.png)
OpenAI appears to have changed ChatGPT Search behavior in a way that materially increases the use of site:-scoped searches, and Promptwatch’s tracking suggests the change landed alongside the GPT-5.6 rollout. The same telemetry also points to a sharp reduction in Reddit sourcing, which makes this more than a cosmetic search tweak.
Promptwatch, a GEO/“Generative Engine Optimization” vendor that automates prompt tracking across ChatGPT, Claude, and Gemini, reported a notable jump in the share of ChatGPT Search fanout queries containing the site: operator. Their aggregate chart hovered around 0.3%–0.5% for weeks, dipped to roughly 0.15% on August 3–5, then jumped to 16%–17% on August 8.
That kind of discontinuity is hard to explain as organic drift. The most plausible reading is a staged rollout or a backend search-policy change that moved site: from rare edge-case usage to a first-class routing primitive for a large fraction of search fanouts.
The caveat matters: these numbers only cover the prompts Promptwatch has instrumented. This is not a global measurement of ChatGPT Search traffic, but it is still enough to detect product-level behavior changes that would otherwise be invisible.
OpenAI’s August 6 GPT-5.6 announcement was vague: “For Plus and Pro users, we’re updating GPT‑5.6 Sol in Chat to be more reliable with facts and provide more focused answers.” That reads like a quality update, but the search behavior shift suggests the implementation changed in the retrieval layer, not just in generation.
The shape that best matches the observed behavior is something like search(query, recency, domains) rather than a model that freely emits site: as part of the natural-language query. In other words, the system likely moved toward explicit domain constraints in the search controller, with the model or router deciding when to narrow the candidate set.
For ML engineers, that distinction matters. Once the retrieval policy is encoded as structured parameters, you can tune factuality, source diversity, and domain preference without asking the model to synthesize search syntax itself. It also makes policy enforcement easier: you can suppress low-trust domains, prioritize official documentation, or bias toward recent pages at the orchestration layer instead of relying on prompt behavior.
Promptwatch followed up on August 18 with a second claim: ChatGPT seemed to have greatly reduced the likelihood of Reddit being used in search results. That aligns with the broader site: shift, because site:-scoped retrieval makes it easier to route factual queries to official sources and away from high-variance community content.
The important part is not “Reddit down” in isolation. The important part is that ChatGPT Search appears to be getting more selective about source classes. If that holds, the system is likely optimizing for answer reliability and citation cleanliness, even at the cost of losing some long-tail, experience-based content.
I would not over-interpret this as a universal anti-Reddit policy without stronger evidence. OpenAI obscures system prompts, and the available leaked prompt collections do not show a confirmed change. But the behavior change is consistent with a retrieval policy update that deprioritizes community forums for many query types.
If you work on search visibility, content strategy, or AI answer monitoring, this is the kind of backend change that breaks assumptions quickly. A model that shifts from broad web fanout to domain-scoped retrieval can change citation share overnight without any obvious product announcement.
For evaluation, I would track three things separately: source-domain frequency, query-class routing, and citation stability under repeated prompts. Aggregate citation counts are too blunt once the system starts applying source filters upstream. You want to know whether the retriever is narrowing to official domains for factual queries, whether forum content is still used for opinion or troubleshooting queries, and whether the same prompt produces different source mixes across time windows.
For production teams building retrieval or answer engines, the lesson is straightforward: source policy is now a first-order product surface. If your system depends on a particular corpus mix, you need instrumentation that can detect routing changes, not just answer-quality regressions.
