Hi everyone, this is Neo.

Quick question: do you know what happens inside ChatGPT before it answers your question? How it actually “searches”?

Until last Saturday (August 16), ChatGPT’s search tool call looked like this — a JSON blob:

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

Four days later, on August 20, I asked the same question on the same account. Here’s what it looked like:

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

Note: I didn’t format this by hand. This is what ChatGPT generated on its own. In four days, OpenAI replaced the fan-out JSON that powered ChatGPT search with a compact pipe-delimited query language.

Why should ecommerce and independent site owners care? Because the entire logic of how ChatGPT picks sources, fetches pages, and cites answers just changed. In this post, I’m combining the latest teardown by technical blogger Suganthan Mohanadasan (part four of his series) with Search Engine Land’s reverse-engineering research, to walk through the five big findings of the new format — and what to do about each one.

1. The background: how ChatGPT search “fans out”

Before we dig into the new format, let’s recap how ChatGPT search works.

When you ask ChatGPT a question, it doesn’t search once. It breaks your question into multiple sub-queries (anywhere from 2 to more than 10) and fetches from different sources for each one — a mechanism the industry calls fan-out.

Search Engine Land’s deep-dive from May spelled this out:

• After OpenAI switched the default model to GPT-5.3 Instant on March 4, the average number of unique domains cited per response dropped from 19 to 15, and unique URLs from 24 to 19 — a decline of more than 20% that never recovered. They named it the “Bigfoot Effect,” after the 2012 Google update where a few domains came to dominate entire result pages.
• ChatGPT now has 900 million weekly active users. The citation pie didn’t grow, but fewer websites are sharing it — sites that don’t make the cut are losing visibility that used to be within reach.
• The underlying search tool is called web.run, and it now supports 12 operations (up from 4), including search_query, open, find, click, screenshot, product_query, plus a separate widget system (genui) for sports, finance, weather, and more.
• GPT-5.4 can chain 5 to more than 10 rounds of search per response, refining queries as it goes; 5.3 Instant typically runs 2 or 3.
• The crawler that actually fetches your page content is ChatGPT-User, not OAI-SearchBot — the latter is mainly for building the index.

That’s the context: ChatGPT’s search machinery has been evolving fast, and this late-August format rewrite is a major change to how it picks sources.

2. What the new format looks like — every field matters

The new query language’s general structure is:

call_type|query|number|domain

For example:

fast|Zendesk AI agents pricing 2026|30|zendesk.com

• fast: the call type, corresponding to web search
• Zendesk AI agents pricing 2026: the query itself
• 30: the freshness window in days
• zendesk.com: the target domain slot

Each search block also ends with a directive line like length|long / length|medium / length|short — an instruction for how much text to bring back per result. The author found long fired on a deep product comparison, medium on most lookups, and short on a final call verifying two venue names.

Here’s a big implication: most searches pull a bounded excerpt of your page, not the whole page. The sentence that answers the question needs to sit where an excerpt starts. If your key answer is buried deep in the page where readers have to scroll to find it, AI may never read it at all.

3. Five findings — every one affects your traffic

Finding 1: Every line is a search, and the fields are readable

The biggest change: a tool call is now a block of lines, one search per line, fields separated by pipes. The query still contains brand names (written by the model itself), a year, and intent words like “pricing.”

What to do: ask the questions your buyers ask, then look at the brand names ChatGPT writes. Those brands are the competitive set ChatGPT already holds for your category; the domains in the slots are the sites it intends to visit. If you’re not in that set, you’ve already lost the first round.

Finding 2: The third field is a freshness window — and it changes with what you ask

Here’s the same field across five question types:

Question type	Freshness window (days)
Stock price	2
Football results	7
Commercial product research (e.g., AI support software pricing)	30
Earnings guidance	90

The author’s read: this number is a recency window in days, matched to how fast the answer goes stale. There’s no official spec yet, but on every commercial question he ran, every brand probe had a 30-day window.

What this means for site owners: if the days assumption holds, a pricing or comparison page that hasn’t changed in a month is competing from outside the default window on exactly the queries where brands get compared. If your product, pricing, and comparison pages haven’t been touched in six months, ChatGPT may not even be considering you.

Action: put the pages that answer buying questions on a rolling 30-day update cycle — real changes, visible dates. Not a date bump to game the system.

Finding 3: Search is now a set of verticals — selling physical goods? Pay attention

The author captured five call types so far:

• fast: web search, the successor to the whole old fan-out
• product: catalog lookup. Appeared on robot vacuums and mattresses, never on software questions. Semicolon-separated product names on the line: product|Roborock Saros 10R robot vacuum;Dreame X50 Ultra robot vacuum;Eufy E28 robot vacuum
• business: places lookup with a location. E.g., business|Dubai, UAE|specialty coffee Dubai Marina;best coffee Dubai Marina — the first call sends search phrases, the second sends actual business names to verify one by one
• image: image search, e.g., image|Dubai Marina specialty coffee shop|30
• genui_run: widget call with JSON arguments, e.g., stock charts and Premier League schedules

This is huge for physical goods sellers: when ChatGPT recommends products, it runs a separate “catalog” channel — it checks the catalog (product cards, offers, merchant names) and searches the web. Whether you’re in that catalog is a completely different game from how good your blog is.

