ChatGPT’s New Search: Inside Its Updated Language
▼ Summary
– OpenAI replaced the JSON-based tool call format in ChatGPT with a compact, pipe-delimited query language between August 16 and August 20.
– The new format eliminates the previous metadata field called search_queries and introduces specific parameters like freshness windows and domain targeting.
– Each line in the new tool call represents a distinct search, utilizing fields for call type, query text, numerical constraints, and optional domains.
– A length directive (long, medium, short) was retained from the old JSON format to control the volume of text returned per search result.
– The author analyzed eight diverse questions to map how this new language influences source selection and content retrieval across different verticals.
On August 16, a web search initiated by ChatGPT utilized a standard JSON structure for its tool calls. By August 20, that same query on the same account triggered a completely different mechanism: a compact, pipe-delimited query language. This shift eliminates the previous search_queries metadata field and replaces it with a more directive syntax.
This evolution marks the fourth installment in an analysis of ChatGPT’s search behavior. After examining source selection and pre-search shortlisting, this investigation maps the new format across eight distinct question types, ranging from commercial products to local news and finance. The new language reveals a sophisticated system where every line represents a specific search intent, governed by readability, freshness windows, vertical-specific logic, and domain targeting.
Every Line Is A Search
The new tool call is structured as a block of lines, with each line representing a single search operation separated by pipes. The recurring pattern includes a call type, the query text, a numerical value, and optionally, a domain. For example:
fast|Zendesk AI agents pricing 2026|30|zendesk.com
Here, fast indicates a standard web search, serving as the successor to the older fan-out method. The query itself resembles previous formats, containing brand names, years, and intent keywords like “pricing.” However, the number and domain are new parameters. Each block concludes with a length directive, such as length|long, which instructs the model on how much text to retrieve per result. This mirrors the old “response_length” parameter but is now integrated directly into the search line. Values vary based on complexity: long for deep comparisons, medium for general lookups, and short for verification tasks. This suggests most searches pull bounded excerpts rather than entire pages, emphasizing the need for key sentences to appear early in content.
Practical AI SEO/GEO Tip Identify the questions your buyers ask and monitor the brands ChatGPT mentions in its queries. These brand names represent the competitive set the model already recognizes for your category. Ensure your site can withstand these targeted probes by verifying that your content aligns with the model’s expectations.
The 3rd Field Is A Freshness Window
The numerical field following the query appears to function as a recency window, measured in days, tailored to how quickly information becomes stale. Analysis of various queries reveals distinct patterns:
- Stock prices: 2 daysWhile OpenAI has not officially specified this field, the correlation between data volatility and the window size is strong. For instance, a stock price requires immediate freshness, while commercial comparisons allow for a month-long buffer.Practical AI SEO/GEO Tip If this assumption holds true, pricing or comparison pages that have not been updated in over a month may fall outside the default freshness window for the very queries where they are most likely to be compared. Update buying-focused pages regularly with visible dates and substantive changes to remain within these critical cycles.
Search Is A Set Of Verticals
ChatGPT employs different call types depending on the nature of the query, effectively segmenting searches into verticals.
Product Lookups For physical goods like robot vacuums, the model uses a product call type, listing semicolon-separated item names. This aligns with the product cards and merchant offers seen in commerce layers. Presence in this catalogue is distinct from organic blog visibility.
Practical AI SEO/GEO Tip If you sell physical products, verify if your items appear with cards and merchant names when ChatGPT makes recommendations. A card indicates catalogue recognition, whereas a text-only mention means you are competing solely on content optimization.
Business and Places Local queries trigger a business lookup that takes a location. For example, searching for coffee in Dubai involves an initial call with search phrases, followed by secondary calls that verify specific business names using quoted fast searches. This mirrors the shortlist mechanics used for brands but applies them to local entities.
Practical AI SEO/GEO Tip For local businesses, the primary unit of optimization is your entity in the places data. Ask ChatGPT about your category in your area and observe the second business call. If your venue name appears, the lookup knows you, and subsequent searches verify your details. Ensure your website matches these listings exactly. If your name does not appear, the issue lies with the places data, which a website cannot fix.
