We already know search changes which brands get recommended: 77% of picks flip when browsing is switched on. This time we held the brands and prompts constant and looked only at the words. With search on, the model reaches for ingredients and specifications. With search off, it reaches for impressions. Retrieval pulls facts. Memory pulls a feeling.
Our browsing study established that turning search on changes 76.9% of recommended brands, far above the 39.1% you would expect from noise alone. That answered which brands the model reaches for. It did not answer what the model says once it has reached for them.
If search genuinely changes what the model is working from, not just which name comes out the other end, that should show up in the vocabulary itself. A model reasoning from a product page it just read should describe things differently than a model reasoning from whatever impression of a brand it already carries in memory.
We ran the same 50 prompts through the model twice, once with search available and once without, holding category constant across both runs. Then we counted word frequency in each condition and normalized by total word volume, so a word that just appears more because search-on responses are longer does not get credit it has not earned.
Every word at the top of the list is a composition or specification term: monohydrate, whey, ashwagandha, chamomile, creatine, and collagen are ingredient names. Leggings and provides are specification language, the kind of word that shows up when a claim needs to be precise rather than pleasant.
This lines up exactly with what we already knew about how confident the model is in each condition. In the browsing study, the noise floor, how often the model changes its own answer for no reason at all, sits at 47% with search off and drops to 27% with search on. The model is not just picking different words when it can read a product page. It is guessing less.
That is the whole story in one line: when the model has something concrete to read, it reports what it read. When it does not, it reaches into memory for a general impression and states that instead, with the same fluent confidence either way. The vocabulary shift is not decoration. It is a visible trace of a model doing two different things depending on whether retrieval is available.