to which only about 37 percent of fast calls carry a freshness parameter at all. If you open your own capture and find no number, you have done nothing wrong. The line without a window is the normal case. Incidentally the capture shows something else that none of the English-language analyses could check: on a German question the search lines stay German. ChatGPT does not translate into English here, it phrases the terms in the language of the question.

Second, in this capture the queries also sit in a metadata field, here search_model_queries with a queries array. Suganthan’s finding explicitly concerns the field search_queries, and that one is absent here too. For tools this means the fan-outs have not disappeared, they sit somewhere else. Whether that field is new or was always alongside, I cannot say from one capture. Further down in the same response, by the way, sits resolved_model_slug with the value gpt-5-6. If you make your own capture, that is where you read off which model just answered you.

The call types from Suganthan’s eight questions: fast is the web search and the successor to the entire old fan-out. product queries a product catalogue, which showed up for robot vacuums and mattresses and not for software. business is a places lookup with a location. image is an image search. genui_run calls a widget such as a stock chart or a fixture list and returns no page but an interface. For SEOs this is the most uncomfortable call type, because a stock chart holds no source to win. Konitzny additionally sees slow in his data, more on that below.

Suganthan laid the number out across five questions from 20 August. The table is his, the examples are shortened.

Source: Suganthan Mohanadasan: “ChatGPT Rebuilt Its Search Tool” of 21.08.2026. Five lines from one account on one day. Not a measurement across many users.

His interpretation comes with its limit in the same paragraph: “I think this number is a recency window in days, matched to how fast the answer goes stale. I haven’t seen any spec for this, so this is my assumption.” A stock price from the day before yesterday is worthless, so two days. A product comparison holds for a month, a quarterly forecast for a quarter. The numbers match how quickly the respective answer goes stale. That is exactly what makes the interpretation plausible.

For pages that get compared, this has an uncomfortable side. On the commercial questions in Suganthan’s capture, every brand line got a 30-day window. If the interpretation is right, a pricing or comparison page competes from outside the window after two months without a change. And precisely on the queries where brands are set against each other. What “outside” means in practice is shown in the next section, because it does not mean “invisible”.

David Konitzny works at the AI visibility provider Peec AI and asked the question that was left open after Suganthan’s post: What does the number really do? His basis is a slice of around 250,000 chats from the company’s own tool, so a vendor analysis. The numbers are his measurement and the interpretations behind them are his reading. I separate the two here the way he does himself.

The distribution of call types. In Konitzny’s dataset, fast appears in almost 90 percent of all search lines. slow appears there in around 14 percent of cases. The two shares do not add up to 100 percent, because one call can contain both types. In his data a freshness parameter is attached to about 37 percent of fast calls and to around 95 percent of slow calls. When both types appear for the same topic, fast usually keeps the 30-day window while slow narrows to one to seven days. Konitzny reads this as two questions about the same topic: first the overall picture of the last weeks, then what is happening right now. Whether different indexes sit behind fast and slow or merely different settings, he does not know by his own account.

The window that occurs most often. 30 days is by far the most common value in Konitzny’s dataset. The second most common is not 7 or 90, but 3,650, meaning ten years. A separate section on that in a moment.

Preference or filter. This is the part that counts for your pages. Konitzny held the publication date that ChatGPT supplies for each retrieved URL against the requested window. If the window were a hard filter, the results would have to fall inside it almost completely. Instead he writes: “For a 30-day freshness request, around 65% of retrieved URLs fall within the requested 1–30 day period. The remaining URLs can be considerably older.” His conclusion: “The freshness value influences retrieval, but does not strictly constrain it.” Around a third of the results in his data came from outside the window.

For news questions the picture gets sharper, meaning prompts with words such as “latest” or “what happened today”. There, in Konitzny’s data, around 60 percent of the retrieved URLs came from the last seven days · 27 percent from the last 30 days · 12 percent from the last 24 hours. So for news ChatGPT does not only look for the newest, but for a week of context around the newest.

When transferring these numbers into my verification report, I first assigned the 27 percent to the 30-day request. That became “only 27 percent inside the window” and it was wrong: The 27 percent belong to the news questions. I noticed it on the second reading of the original and before a single line of it stood here. I mention it for one reason: It shows how easily such an analysis yields a number with the wrong reference as soon as it leaves the text in which it was collected.

The second most common window in Konitzny’s dataset is ten years and he first took it for an artefact. Then he looked at which domains are retrieved with this window: In this subset Reddit stands at around 60 percent of retrievals · G2 at about 20 · LinkedIn at 12 · Trustpilot and YouTube at around 10 percent each. He explicitly adds that these are not overall shares, but the distribution within the retrievals that use the broad window.

