Research · published 2026-08-19
How often AI assistants actually search the web before answering a buying question
When you ask an AI which product to use, is it looking anything up, or answering from memory?
Short answer
38.5% of answers to commercial buying questions involved no web retrieval at all, and it depends enormously on how the question is phrased: 91.5% of category questions ("what are the best AI sales agents") triggered a search, against 32.4% of problem questions ("how do I stop losing deals to slow follow-up"). The answers that skip retrieval are drawn entirely from what the model already believes about the market, and no amount of work on your own website reaches them.
The measurement
Across 608 answers to 294 buying questions, 61.5% involved any web retrieval. The remaining 38.5% were produced from what the model already believed about the market.
| Surface | Answers | Searched | Rate |
|---|---|---|---|
| Claude with web search how retrieval was detected: reported directly as a web_search tool call | 294 | 215 | 73.1% |
| OpenAI ChatGPT-search API how retrieval was detected: inferred from whether anything was cited | 20 | 20 | 100.0% |
| OpenAI GPT-5 with web search how retrieval was detected: reported directly as a web_search tool call | 294 | 139 | 47.3% |
| Question type | Answers | Retrieval rate |
|---|---|---|
| capability | 192 | 75.5% |
| problem | 148 | 32.4% |
| category | 82 | 91.5% |
| industry | 76 | 63.2% |
| longtail | 64 | 57.8% |
| comparison | 46 | 45.7% |
Why it matters
Every recommendation in this field assumes the assistant will go and read your page. On a third to a half of answers it does not. For those, the only thing that determines whether a company is named is whether the model already knew it existed, which is a much slower and much older signal than anything a content programme can move in a quarter. It also means any reported 'AI visibility' figure that silently drops the no-retrieval runs is inflated by roughly the size of this number, and most published work drops them.
Method
- The same frozen question universe and the same two surfaces as the citation study.
- For each response we recorded whether the surface performed any web retrieval, independently of whether it cited anything.
- Responses with no retrieval were kept. This is the whole point of the measurement: published GEO research routinely discards them, which conditions every reported rate on a non-random subset.
- Rates are reported per surface and per question intent, because the intent turns out to be the single largest factor: a category question and a problem question differ by nearly sixty percentage points.
What this cannot tell you
Limitations
- Retrieval is inferred differently on each surface. On the Claude surface it is observed directly from the tool call; on the OpenAI search surface it is inferred from whether any citation was returned, which will undercount a retrieval that produced nothing worth citing.
- As above: these are API surfaces and not the consumer products, and consumer products may retrieve far more or far less.
- This measures whether retrieval happened, not whether it helped.
The data
Published under CC BY 4.0. Use it, and cite DFX Intelligence with a link to this page.
A note on who published this
DFX Intelligence sells an AI operator, so we have an obvious interest in this subject. Two things follow from that and we would rather state them than have them noticed. First, the measurement includes our own product and reports its results without adjustment: at the time of publication we appear in none of the answers in this corpus. Second, the question set is ours and is weighted toward the categories we operate in, which is stated in the limitations rather than buried.
Report published 2026-08-19, updated 2026-08-19. Product facts referenced on this page were last verified 2026-08-19.
Facts on this page were last checked against the running product on . Prices are read from the live billing configuration, so this page and your invoice cannot disagree. The machine-readable version of this record is at /ai/entity.json.