Javo raises sales outcomes in live website tests

Can training on real sales conversations change what visitors buy, rather than just how an agent scores with simulated customers?

The paper connects those questions by testing Javo on offline sales tests and randomized live website traffic, including comparisons with no agent, another model, and site associates.

Javo trains an open-weight sales agent on full commercial conversations, then tests whether offering it changes live-site purchasing outcomes.

Javo is an open-weight sales agent trained on whole commercial conversations. The authors first measured its buying-intent and selling-performance scores in simulated conversations, then randomized website visitors to Javo or alternatives.

Against no agent, Javo changed net revenue per randomized visitor by +10.0% across 85 transactional sites, including visitors who never chatted.

That comparison answers whether offering Javo changed short-term sales outcomes on those sites.

Separate tests found higher engaged conversion than the tested frontier agent and non-inferiority to existing site associates under a prespecified margin, though those comparisons had different conditions.

The sites consented to participate and were not a random sample of the web. The field results cover five months, and the frontier comparison tested one model version; they do not establish longer-term effects or results across other sites and models.