How to check: ask ChatGPT about your category and see whether recommendations come back with product cards (price, rating, sellers). Cards mean the catalog knows you. A text mention with no card means you’re only in the “content game” — and in that case, no amount of page work changes what the product call returns.

Finding 4: The domain slot means ChatGPT visits sites it already knows

In the old days, when ChatGPT wanted to search one specific website, it wrote site:intercom.com in the query. In the new format, “website” is a dedicated slot right next to the query — and the model fills it from memory.

That introduces a new risk: if the domain in its memory is wrong or stale, the probe searches the wrong place.

The author watched this happen live: on the query “best LLM visibility tool and what does each one cost,” the model wrote site:profound.ai pricing AI visibility platform — but Profound doesn’t run on profound.ai; they run tryprofound.com. The probe returned zero pages. The brand survived anyway (the broad search found the right domain, and it got cited six times) — but the author explicitly says he wouldn’t build a strategy on that rescue repeating.

Action: capture your own logs and check whether the domain in the slot is one you actually run. The cases that produce a wrong slot: rebrands, migrations, and country-domain switches (e.g., moving from .com to .de or .co.uk). The model fills the slot from memory, and memory lags.

Finding 5: Some answers are becoming widgets — nothing to win in them

What genui_run means: for some questions, ChatGPT’s answer is a direct function call — a Premier League schedule, an NVDA stock chart — typed arguments, not a search at all.

This is a brutal signal for content creators: any answer type that can be widgetized (scores, stock prices, exchange rates, weather, flight times) may soon have no citations to win. No matter how good your “Premier League weekend fixtures” article is, it can’t beat a live component.

This isn’t fearmongering. Search Engine Land’s research confirmed the genui widget system exists — sports, finance, and weather are already in there. Part of the SEO battlefield will simply disappear.

4. Old vs. new format at a glance

Dimension	Old format (before Aug 16)	New format (from Aug 20)
Structure	Structured JSON	Compact pipe-delimited text lines
Site restriction	site: operator inside the query	Dedicated domain slot
Freshness control	No explicit parameter	Third numeric field (2–90 days)
Verticals	Single web search	fast / product / business / image / genui_run
Response length	response_length field	`length
Content fetching	Full pages	Mostly excerpts (depends on length)

5. Action checklist for site owners (Neo’s version)

1. 30-day content freshness: list every page that answers a buying decision — product pages, pricing pages, comparison pages — and put them on a rolling 30-day update cycle with real changes and visible dates.
2. Catalog self-check: ask ChatGPT about your category (include words like “best” and “top”) and see if your products come back with cards. No cards? Research how to get into its product data — for Shopify sellers, that means Merchant Center and product structured data.
3. Brand probing: ask questions in your customers’ words, capture ChatGPT’s calls (tools like FanoutFox can help), and check whether you’re in the brand set and whether the domain slot points to your site.
4. Domain slot calibration: if you’ve changed domains, rebranded, or switched to a country domain, verify what AI remembers. One wrong slot = one wasted probe.
5. Put the answer where the excerpt starts: since searches mostly pull excerpts, put each page’s core answer in the opening 50–100 words. Don’t bury it deep in the page.
6. Avoid the widget battlefield: if your core keywords are widgetizable (stock prices, scores, exchange rates), pivot early — move content toward analysis, interpretation, and comparison that widgets can’t replace.

Neo’s take

Honestly, this is the most valuable ChatGPT mechanism teardown I’ve seen in months — not because the format details are cool, but because it turns the black box of “how ChatGPT picks you” into readable instructions, line by line.

Three things matter most to me:

First, if the freshness window is real, it changes the economics of content updates. We’ve always said “keep content fresh” as a feeling. If ChatGPT really has a 30-day window, updating stops being a superstition — commercial pages untouched for over 30 days fall outside the default retrieval window. That turns “regularly update key commercial pages” from advice into a matter of survival. It also explains a puzzle: why some high-authority old pages never get cited by AI. Maybe it’s not that AI doesn’t trust you — your content is just sitting outside the window.

Second, the split between product catalog and web search is the most underrated shift of 2026. Many site owners are still grinding on “content + backlinks,” but ChatGPT’s recommendation path for physical goods doesn’t go through the web at all — it checks the catalog first. That means: whether your product data makes it into the catalog (structured data, Merchant Center inclusion, complete offers and ratings) matters more than 100 blog posts. Content players and catalog players are becoming two completely different races.

Third, the domain slot exposes a brutal truth about the AI era: a model’s memory of your brand can be wrong — in a very dumb way, by visiting the wrong domain. But here’s the upside: AI memory can be corrected, as long as you know it’s wrong. Proactive monitoring and calibration will become the new routine for AI-era independent sites.

One honest caveat: OpenAI rewrote the underlying format in four days, so ChatGPT search is far from settled. The specifics in this post may be outdated in a few months — but the directional trends won’t change anytime soon: the model writes a brand set before searching, freshness windows matter, physical goods go through catalogs, and answers are getting widgetized. Ride the direction, not the format.