OpenAI Quietly Rewrote How ChatGPT Searches the Web — Here's What Changed for GEO

Sometime between August 16 and August 20, 2026, OpenAI replaced the JSON format ChatGPT used to run background web searches with a compact, pipe-delimited query language of its own. No announcement, no changelog entry. A researcher at Search Engine Journal caught the change by comparing raw network traffic from the same account, four days apart, on the same question.

If you build or manage SEO strategy for clients running on AI-driven visibility, this is the kind of shift that matters more than another core update post-mortem. It changes what a search engine "query" even looks like, and it changes what you can measure.

What actually changed

The old format wrapped every background search in JSON, something like this: `{"system1_search_query":[{"q":"site:intercom.com pricing 2026"}]}`. The new format writes each search as one line, fields separated by pipes:

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

Four parts, and each one tells you something.

• Call type (`fast`, `product`, `business`, `image`, `genui_run`): which retrieval system handles the query.
• The query itself: the words ChatGPT decided to search.
• A number: a freshness window in days, matched to how fast the topic goes stale. A stock price query gets a window of 2 days. A product comparison gets 30. A Reddit opinion check on that same comparison can get 365, or in one captured case, 3,650.
• A domain: when ChatGPT wants one specific site, it now has a dedicated slot for that, instead of writing `site:domain.com` into the query text.

The metadata field that used to mirror this activity for outside tools to read is gone. Browser extensions and GEO trackers built against the old field stopped working the day the format flipped.

The freshness window is the real story for agencies

Here is the part worth building process around. Every commercial or comparison query the researcher captured used a 30-day freshness window. That means ChatGPT treats anything older than a month as stale when it decides what to pull for a buying question.

If a client's pricing page, comparison page, or "vs." page has not changed content or visible dates in the last 30 days, it is competing from outside the window ChatGPT actually searches on the exact query type where brand comparisons get decided. That is a concrete, testable reason to keep commercial pages on a monthly refresh cycle, not just SEO best practice dressed up as urgency.

Search now runs on separate lookup systems, not one engine

The pipe format also exposed that ChatGPT is not running one search function. It routes queries to different backends depending on what you asked.

• `fast` is the general web search, the direct successor to the old fan-out.
• `product` hits a catalogue lookup for physical goods, separate from anything a blog post can influence. If you sell a product, whether it shows up with a card, price, and merchant name depends on catalogue presence, not page copy.
• `business` is a places lookup for local queries. It sends search phrases first, gets back real business names, then verifies them with quoted searches. For local clients, the thing to optimize is the entity in that places data, not the website.
• `genui_run` does not search at all. It calls a widget directly, a stock chart or a sports schedule rendered by OpenAI itself, with no link and no citation opportunity inside it. If a client's traffic depends on ranking for something that now resolves to a widget (match times, stock prices, weather), that query has left the citation game entirely. No amount of content work wins it back.

The Reddit signal got more complicated, not simpler

The same research caught ChatGPT running a dedicated Reddit search after it had already picked its shortlist of brands, using a much wider freshness window (up to ten years) than the brand searches sitting next to it. In one captured conversation, 84 of 221 retrieved sources were Reddit threads, and none of them received a citation in the final answer.

This lines up with separate tracking data from Promptwatch showing Reddit's share of ChatGPT Search citations falling from roughly 3.83% to under 1% in mid-August. The honest read: Reddit discussion may still be shaping the verdict ChatGPT writes, just without the credit. Promptwatch's own numbers complicate the simpler theory that a single fan-out change on August 8 explains the whole drop, so treat any confident explanation, including this one, as provisional. What is not provisional: citation counts on Reddit threads are no longer a reliable proxy for whether that content is influencing the answer.

What to actually do with this

• Put commercial and comparison pages on a 30-day content refresh cycle. Real changes, visible last-updated dates. Treat this as a ranking factor for AI answers, not just a hygiene task.
• Check whether your queries have moved to a widget. Run the searches your clients care about and watch for answers that render as an interface (chart, schedule, map) instead of prose with sources. There is nothing to optimize inside those; redirect effort to queries that still cite pages.
• Verify domain accuracy everywhere a brand is mentioned. ChatGPT fills its domain-targeting slot from memory. A stale or wrong domain on record, after a rebrand or migration, wastes one of a handful of searches on a client's most important comparison queries.
• Stop treating Reddit citation share as the whole picture. A drop in citations does not necessarily mean a drop in influence. If a client has genuine community presence, the incentive to keep it did not go away, only the way you would measure its payoff got harder.

The pattern underneath all four of these: the systems doing AI-driven search are changing their internal mechanics faster than most GEO tooling can track them, let alone most content calendars. Programmatic and AI-assisted SEO work only holds up if what gets published keeps pace with the retrieval layer underneath it. That is exactly the gap TrafficForge's quality scoring is built to close: pages that pass 16 structural and freshness checks before they publish, not content that ages out of relevance the month after it goes live.