Knowledge baseTermsSearch decision

Search decision term/search-decision

Terms

What it is

Before ChatGPT answers, it decides whether to query the web or answer from memory. Until summer 2026 that decision was readable in the conversation object: metadata.sonic_classification_result carried three probabilities (simple_search_prob, complex_search_prob, no_search_prob), the thresholds (no_search_threshold 0.12, complex_search_threshold 0.4) and the outcome. Since our measurements began on 20 Aug 2026 the field is absent. What remains is the outcome: if the conversation contains a message addressed to web or web.run, a search happened (routing.searched).

Measured 22 Aug 2026: 18 of 26 questions searched in standard mode. No search for evergreen knowledge (K4, "why is the sky blue"), long-tail problem descriptions (K12, sourdough, barking dog), the introspection prompts (K14) and both health questions (K9).

External evidence

For visibility

A question ChatGPT does not search on is closed to every website, regardless of content, authority or technology. The search rate per topic is therefore the first number an SEO needs before thinking about optimisation. Blyskal measures a 36.6 % search rate for Claude across all prompts and considerably more for ChatGPT; our German prompts sit at 69 %, with clear topical patterns.

The decision is not fixed: the same question can search one day and not the next. Only a time series shows whether the regime changes or merely rolls dice. That is the search-rate metric.

External sources

green-2025-08-09 · mohanadasan-2026-06-24 · blyskal-2026-07-22

Change history

No event has touched this object yet.

Open

Related

Search rate Query fan-out web.run syntax (pipe format) Retrieval pipelines (Labrador, scrape)