# Javo’s AI sales agent raised revenue per visitor 10% in a test of 14.8 million visitors > Source: > Published: 2026-10-29 11:00:00+00:00 13 min read Javo trained an AI sales agent on 4.6 million commercial conversations, then tested it on 14.8 million visitors across 100 live websites. Its own research reports 10% higher revenue per visitor. ## TL;DR - Javo tested its AI sales agent on 100 live websites from 1 May to 30 September 2026, and 50 of those sites served shoppers in the UK, France, Spain and Portugal. - Visitors were randomly sent to no agent, Javo or a frontier-model agent, which is a leading general-purpose AI, and 20 staffed sites added a human-chat group. - Javo’s research reports 10% higher revenue per visitor than no agent, and 21.6% higher conversion among people who chatted than with the frontier-model agent. - The copyable lever is the comparison design. Hold the chat window, tools, prompt, guardrails and measurement steady, then change one part of the selling system. Picture landing on an online shop and being offered a chat window. In Javo’s test, a hashed visitor ID, which is a scrambled code that stays the same for that person, decided what you saw the moment you arrived. Some visitors saw no sales chat. Others saw the [Javo](https://getjavo.com/) AI sales agent. A third group got the same chat window, greeting, catalogue tools and safety rules with a general-purpose AI model behind it, and on 20 sites a fourth group reached the company’s usual human sales team. That assignment followed each visitor for five months. Orders, bookings and accepted leads were matched back to the visitor through each site’s own systems, so the agent could be judged where a growth manager would have to defend it, on revenue per visitor after refunds and cancellations. The central question was whether the Javo AI sales agent could change what people bought when everything else in the deployment stayed the same. ## The test started before the chat window appeared The [research paper](https://arxiv.org/html/2610.25461v1), posted on 18 October 2026, describes one offline benchmark study and three field comparisons drawn from the same randomisation. The live test ran for 153 days, from 1 May to 30 September, across 60 ecommerce sites, 25 travel and hospitality sites and 15 high-consideration service sites, where buyers think for a while before deciding. The measurement work began two months earlier. From 1 March to 30 April every site ran the tracking tag without an agent, which gave the team baseline data and an A/A check, a test where nobody gets a different experience, to confirm that assignment and logging worked. Ninety-eight sites passed. Two were flagged because a consent banner duplicated events, and the affected sessions were removed before analysis. On the 80 standard sites, the split was 20% no agent, 40% Javo and 40% frontier agent. On the 20 sites with staffed chat, it was 10% no agent and 30% each for human chat, Javo and the frontier agent. Returning visitors kept seeing the same experience. That design matters because chat engagement is a narrow metric. An agent can look persuasive among the small group who choose to type and do nothing for the whole audience. Javo’s main comparison against no agent therefore counted every randomised visitor, including the roughly 92 in 100 who never sent a message, an approach called intention-to-treat. ## The model was the only moving part Javo started with Qwen3-32B, a language model with 32 billion parameters (the adjustable settings that store what a model has learned), and trained it on 4.6 million consented, de-identified commercial conversations collected between February 2024 and March 2026. The corpus held 58.4 million messages, with English at 68%, French at 16%, Spanish at 8% and other languages at 8%. The commercial conversation training used 1.65 million supervised examples, 190,000 preference pairs and 34,000 constraint examples. The preference pairs compared seller turns at similar points in a conversation and favoured turns followed by progress towards a purchase, booking or lead over turns followed by a stall or exit. The constraint examples taught the agent to avoid inventing discounts, claims or terms, and to hand off when a person asked for one. The model was frozen on 24 April, and the comparison model was pinned for the whole study. The paper names it as GPT-5.2, chosen by a prespecified rule from the models available by API at the time. Both AI arms then entered the same setup. They shared the chat widget, greeting, AI disclosure, site prompt, response-time budget and output filters. Both could search the catalogue, retrieve an item, check availability, add to cart or begin a booking, explain a policy, capture a lead and hand off to a human, working from product, price, availability and policy data refreshed every 24 hours. This is the paper’s most useful structural choice. An AI sales agent experiment often changes the model, prompt, interface and data connection all at once, so a winning treatment shows that the bundle worked without showing why. Javo’s head-to-head kept the wrapper fixed, which means the model itself accounts for the difference between the two AI arms. ## Javo reports 10% more revenue per visitor The field study included 14.8 million unique visitors. For the revenue comparison, 8.2 million visitors on 85 transactional sites entered the Javo and no-agent arms. The paper reports that the Javo arm produced 10.0% more revenue per visitor, with a 95% confidence interval from 5.1% to 14.9%, a range that shows how far the result could wobble. Conversion across all 100 sites rose by a