How often does AI actually change a buying decision? We looked at real conversations

A recent Exploding Topics and Semrush survey produced an almost perfect split: 57.51% of AI users said a chatbot had talked them into buying something, while 57.5% said one had talked them out of a purchase. It is a striking result, but it answers a person-level retrospective question. It does not tell us what fraction of individual buying conversations end with either outcome.

So we looked at the conversation record instead. The new paper, Purchase Advice and Observable Buyer Responses in Real AI Conversations, studies historical interactions from Aiso's proprietary research database of licensed, consent-based and de-identified conversations with commercially available AI assistants.

Recommendations are easy to see. Decisions are not.

We screened 317 historical interactions and retained 67 purchase-directed episodes for detailed analysis. In 52 of the 67 episodes, or 77.6%, the assistant provided commercially meaningful recommendation content: candidate products, accommodation options, acquisition channels or conditional preferences between alternatives.

The record becomes much thinner immediately after the recommendation. Only 18 of the 67 episodes, 26.9%, contained a later user message that continued the same purchase-related mission. Twenty-three episodes contained some later user message, but several had already changed subject. Treating any later turn as commercial follow-up would therefore overstate useful follow-up availability by 27.8%.

Across the 47 same-mission follow-up messages retained for outcome coding, we did not observe an explicit post-advice purchase commitment, a confirmed completed purchase or an explicit abandonment of the purchase category.

A recommendation is not a conversion

Traditional web analytics often gives marketers a technical path from search or ad impression to click and transaction. Conversational AI breaks that chain. The assistant may influence the consideration set, but the final transaction can happen elsewhere, later and without another message to the assistant.

This creates a denominator problem. Brand mentions are observable. Recommendations are observable. Buyer resistance is sometimes observable. Completed sales are often missing from the conversation record. Counting the first three as if they directly measured the fourth produces a much cleaner number than the evidence deserves.

Negative influence is usually subtler than “don't buy”

We expected more explicit dissuasion. Instead, the clearest negative behavior was narrower: an assistant redirects a buyer away from a specific candidate because another option better fits the stated constraints. We did not observe a case where the assistant clearly advised abandoning the whole purchase category.

For a brand, that narrower effect can still matter. A hotel disappears after the user adds a budget. A product loses when warranty becomes important. Negative reviews trigger a comparison. A competitor becomes the stronger fit. None of those requires the assistant to say “do not buy this.”

What marketers should measure instead

A single sentiment score is too coarse. The useful object is the buying journey. Start with whether the brand appears, then track whether it survives as the user adds constraints and the assistant updates the shortlist.

• Was the brand included in the first answer?
• Was it actually recommended, or only mentioned?
• Did it survive a budget, location or use-case constraint?
• Did the user push back on the recommendation?
• Did another brand replace it?
• Did the user ask how to buy, book, contact or compare it?

Those are different commercial events. Collapsing them into one visibility score hides where the recommendation was won or lost.

Our fourth paper on real AI conversations

This paper extends three earlier Aiso studies. Answer-Reconstruction Search Density measured the query and source work compressed into conversational answers. The Prompt Is Not the Query studied how the user's request state develops across turns. Beyond the Final Prompt measured how removing within-conversation context changes answers.

This fourth paper asks what happens after commercial advice appears. The main result is less tidy than “AI converts X% of buyers,” but more useful: we can observe AI recommendations much more often than we can observe the decisions they eventually influence.

The full 16-page paper includes six figures, four tables, operational definitions, text-free annotations and reproducibility code. Read it on arXiv.