Suganthan had seen the same pattern in individual lines. In his live chat example the Reddit line got 365 days, while the brand lines in the same call all sat at 30. In the coffee question there was fast|site:reddit.com Dubai Marina specialty coffee Dubai|3650|reddit.com. Ten years of threads on a question for which news sites got seven days.

Both offer the same explanation and it is an interpretation: On these platforms a post does not lose its value with age. A Reddit thread from 2019 can contain the answer, a G2 review from three years ago still describes the product. The search system gets a wide window for such sources and a narrow one for news. Plausible and not confirmed by OpenAI. Konitzny explicitly frames his further conclusions on this as a supposition.

What remains for you: If your topic touches opinions and experiences, your page does not compete with last week’s Reddit thread, but with ten years of Reddit. That is not an argument against your own content. It is an argument for having the experience part where ChatGPT looks for it. According to this data that is the platform rather than the blog.

On 04.09.2026 I searched the documentation of the web_search tool in OpenAI’s Responses API for “fresh”, “recent” and “days”. Two controls are documented that match what is in the traffic: first filters with up to 100 allowed_domains or blocked_domains as a domain slot and second search_context_size with the levels low, medium and high, which according to the same docs controls how much search context the model receives. I did not find a parameter for a time window in this version. That is a statement about these docs on that day and not about OpenAI’s systems.

The docs describe the API and not ChatGPT. That the consumer interface works internally with a window that is not offered to developers would be no contradiction. It only means: Outside OpenAI there is no statement on how the window is applied. The two analyses here are therefore observations, robust observations with numbers and a method, but not a vendor statement.

Part of the context is that the syntax is only the latest of several rebuilds. Search Engine Land evaluated the model switch of 04.03.2026 in May 2026: The number of domains cited per response fell from 19 to 15, measured on 400 prompts daily over 14 weeks with data from Meteoria. Chris Long of Nectiv ran around 4,000 prompts on GPT-5.6 Sol in August and in his data an average of 7.61 fan-out queries per prompt came out, the year before it was 2.17 for him. In 64 percent of the queries at Nectiv there was a site:. Both are vendor and editorial measurements.

How solid such numbers are is shown by a counter-example from 25.08.2026. An analysis at dejan.ai covering 196,692 fan-out queries from the OpenAI API arrives at a different picture: the site: share there sits at 0.14 percent in August and the number of fan-outs per response fell over the measurement window from 6.16 in March to 1.10 in August. Against Nectiv’s 64 percent and 7.61 fan-outs that is not a difference in detail, but a different result.

The explanation lies in what was measured in each case. Both analyses run through the API rather than the ChatGPT interface, but across different models: Nectiv measures GPT-5.6 Sol, while dejan.ai carries its August traffic on GPT-5.5. By dejan.ai’s own breakdown, the older models 5.2 and 5.4 scope roughly one query in eighteen to a domain and 5.5 only one in 647. On top of that comes the type of prompt: short monitoring prompts produce a site: share of 0.44 percent there, longer research prompts one of 6.29 percent.

For you this means less than it sounds like, and something different. None of these numbers describes “ChatGPT”. Each describes one model with one sort of prompt in one period. Anyone reading a fan-out statistic should therefore ask first about the model, the access path and the type of prompt and only then about the percentage. The direction all four measurements agree on: the search query alone is no longer the whole instruction.

That these lines can be read off the network traffic was described by Suganthan Mohanadasan in his post. The walkthrough below is my own capture of 04.09.2026 including the places where it went wrong. The traps are the actual point here, because all three of them feel like “there is nothing to see”.

Step 1: sign in. My first attempt ran in a signed-out session. The network list then held only prepare?lightweight_authenticated=0 and updates?lightweight_authenticated=0, plus a few CSS and JS files that happen to carry “conversation” in their name. The answer arrived and it even had sources, the payload with the search lines was nowhere. Without a sign-in a different path runs.

Step 2: find the right row. Open DevTools with F12, Network tab, filter field conversation. Then send your question and wait until the answer stands complete. The list now holds many rows sharing that name fragment. You want the one with the large payload, not the files ending in .js and .css. The fastest route is a click on the Size column. In my case it was the request to /backend-api/f/conversation, visible in the field X-Openai-Target-Path.

Step 3: search in the right pane. This is the trap that cost me the most time. DevTools has its own search field on the left edge and that one searches all loaded files. For fast| it reliably reports “No search results”, because the lines do not sit in a file but in a response. The right way: click the row, tab Response, click into the pane, then Ctrl+F. If there is a tab EventStream instead, the answer is a stream and the events are listed there one by one.

