When ChatGPT recently ran a web search for me, the backend tool call was formatted as standard JSON. But just four days later, I asked the exact same question on the same account, and the payload was completely different. Somewhere in those four days, OpenAI deployed a massive ChatGPT search tool update that replaced the standard JSON with a highly compact, custom query language.
Furthermore, the search_queries metadata field—the field developers previously relied on to parse AI logic—is gone from the payload entirely.
This ChatGPT search tool update fundamentally alters how the AI retrieves data, ranks brands, and cites sources. By analyzing the raw browser traffic from my ChatGPT Plus account, I mapped out exactly what this new query language is saying. Here are five shocking insights you need to know to adapt your AI SEO strategy.
1. Every Line Is a Highly Specific Search
Under the new ChatGPT search tool update, a tool call is now a block of lines, one search per line, with fields separated by pipes (|).
The repeating pattern is: Call Type | Query | Number | Domain. For example: fast|Zendesk AI agents pricing 2026|30|zendesk.com
fastis the call type.- The query reads like standard search text, often including brand names the model wrote in itself.
- The number and the domain are brand new parameters that the old JSON format had no place for.
Each block closes with a directive line, such as length|long or length|short. This appears to be an instruction for how much text to extract per result. If this reading is correct, most searches pull a bounded excerpt of your page rather than reading the entire page. Therefore, the exact sentence that answers the user’s question must sit right where the excerpt starts.
2. The 3rd Field Is a Dynamic Freshness Window
When reviewing the raw output of the ChatGPT search tool update, I noticed the number following the query changes drastically based on the subject matter:
- A stock price search got a 2.
- Football results got a 7.
- Commercial product research got a 30.
- Earnings guidance got a 90.
This number is a recency window calculated in days, matched directly to how fast the answer goes stale. For commercial questions, every brand probe had a 30-day window. If your pricing or product comparison page hasn’t been updated in a month, you are competing from outside the default window. To win in this environment, you must update the pages that answer buying questions inside that 30-day cycle with real, visible dates.
3. Search Is Now Segmented by Verticals
The ChatGPT search tool update reveals that ChatGPT no longer executes a single type of web search. Instead, it segments searches into distinct verticals:
- Fast: The standard web search.
- Product: A direct catalog lookup. For physical goods (like robot vacuums), ChatGPT looks up products in a specific catalog rather than searching standard web pages. If your product doesn’t return a product card, the catalog doesn’t know you exist.
- Business: A local places lookup that takes a location. It sends a search phrase, returns specific business names, and then verifies them one by one. If you run a local business, the unit of optimization is your entity in Google’s places data, not just your website.
Just as small businesses are adapting to workflow automation via platforms like Wix Symphony AI, they must also adapt to how generative search engines query local data entities.
4. The Domain Slot Relies on AI Memory
Until this ChatGPT search tool update, when ChatGPT wanted to search a specific website, it wrote the classic site:intercom.com in front of the query. Now, it has a dedicated slot for it.
The last field on a line is a target website, and that search only looks there. The AI model fills this slot itself based on whatever domain it already links to the brand in its memory. If the domain in its memory is wrong or stale (such as a recent rebrand or migration), the probe searches the wrong place.
Make sure you read the slot on your own captures to verify the domain is one you actually run. If the AI is looking at the wrong URL, you are wasting valuable search probes. For more details on technical domain migrations and search logic, review the Google Search Central Blog.
5. Answers Are Becoming UI Widgets (With No Citations)
Perhaps the most shocking revelation of the ChatGPT search tool update is the genui_run command. These lines do not return web pages; they return interface widgets.
For queries regarding sports schedules, stock charts, or local maps, the answer is a component that OpenAI renders itself. A stock chart widget has no link inside it for you to “win.” Whole classes of queries are moving to interactive surfaces where the citation game—the exact metric every GEO (Generative Engine Optimization) tool measures—simply does not exist. Stop counting citation opportunities on answer types that no longer offer any.
The Reddit Situation: Fetching Without Crediting
I want to walk you through how the ChatGPT search tool update handles Reddit, because it is fascinating.
In my capture for live chat software, the query looked like this: fast|best AI live chat support Intercom Gorgias Zendesk Ada Tidio Crisp reddit|365|reddit.com
This search did not ask Reddit who the good live chat tools are. It asked Reddit what people say about the six tools the model had already picked from its memory. Even wilder, on a local coffee shop query, it set the window to 3650 days—a full decade of Reddit threads!
However, while 84 of the 221 entries in the retrieval pool were Reddit threads, exactly zero of them got credited in the final answer. OpenAI is fetching opinions about its shortlist, using them to validate the answer, and then attributing the opinions to nobody. This indicates that Reddit has become an invisible input.
This move might be a deliberate anti-manipulation tactic. By expanding the query window to a full decade and stripping the visible citations, OpenAI renders recent Reddit spam campaigns virtually useless. If your strategy was to spam Reddit to manipulate AI visibility, this ChatGPT search tool update is terrible news. For official insights into data licensing deals, you can track announcements via the OpenAI Press Page.
Conclusion: Adapting to the New Payload
The old JSON recipe is dead. The queries have moved into the web.run tool-call messages as pipe-delimited lines. To check your own performance under the new ChatGPT search tool update, you must open DevTools on the Network tab, filter for conversation payloads, and manually search for fast|.
If you do this kind of work regularly, keep dated snapshots of your traffic and diff them after every capture session. The AI SEO landscape is shifting underneath our feet, and the marketers who can read the raw data will be the ones who survive.
Frequently Asked Questions
What was the recent ChatGPT search tool update?
The ChatGPT search tool update replaced the standard JSON payload with a compact, pipe-delimited query language. This new format introduces targeted freshness windows, distinct vertical searches (like products and local businesses), and explicit domain slots for brand probes.
Why did ChatGPT stop citing Reddit in its answers?
According to raw traffic analysis, the ChatGPT search tool update still pulls heavily from Reddit to validate opinions, but it has stopped providing direct clickable citations in the final answer. This may be a tactic to deter SEO manipulation and spam on the Reddit platform.
How can I check how ChatGPT searches for my brand?
You can monitor this by opening Chrome DevTools, asking a commercial question related to your brand in ChatGPT, and analyzing the Network tab. Look for the conversation payload and search for fast| to see the exact queries, freshness windows, and domains the AI is utilizing.