Images and Widgets Image searches use the image call type, often running alongside places lookups. Additionally, some answers are generated via genui_run calls, which request widgets like stock charts or schedules using JSON arguments rather than traditional searches.
Practical AI SEO/GEO Tip Monitor your queries for genui_run lines. These indicate that the answer is a component rendered by OpenAI, such as a map or chart. In these cases, there are no citations to win, and traditional GEO tactics may not apply.
The Domain Slot
Previously, targeting a specific website required adding the `site:` operator to the query string. The new format includes a dedicated domain slot at the end of the line. This allows the model to scope a search to a known domain without altering the query text.
For example, a probe might target `zendesk.com` directly. This step is now standard; the model fills the slot from its memory of the brand’s domain. However, if the model’s memory is outdated, it may search the wrong place. In one observed case, a probe targeted `profound.ai` instead of the correct `tryprofound.com`, resulting in zero pages from the pool. Fortunately, a broader discovery search rescued the brand by finding the correct domain.
Practical AI SEO/GEO Tip Review your own capture data to ensure the domain in the slot is one you actually control. Rebrands, migrations, and country-specific domains often cause mismatches because the model relies on memory. A wasted probe due to an incorrect domain costs you a search opportunity, so verify your presence regularly.
Some Answers Are Becoming Widgets
The genui_run calls do not return web pages; they return interfaces. In financial answers, stock charts appear as references to widgets hosted on OpenAI’s CDN. Similarly, local answers may include map widgets wired to businesses identified in the places lookup.
This shift means that for sports, stocks, and local venues, part of the answer is a component rendered by OpenAI. There are no links inside these widgets to optimize, and no citations are attached to them. Entire categories of queries are moving to surfaces where the citation game no longer exists.
Practical AI SEO/GEO Tip Run your target queries and identify which ones return widgets. If your traffic depends on being the page that answers a specific question, and that answer is now a widget, you must adapt. Stop counting citation opportunities for answer types that no longer offer them.
The Reddit Situation Is Very Very Interesting
A particularly notable anomaly appeared in a query about AI live chat support. The search line included six brand names, pointed to reddit.com, and used a 365-day freshness window:
fast|best AI live chat support Intercom Gorgias Zendesk Ada Tidio Crisp reddit|365|reddit.com
This search did not ask Reddit who the best tools were. Instead, it asked Reddit what people said about the six tools the model had already selected from memory. The candidate set came from internal knowledge, and Reddit was fetched to gather opinion on those candidates.
The results showed that 84 of the 221 entries in the retrieval pool were Reddit threads, more than any other source. However, none of these 84 threads were credited in the final answer. All citations were bound to vendor pages.
In another instance, a local coffee query used a 3,650-day window for Reddit, accessing a decade of threads, while news sites received a seven-day window. Although Promptwatch reported a collapse in Reddit citations from 3.83% to under 1% in mid-August, the wire data suggests a more nuanced change: Reddit is being used as an invisible input for verdicts, while credits go elsewhere.
Practical AI SEO/GEO Tip If Reddit has been part of your AI visibility strategy, audit your campaigns. Check if the threads you invested in still receive citations. While measurable citation payoff has collapsed, engaging with Reddit communities and building brand trust remains valuable. Years of threads feed the training data that builds the model’s memory, and human readers continue to influence buyers. However, spamming Reddit for manipulation is no longer effective, as the data is used upstream for evaluation without direct attribution.
What Breaks, And What I’d Do
The old method of checking DevTools for the `queries` field is obsolete. That field no longer exists. The searches are now embedded in the `web.run` tool-call messages as pipe-delimited lines. Tools like FanoutFox will need updates to parse this new format.
To manually check your visibility:
- Open ChatGPT in Chrome with DevTools on the Network tab.
- Ask a commercial question relevant to your business.
- Filter for `conversation` payloads and select the largest response.
- Search the response body for `fast|`.
- Analyze the lines for your brand presence, domain slot accuracy, and freshness window.
This manual process is tedious. Extensions like FanoutFox can automate the fetching and citing analysis. Keep dated snapshots of your traffic, as these formats can change rapidly. What was stable on one day may be altered by the next.
(Source: Search Engine Journal)
