How people really search
inside AI conversations
We publish the measurement work behind Aiso as preprints, so you can check the method before you trust the numbers. Every paper has a plain-language write-up.
Answer-Reconstruction Search Density
The median information-seeking answer packed 11 retrievable facts into 3 facets across 183 conversations. One AI answer can absorb a bundle of searches that query analytics count once.
The Prompt Is Not the Query
Across 8,133 conversations the final prompt carried a median 36% of the session's user vocabulary, and about half the time a need stated earlier never reached that last prompt.
Beyond the Final Prompt
In 180 paired conversations, answers built from the full history differed materially from final-prompt-only answers in 44.7% of cases. Conversation context changes what the model says.
Purchase Advice and Observable Buyer Responses in Real AI Conversations
Recommendations appeared in 77.6% of purchase-directed episodes, but only 26.9% carried a later message that continued the same buying mission.
Scoring With the Engine
Across the same prompts, four live AI engines barely cited the same URLs. Mean cross-engine overlap was 0.8%, and 84.9% of engine pairs shared no cited URL.