Search Does Not Just Change Who AI Recommends. It Changes the Words It Uses.

This study tests whether enabling web search changes not only which brands an AI model recommends, but also the vocabulary it uses to describe them.

We reused the same 50 buyer-intent prompts across two conditions: web search enabled and web search disabled. Category and prompt set were held constant, allowing the analysis to isolate changes in the language used by the model.

The results show that search changes the model's vocabulary as well as its recommendations. With search enabled, the model shifts toward composition, ingredient, and specification terms. With search disabled, it relies more heavily on general and unverifiable impression language.

The strongest observed shifts included monohydrate (21.0×), provides (20.0×), whey (17.0×), leggings (15.0×), and ashwagandha (14.0×). Smaller but measurable shifts included chamomile (5.9×), creatine (5.7×), collagen (4.9×), and third party (3.3×).

The frequency ratios are based on normalized word-frequency counts, calculated by dividing the normalized frequency of each word with search enabled by its normalized frequency with search disabled. This controls for differences in response length between conditions.

The companion browsing study found that 76.9% of recommended brands changed when web search was toggled. This study extends that finding by showing that the language itself also changes. The noise-floor figures of 27% with search enabled and 47% with search disabled are reused from that companion study as corroborating context rather than re-measured here.

The experiment used gpt-4o and the same 50 prompts in both conditions. The nine words reported in the dataset represent the clearest observed shifts and are not an exhaustive list of every word whose frequency changed.

The package contains the aggregate word-frequency data, methodology, and primary study chart. No row-level or per-call raw model responses are included.