ChatGPT’s New Search Language Changes GEO Strategy

TL;DR: OpenAI silently replaced ChatGPT’s JSON search format with a pipe-delimited query language in August 2026. The new format reveals freshness windows tied to query type, vertical-specific call types for products and places, and direct domain probing from the model’s own memory. Operators who don’t align their content and listings to these mechanics will be invisible in AI-generated answers — regardless of how well they rank on Google.

The Format Switched in Four Days

On August 16, 2026, ChatGPT’s web search tool called out in standard JSON: {"system1_search_query":[{"q":"site:intercom.com Fin AI Agent pricing 2026"}]}. By August 20, the same query on the same account produced a block of pipe-separated lines:

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

Each line is a search. The fields are: call type, query, a number, and an optional domain. The block closes with a length directive — long, medium, or short — which instructs the tool on how much text to retrieve per result. That directive existed in the old JSON as "response_length", so it was renamed, not invented. Everything else is new structure.

The implications for operators investing in AI search visibility are significant. A full marketing audit of your current GEO posture is a reasonable starting point before adjusting anything else, because the mechanics exposed by this format reveal multiple distinct failure modes — and fixing the wrong one wastes budget.

The Third Field Is a Freshness Window

The number sitting between the query and the domain is not arbitrary. Mapped across five query types, a clear pattern appears:

• Stock price: 2 days
• Sports results: 7 days
• Commercial product research: 30 days
• Earnings guidance: 90 days
• Reddit community signal: 365 days

This is a recency window in days — how far back the model is willing to look for a fresh answer. The number matches how fast the answer goes stale. A pricing page that hasn’t been touched in six weeks is competing from outside the 30-day window on exactly the queries where brands get compared against each other.

For operators in high-CAC verticals — forex brokers, crypto exchanges, iGaming platforms, personal injury law firms — buying-intent queries are where CAC is won or lost. If a competitor’s pricing or comparison page was updated 12 days ago and yours was updated 45 days ago, the model’s freshness window already disqualifies you before any content quality question gets asked. Update pages that answer buying questions inside a 30-day cycle, with visible dates and real changes. A thin date-stamp with no content change likely won’t satisfy the retrieval criteria.

Five Call Types, Each a Different Game

The new format exposes that ChatGPT’s search is not one tool — it is a set of vertical-specific lookups. Understanding which call type fires for your category determines what you can actually optimize.

fast is the standard web search, successor to the old fan-out. Most content optimization targets this call type. product fires on physical goods and returns catalogue cards with merchant names and offers — it does not fire on software or services. If you sell physical products and your item doesn’t return a card when ChatGPT recommends it, no amount of blog content changes what the product call returns. Catalogue presence is a separate game. business is a places lookup with a two-pass structure: the first call sends search phrases, the second sends specific business names retrieved from the first call and verifies them against the web. The unit of optimization for local operators is entity data in places listings, not the website itself. If your business name never appears in the second business call, your problem is listings — the website is irrelevant to that failure. image and genui_run round out the set; genui_run fires widget functions like stock charts and sports schedules as typed JSON calls, returning structured data with no web citation at all.

Operators running performance ad campaigns in verticals adjacent to these widget categories — financial products, sports betting — should note that genui_run answers carry no citation opportunity. The answer exists as a rendered widget, and no page wins attribution for it.

The Domain Slot and Memory Probing

In the old format, targeting a specific site required a site: operator inside the query string. The new format has a dedicated domain slot on every search line. Checking a brand’s own website is now a standard step in the tool, not a special-case operator. The model fills the domain slot itself, from whatever domain its training associates with that brand name.

That creates a specific failure mode. If ChatGPT’s memory links a brand name to the wrong domain, the probe searches the wrong place and returns zero results. One documented example: a model probed profound.ai for a tool that actually runs on tryprofound.com. The probe returned nothing. A broad fast search recovered the brand, but that recovery is not guaranteed. If your brand has changed domains, been through a rebrand, or operates on a non-obvious domain pattern, you need to verify what domain ChatGPT associates with your name — and whether that domain returns your current content.

For forex acquisition campaigns, this matters directly. Broker brands that have rebranded, merged, or migrated infrastructure in the past 18 months are candidates for domain-slot mismatches. The model’s memory of the old domain persists independently of any redirect or canonical tag you’ve implemented on the technical SEO side.

The same issue surfaces for iGaming operators running under holding company structures with multiple brand domains. ChatGPT may probe the holding company’s domain when a user asks about a specific product brand, returning generic corporate content rather than the conversion-optimized landing page for that product.

What This Means for High-CAC Vertical Operators

The operators most exposed to these mechanics are the ones spending $10K or more per month on paid acquisition in verticals where the buying cycle involves research: forex brokers, crypto exchanges, iGaming platforms, personal injury law firms. These are exactly the categories where ChatGPT’s commercial query handling — with 30-day freshness windows and direct domain probing — applies most aggressively.

Three concrete actions follow from what this format reveals:

Audit your competitive set as ChatGPT sees it. The brand names written into the fast queries are the competitive shortlist the model already holds for your category. If your brand isn’t in those queries, you are not in the comparison set before the search even starts. Tools that surface ChatGPT’s fan-out queries for your category can identify this gap. A structured precision targeting review of your AI search presence covers this as a first step.

Fix your buying-intent pages first. Content freshness prioritization should start with pricing, comparison, and feature pages — the pages that answer the questions commercial queries ask. Those pages sit in the 30-day window. General blog content and thought leadership operate on longer freshness tolerances and should not compete for the same update budget.

Verify your entity data independently of your website. For operators with a local component — law firm offices, financial advisory branches, physical retail — the business call type means your Google Business Profile, Apple Maps listing, and third-party review aggregators are the actual retrieval source. The website only gets probed to verify details that the listing already returned. Mismatches between listing data and website data create verification failures, not citations.

Operators deploying AI agents for lead qualification should also audit whether the AI answers users receive about your brand are sourced from your own pages or from third-party review sites and aggregators — because the domain slot in ChatGPT’s new format makes that distinction explicit and traceable.

Separately, operators running crypto lead generation programs should map which call types fire for their product category. Software and exchange products appear to stay in the fast vertical — no product card, no merchant catalogue. That means content and entity authority remain the primary levers, and the freshness window still applies to comparison and rate pages.

The broader point is that ChatGPT’s search mechanics are now readable. The format tells you what the model looks for, what freshness threshold it applies, and what domain it intends to probe. Operators who treat GEO as a content volume game will continue to miss these structural factors. The ones who align page freshness, domain integrity, and entity data to what the tool actually calls will see the difference in citation rates within one to two update cycles.

Originally reported by Search Engine Journal, August 2026.