Once you have landed on fast|, there is more around those lines. The fields I met in my own capture and what they tell you:

And now the part that makes the effort worthwhile. Four questions to put to your own capture:

• Are there brands in the search line that you did not name? Then ChatGPT held that candidate set before it searched for anything. That is your competitive set as the model sees it, in black and white.
• Does the last field hold your real domain? If an old one, a wrong one or none at all sits there, the targeted check of your brand runs into nothing.
• Is there a number at all? In my case there was not. Then your category runs without a recency preference, and the maintenance cycle is a question about your users rather than about this parameter.
• Which model answered? Note the slug together with the date. Next time you will see whether something changed or only the model did.

1. Make the date visible and the change real. Konitzny’s measurement compares first publication dates and not updates. Whether ChatGPT reads a “last updated” is open. What I would do anyway: keep the date in the visible text and in the structured data and change it only when the content has changed. A date change without a text change is the kind of signal a search system eventually ignores. That is my assessment and not a rule from any of the sources.

2. Know your 30-day pages. Prices and comparisons and feature overviews are the pages with the 30-day window in Suganthan’s capture. If you have such pages, their maintenance cycle belongs under one month. A comparison that has stood unchanged since spring competes, according to Konitzny’s data, from the last third of the results and not from invisibility.

3. Check the domain slot. ChatGPT fills the domain behind the search line from its own memory. Suganthan saw a search aimed at profound.ai, although the vendor runs on tryprofound.com. The search came back with zero pages. The brand survived only thanks to a broad search. After a rebrand or a migration or with country domains, that is exactly the spot that tips. Ask ChatGPT about your category and check which domain it enters for you.

4. The answer sentence up front. The length line suggests that ChatGPT fetches a limited excerpt per result and not the whole page. If that is right, the sentence that answers the question has to sit where an excerpt starts. So at the start of a paragraph or under a subheading and not after three sentences of lead-in. For Google’s query fan-out I described the same thing in the post on AI Mode. The foundation for it is in the GEO post.

5. Read Search Console. Lily Ray documented in August that site: queries from fan-outs show up in Search Console, one of them with around 197,000 impressions and exactly one click. You can check this with her filter directly in Search Console. In the performance report filter on queries, custom, regular expression, and enter (?i)(site:.*official|official.*site:) there. If your impression curve shows such outliers, that is machine traffic rather than demand. How to separate such patterns from data errors is in the post on the GSC impressions bug. What Google now delivers in AI data of its own is in the post on the GenAI reports.

6. If you sell products, you are playing on two fields. The call type product queries no text but a product catalogue. Suganthan saw it on robot vacuums and mattresses and on not a single software question. His conclusion, and I share it: anyone selling physical goods additionally has to appear in that catalogue, and the best product page changes nothing about that. The test is simple. Ask ChatGPT about your product category and check whether your items come back as a card with a price and a merchant name or only as a mention in the text. Only the card means the catalogue knows you.

What counts for me about the new syntax is not the syntax. It is the fact that the search query is now only one part of the instruction. ChatGPT now tells its search system which kind of retrieval it wants · which time window · which domain. The question “how do I show up in ChatGPT” thus has three new variables and none of them is in any documentation.

The data that exists is enough for a working rule and too thin for a promise. A window that holds two thirds of the time is a preference. For a pricing page that has been unchanged for months it is nevertheless the difference between the first and the last third of the results. That is reason enough to know the maintenance cycle of these pages. It is no reason to bump dates without a change.

Sources

What is linked in the text is listed here once more with date and role.

• Primary analyses: Suganthan Mohanadasan: ChatGPT Rebuilt Its Search Tool. I Read the New Language It Speaks. of 21.08.2026 (syntax · window · verticals · domain slot · own capture from one Plus account) · David Konitzny (Peec AI): Inside ChatGPT’s New Search Language: How Fresh Does Content Need to Be? of 01.09.2026 (vendor analysis · around 250,000 chats)
• OpenAI: Web search · Responses API · retrieved on 04.09.2026 (documented parameters of the search tool · vendor documentation)
• Search Engine Journal: Republication of Suganthan’s analysis of 25.08.2026
• SEO Südwest: ChatGPT sucht jetzt anders: zeilenbasiert und mit Freshness Windows of 23.08.2026 · Wie die neuen Freshness Windows bei Fanout Queries funktionieren of 02.09.2026 (German-language coverage)
• Search Engine Land: Inside ChatGPT Search: how web.run and fan-out queries shape AI visibility of 14.05.2026 (domains cited per response after the model switch · data from Meteoria)
• Nectiv: ChatGPT Tripled Its Fan-Out Queries of 13.08.2026 (vendor analysis · around 4,000 prompts)
• dejan.ai: Our OpenAI fanout data shows no rise in site: operator use of 25.08.2026 (counter-measurement across 196,692 fan-out queries from the OpenAI API)
• Lily Ray: What We Can Learn from Evolving ChatGPT Fan-Out Queries of 17.08.2026 (site: queries in Search Console)