How ChatGPT search works: the language changed in 4 days

Between August 16 and 20, 2026, ChatGPT replaced the internal format of its search calls, moving from JSON to a pipe-delimited one-line language. Each line carries the call type, the actual search query, a number read as a freshness window, and a slot that pins the search to a specific domain. For anyone trying to get a brand into AI answers, observations this concrete about what the engine looks for, and under which conditions, are rare.

This article rests on a single source: Suganthan Mohanadasan's observation piece in Search Engine Journal^[1], cross-checked against the original analysis on his blog^[2]. The author read his own browser traffic across eight questions, on one ChatGPT Plus account in one location, so the sample is small, and he states plainly that reading the number as a freshness window in days is his assumption, with no spec to confirm it. We operate a product that tracks AI search visibility, so we have an interest in this topic and say so upfront.

What changed: JSON calls became a single pipe line

In the August 16 capture, ChatGPT's search call was a JSON structure. In the August 20 capture, the same role appeared as a pipe-delimited line, and the author places the switch somewhere in those four days^[1].

Before (observed August 16)
{"system1_search_query":[{"q":"site:intercom.com Fin AI Agent pricing 2026"}]}

After (observed August 20)
fast|Zendesk AI agents pricing 2026|30|zendesk.com
 call    query                      win  domain slot

The two fields after the query are optional, and the block ends with a directive line such as length|long. It looks like a cosmetic change, but as the sections below show, this one line encodes the conditions of the whole search behavior.

Five call types: some answers have no citation slot at all

Five call types were observed^[1].

The one with immediate practical weight is genui_run. Questions like stock prices or sports results come back as widgets instead of web pages, and those answers contain no citation slot an external site could win. For some question types, no amount of content optimization earns a place, which matches the direction we covered in how generative UI changes citations.

Freshness windows vary by question: pricing research gets 30 days

The third field of the pipe line varied with the nature of the question. The author reads the number as a freshness window in days, and flags in the original analysis^[2] that this is an assumption without an official spec.

Table 1. Freshness window estimates observed across eight questions. These are per-question observations, not official per-category rules^[1].

If the reading holds, the implication is large. If the engine pushes documents older than 30 days out of the window for commercial comparison questions, last year's well-made comparison page quietly drops out of this year's answer pool. That freshness matters for citation matched earlier observations; what is new here is seeing the size of the window per question type.

Domain slots: the engine searches sites it already remembers

The fourth field carries specific domains such as zendesk.com or investor.nvidia.com. The engine does not scan the open web from a blank slate; it recalls which site answers which question and runs a pinned search inside it^[1].

User question
   |
   v
Model recalls brands and domains from memory   <- brand position in trained knowledge (GEO)
   |
   v
fast|query|window|domain                        <- pinned search on that domain
   |
   v
Retrieval pool -> selection -> answer and citations   <- real-time citation contest (AEO)

Getting named in the domain slot is decided before any real-time search happens, inside the model's trained knowledge. That contest sits upstream of document quality at search time, and the call format itself confirms the structure: making the model remember a brand during training is the precondition for the real-time citation contest.

Entering the retrieval pool is not being cited: Reddit went 84 for 0

When the author checked one conversation's retrieval pool against its citations line by line, 84 of the 221 pool entries were Reddit threads, more than any other source, and zero were cited (SEJ article^[1]). The 84 entries were snippets, with no sign the threads were read in full. Over the same period, Promptwatch's chart, cited in the article, shows Reddit's share of ChatGPT citations collapsing from about 3.83% to under 1% in mid August.

Being retrieved and being cited are different stages. Judged by pool entry, Reddit is thriving; judged by citations, it collapsed. Measure the wrong stage and you reach the opposite conclusion. This is the same lesson as in our analysis of 70,000 Perplexity queries: user questions and the searches the engine actually runs are different things, and each stage needs its own measurement to see where you are losing.

Why a tracking tool went blank overnight

The format switch removed the old JSON field, and the tool that parsed it to display fan-out queries went blank on new conversations. FanoutFox, the free extension built by the article's own author, shows an empty card until its patch ships, as he discloses himself^[1].

Tracking query fan-out to set content priorities remains valid, but the internal format that tracking depends on can be replaced wholesale in four days, as it just was. No measurement setup can be assumed to stay correct; the measurement method itself needs continuous verification.

Limitations

• This is a browser-level observation from one account, one location and eight questions. Rollout may differ by model, plan and region.
• Reading the numeric field as a freshness window in days is the author's assumption. There is no official spec.
• The zero Reddit citations come from a single conversation checked line by line. Promptwatch's aggregate supports the direction, but its sampling design is not public.
• Internal formats change without notice; the JSON-to-pipe switch itself took four days. These observations date to August 2026, and anything quoted from them should carry that date.

What to do with this

First, shorten the update cycle on pages competing for commercial answers. If the 30-day reading holds, product, pricing and comparison pages need substantive updates at least monthly. Changing the date alone is not an update.

Second, measure retrieval and citation separately. Whether your site fails to enter the pool, or enters and loses the citation, calls for entirely different fixes.

Third, aim for the domain slot. Becoming the site the engine pins for a question is decided before the real-time contest, and building your brand's position inside the model's trained knowledge is the path there.

Fourth, do not treat measurement as a one-time setup. As this switch invalidated a tracking tool overnight, the internals of AI search change week by week. This is also why TRAIL Search is built around tracking prompt sets on a schedule: only a measurement that keeps running catches the switch itself in the data.