Original insights by Suganthan Mohanadasan
30-second rundown
Key learnings:
- ChatGPT’s search tool appears to set a freshness window: stock-price searches looked back two days, commercial comparisons 30, and some Reddit searches ten years. OpenAI has not confirmed whether this is a hard cutoff or which page date it uses.
- Freshness appears to follow how quickly an answer goes stale: buying pages may need frequent material updates, while useful community discussions can remain relevant for years.
- ChatGPT does more than send keywords: it can target one website, search product catalogues, find local businesses, or return widgets such as stock charts and football schedules. Widgets may answer the question without linking to a webpage, leaving no citation to win.
This week: Choose one commercial question customers ask. In Chrome, open Developer Tools, select Network, ask ChatGPT the question, then filter requests for conversation and open the largest result. In its Response tab, search for fast|. In a line such as fast|Zendesk pricing 2026|30|zendesk.com, the number after the query is the suspected freshness window in days. Record the brands, targeted domains, and windows.
Coming 3 months: Clearly display the original publication date and last materially updated date on every important commercial page. Use ChatGPT’s search lines to estimate how fresh each page needs to be, then create an automated workflow that alerts its owner before that freshness window expires.
Suganthan Mohanadasan watched ChatGPT swap its search machinery while the industry argued about the paint. Between 16 and 20 August 2026, the same question stopped producing familiar JSON instructions and began speaking a pipe-separated language containing a query, call type, number, and sometimes a domain.
fast|Zendesk AI agents pricing 2026|30|zendesk.comThat number appears to be a freshness window in days. The evidence comes from eight questions on one account in one location, so it is an observation rather than an OpenAI specification. Still, the pattern is too orderly to dismiss as desert static.
1. Different questions receive different clocks
A stock-price question received a two-day window. Premier League results got seven days, commercial software research got 30, earnings guidance got 90, and Reddit searches stretched to 365 days or, in one local example, 3,650 days.
The clock follows the answer’s shelf life. A price can rot before lunch, while an old discussion about a product problem may remain useful years later. ChatGPT appears to match retrieval recency to the evidence it needs, like a pit boss assigning expiry dates to every stack of chips.
For a company selling running shoes, this means keeping pages about current models, prices, stock, comparisons, and specifications genuinely current. Every commercial brand probe in Mohanadasan’s sample used a 30-day window.
Mohanadasan inferred that the values represent days because they matched the questions, but found no public specification confirming it. Treat the windows as observational evidence, not scripture smuggled from the server rack.
2. ChatGPT can choose the street before it starts driving
The new format contains a separate domain slot. ChatGPT can name a brand, choose the website it associates with that brand, and search that property directly instead of adding a visible site: operator to the query.
A stale domain association can waste the search. ChatGPT targeted profound.ai for Profound, although the company operates at tryprofound.com. A broader search rescued the brand, but relying on rescue is how companies end up sleeping beside a burning roulette table.
Search is also splitting into verticals. product looks in a catalogue, business handles places, image finds pictures, and fast searches the web. If a product never receives a product card, rewriting the blog may not help because catalogue and web search are separate routes.
The practical check is simple: use FanoutFox, Suganthan Mohanadasan’s free Chrome extension, to inspect the generated searches for an important customer question. Confirm that ChatGPT names the right brand, targets the correct domain, and reaches the page containing the current facts.
3. Some answers have no citation prize
The strangest line is genui_run, which can generate a stock chart, football schedule, or local map. These components return an interface rather than webpages, so the familiar GEO prize, the linked citation, may vanish.
Some answer types offer no link to win. If a question has moved into a widget, stop counting every appearance as an available citation opportunity. The game changed tables while the scorecard was ordering drinks.
Reddit exposes another trap. In one conversation, ChatGPT fetched 84 Reddit snippets while citing vendor pages. Those snippets may influence the verdict without receiving a citation, so analyse both the pages ChatGPT retrieved and the pages it ultimately cited.
That distinction should discipline the whole operation. Use FanoutFox to track whether a source was searched, retrieved, cited, and capable of sending a qualified visitor as separate events. A hidden input may shape the answer.
ChatGPT’s new language reveals a retrieval system making choices about time, source, format, and destination before the answer appears. Keep commercial facts current, make the correct domain unmistakable, and stop treating every prompt as one ordinary web search. The query is only the first instruction in the envelope. The rest decides whether your page enters the casino or dies in the parking lot.
