In mid-August 2026, ChatGPT rebuilt how its search tool talks to the web. Between 16 and 20 August 2026, the model stopped issuing searches as JSON objects and switched to a compact, pipe-delimited query language — each background search is now one line: a call type, the query, a freshness window in days, and an optional target domain. This reference lays out the syntax field by field and what each part changes for getting cited.
Last reviewed: 23 August 2026. The format below was documented by search consultant Suganthan Mohanadasan (suganthan.com), who captured ChatGPT's browser traffic across a small set of questions on a single Plus account between those dates. It is a single-observer, directional teardown of a tool OpenAI changes without announcement — the structure is directly observed, but the exact values (freshness windows especially) come from a small sample. The direction is corroborated by Search Engine Land's reporting on web.run and by Promptwatch's fan-out tracking. Attribute "Suganthan, August 2026" and date any figure you reuse. For the underlying concept of one prompt becoming many searches, see how query fan-out works.
What does ChatGPT's new search syntax look like?
One line per search, fields separated by pipes. The change is easiest to see side by side — the same intent, before and after the switch:
Old (JSON, ~16 Aug 2026):
{"system1_search_query":[{"q":"site:intercom.com Fin AI Agent pricing 2026"}]}
New (pipe-delimited, ~20 Aug 2026):
fast|Intercom Fin AI agent pricing 2026 live chat support|30|intercom.com
Each line reads as an ordered set of fields, and a search sequence often closes with a separate length|long (or medium/short) directive that sets how much the model wants back:
| Field (position) | What it holds | Example |
|---|---|---|
| 1 — call type | Which search vertical to use | fast |
| 2 — query | The rewritten sub-query text | Intercom Fin AI agent pricing 2026 |
| 3 — freshness window | Recency limit, in days | 30 |
| 4 — domain (optional) | A single domain to target | intercom.com |
| closing directive | Desired answer length | length|long |
The takeaway: a fan-out sub-query is no longer just a keyword string. It is a structured retrieval instruction that already encodes what kind of search, how recent, and which domain — before it ever touches the open web. This is a return to form, not a brand-new idea: analysts note ChatGPT used a similar pipe format before GPT-5.3, moved to JSON, and has now moved back to a compact language.
What are the search verticals (call types)?
Five, each routed to a different index. The first field decides which one, and that decision determines whether a citation is even possible:
| Call type | What it searches | Citation opportunity |
|---|---|---|
fast |
General web search | Yes — standard web pages |
product |
Catalogue lookups for physical goods | Product cards / merchant offers |
business |
Places and local lookups (with location) | Local listings / business data |
image |
Image search | Image sources |
genui_run |
Generated UI widgets (charts, schedules) | None — OpenAI-hosted, no web page cited |
Source: Suganthan, August 2026. The practical read is that different query types are won on different surfaces. A commercial "best X" question flows through fast and rewards clean web pages; a "show me trail shoes under $120" flows through product and rewards a correct product feed with merchant offers; a "coffee near me" flows through business and rewards your places data, not your blog. Knowing which vertical your target queries hit tells you which asset to fix.
What is the domain slot, and how is it different from site:?
The domain slot is the fourth field, and it is the structured successor to the site: operator. Earlier in 2026, analysts measured ChatGPT increasingly appending site: to its fan-out sub-queries to pin a search to one domain — a behaviour we documented in what ChatGPT's query fan-out targets. The new format promotes that targeting from text-inside-the-query to its own dedicated field.
Why it matters: the tool now routinely pre-selects a domain before searching, and that domain reflects the competitive set already in the model's memory. If it targets your domain and your page hides the answer behind JavaScript, images, or vague marketing prose, the sub-search returns nothing citable and the model moves on. If it targets a competitor's domain and skips yours, you were never in the consideration set for that query. The domain the model writes is a free read on how it perceives your category.
What does the freshness field do?
It caps how old a page can be and still be preferred for that search. The third field is a number of days, and the observed values scale with how time-sensitive the topic is — roughly 2 days for stock prices, 30 days for commercial products, up to thousands of days for evergreen forum content. Because the values vary so much by query type, we cover them in their own piece: see how fresh your page needs to be for ChatGPT and the query freshness window definition. The short version: for commercial and comparison pages, the observed window is about a month, so a pricing page that has not been genuinely updated in a quarter can fall outside it.
Why do genui_run widget answers matter for citations?
Because they close the citation door entirely. A genui_run line returns a rendered interface component — a stock chart, a sports schedule, a weather card — built from OpenAI-hosted data, not assembled from web pages.
When a query resolves to a
genui_runwidget, there is no web page in the answer and no link to earn — the citation slot you were optimising for does not exist for that question.
For visibility measurement, this is a real trap. If you track a basket of target queries and some of them now return widgets, those slots are unwinnable and will drag your apparent citation rate down for no fixable reason. The fix is to identify the widget-returning queries and stop counting them as citation opportunities — measure the queries where a citation is actually possible.
What should brands do about the new search language?
Treat each field as a diagnostic, and fix the asset that matches the vertical:
- Read the domains ChatGPT writes. Ask your target questions, watch which domain slots the model fills, and confirm they include yours. A wrong or missing domain — common after rebrands, migrations, and country-domain splits — means wasted probes and a fixable visibility gap.
- Match the asset to the call type.
fastrewards clean, server-rendered web pages;productrewards a complete product feed;businessrewards accurate places data. Fix the one your queries actually hit. - Respect the freshness field. Keep commercial and comparison pages genuinely current, because the observed commercial window is about 30 days — see the freshness-window playbook.
- Discount the widgets. Flag
genui_runqueries and remove them from your winnable-citation count so your measurement stays honest. - Measure the retrieval, not just the answer. The format proves how much happens before a citation appears; watching only the final links misses where you are being dropped.
The durable point behind the volatile syntax: ChatGPT's search tool is a precise, structured retrieval system, and each field tells you something about how it sees your brand. Reading those signals across hundreds of real questions — which verticals your category triggers, which domains get pre-selected, how fresh your pages need to be — is exactly what Buffy Intel is built to measure over time. Questions: [email protected].