ChatGPT Search
By Şamil Quliyev · Updated 2026-09-15
ChatGPT Search is the browsing mode of ChatGPT: when a question looks like it needs current information, the model issues a web search and writes its answer from the retrieved pages, citing them inline. Without that retrieval step the same question is answered from training memory, which is a different measurement entirely.
How it picks sources
Retrieval returns a small set of pages, and the model then writes a synthesis that leans on the two or three it finds most usable — typically pages that state the fact plainly, near the top, in a structure that survives being quoted out of context. Pages that bury the answer under narrative, or that require the model to infer a claim from a table it cannot parse, are retrieved and then passed over.
Grounding behaviour
Grounding is conditional, not guaranteed. The same prompt can produce a grounded answer in one run and a memory answer in the next, and the two look similar to a reader while meaning opposite things for measurement: one reflects what the web says about you today, the other reflects a training snapshot. Any serious tracking records the grounding status per answer and keeps memory answers out of visibility rates.
How to track it
Query it through the API with the web-search tool enabled, sample each prompt several times per cycle, and store the full response plus its citation list. Rates are computed only over answers that actually retrieved. A tool that cannot tell you whether a given answer searched is reporting a blend of two different things.
Measurement mistakes
- Checking a prompt once in the chat interface and treating the result as a measurement — personalisation, memory and retrieval variance all sit in that single answer.
- Counting memory answers as visibility, which measures the training snapshot rather than the current web.
- Ignoring the citation list. Being described without being cited is a different problem from not appearing at all, and only the citation list separates them.
What TAMGIO measures here
This engine is part of TAMGIO ongoing tracking: prompts are sampled on it every cycle.
Frequently asked questions
- Does ChatGPT always search the web before answering?
- No. Retrieval is triggered when the question looks like it needs fresh or specific information; otherwise the model answers from memory. That is why grounding status has to be recorded per answer rather than assumed.
- How many samples per prompt are enough for ChatGPT tracking?
- Enough that one unusual answer does not move the rate: three per prompt per cycle is a workable floor, and the sample size should be reported next to any rate derived from it.
