If someone never leaves your AI assistant to search the web, does that mean they're satisfied — or just stuck?

This explores whether an assistant session that never sends the user out to search or browse means they were satisfied. The short answer from this collection: probably not on its own. No paper here tests that link directly. What the collection does have points to two problems with reading a self-contained session as a happy one: the session may not be self-contained in the first place, and stopping is a weak signal of success.

Start with where these sessions sit in a person's day. A cross-surface panel study Do people use AI assistants before or after searching? found that assistant use comes *after* search and browsing much more often than before it, by 20.6 percentage points. That runs against the popular idea of the assistant as an 'answer engine' that replaces search. The same study found that assistant-only sessions are more common than search-only sessions among the same users. So a session with no web steps inside it may still be the last step of a longer journey that began on the web. The person may have arrived already informed, and the assistant was used to synthesize or finish the job. 'No web steps' describes where the session's boundary falls. It doesn't tell you whether the assistant met the need by itself.

The second problem is with satisfaction. Work on the STORM system Does user satisfaction actually measure cognitive understanding? found that people report being satisfied even while they're still confused, especially when they don't know what they're missing. The better sign of real understanding was continued engagement, not a quick sense of being done. This suggests a counterintuitive reading: a user who never goes looking for more may sometimes be someone who doesn't realize there was more to find. A tidy, contained session fits that pattern just as well as it fits a satisfied one.

Other studies in the collection show why one behavioral signal rarely captures 'it went well.' On phone agents Do phone agents succeed at all three critical tasks equally?, task success, privacy-safe completion and reuse of saved preferences turned out to be separate abilities, and ranking agents by success alone didn't predict the other two. Satisfaction also changes over time. Long-term research on personalized chatbots Does chatbot personalization build trust or expose privacy risks? shows each good interaction raises what users expect next, so the same session can satisfy someone early on and disappoint them later. Some failures also leave no visible trace. When tool-using assistants chain actions together silently When should AI agents ask users instead of just searching?, they can drift from what the user meant without the user ever stepping in. A session can look smooth and contained while answering a slightly different question than the one asked.

What you might not have expected: within this collection, an assistant session with no web steps is better read as a clue about *where a person is in their search journey* than about how satisfied they are. Testing satisfaction properly would mean pairing that behavior with something else, such as whether the person comes back, keeps exploring, or actually understood the answer. That pairing is the gap the corpus leaves open.