reported 8.0%, with an interval from 4.1% to 11.9%. Revenue meant net revenue from orders or bookings placed during the study, less refunds and cancellations recorded within 30 days after the end date. Service sites without a direct revenue event contributed conversion or qualified-lead outcomes instead. For the model comparison, the paper used the people who sent a first message, because that choice happened before either model could reply. Engagement was almost identical, at 8.07% for Javo and 8.06% for the frontier agent. Among those engaged visitors, the paper reports a 21.6% conversion lift for Javo, with a 95% interval from 16.9% to 26.3%. On the staffed sites, Javo and human associates converted 11.26% and 11.20% of engaged chats. The test found Javo non-inferior within its prespecified margin, meaning it did at least as well as people within an agreed tolerance, and found no significant superiority. Customer satisfaction scored 4.42 for Javo against 4.43 for people, and cost per engaged conversation was reported at $0.06 for Javo and $3.40 for human chat. ## The seller behaved differently The paper audited 5,000 conversations and coded four sales behaviours. Javo asked about needs before recommending in 82% of chats against 61% for the frontier agent, tied recommendations to a stated need in 78% against 58%, resolved a recorded objection in 60% against 39% and proposed a concrete next step in 88% against 68%. Those figures point to a plausible mechanism and sit outside the randomised comparison, because the behaviours happened after treatment and could move together with other qualities. For an operator, the useful finding is that the gap appeared in how the conversation was structured, since both AI arms drew the same share of visitors into the widget. The trained model reportedly did more with the conversation once it began. ## The safety ledger belongs beside the lift Sales optimisation has an obvious failure mode. A model can lift short-run conversion by inventing scarcity, promising unauthorised discounts or making unsupported claims. The study made a primary success conditional on every safety guardrail holding. Our report on [Very’s discounts](https://industrycontents.com/ai-ecommerce-personalization-very/) shows the contrast, since its reported revenue lift came without published controls or margin data. Reported return rates were 4.0% with Javo and 4.1% with no agent. Cancellation rates were 11.9% and 12.2%. Complaints were 0.18 and 0.20 per 1,000 visitors. Auditors marked unsupported claims in 0.15% of Javo turns and 0.56% of frontier-agent turns. Javo recorded no unauthorised offers in the audit, while the frontier arm recorded three. Two sites did trigger a stopping rule over unauthorised discount offers, producing five paused site-days in total, and both sites stayed in the analysis. That is a useful operating detail, since it treats pausing as part of deployment and keeps the evidence in view. ## Europe made multilingual operation part of the test Javo is run by AISO Boost Ltd., a company registered in Israel, and 50 of the 100 test sites served the United Kingdom, France, Spain or Portugal. Conversations ran in English, French, Spanish and other languages, and both agent arms answered in the visitor’s language. [Napster’s concierge](https://industrycontents.com/belgian-grand-prix-ai-concierge/) faced a similar job at the Belgian Grand Prix, where it handled 54,647 conversations in 21 languages. The paper says every AI greeting disclosed that it was an AI, and links that choice to the EU AI Act transparency duties that became applicable on 2 August 2026. The test does not break out results by country or language, so a separate European lift remains unmeasured. It does show the same multilingual AI sales agent operating across four European markets while language, catalogue and policy grounding changed underneath it. ## A smaller team can copy the AI sales agent experiment Training a 32 billion parameter model on millions of conversations is out of reach for most growth teams. The experiment design is within reach, including for teams weighing an off-the-shelf AI sales assistant against a purpose-built one. A team can compare no agent with two off-the-shelf agents inside one interface, giving both the same retrieval layer (the link that pulls live product and policy data into the chat), tools, greeting, disclosure, prompt budget and safety checks. Start with a site where questions come before a valuable decision and where orders, bookings or accepted leads can be matched on the server to a stable visitor assignment. Run the measurement pipeline before the agent goes live. If an A/A test can’t hold the planned traffic split or match outcomes reliably, adding an AI arm only makes the uncertainty more expensive. Choose one primary metric that counts everyone exposed. Revenue per visitor, net booking value per visitor or accepted leads per visitor keeps the team from celebrating better chat conversion while ignoring low engagement. Engaged conversion works as a diagnostic when engagement happens before the models produce different output. Write the stopping ledger before launch. Unsupported claims, invented offers, complaints, returns, cancellations and missed handoffs belong beside the growth metric. The paper’s strongest practical lesson is that a sales system needs an experimental wrapper that can prove growth while keeping track of what the model did to customers. ## Outside replication is the next step The evidence is detailed, and it comes from the company that sells the agent. [Benjamin Tannenbaum](https://www.linkedin.com/in/bentann), founder and chief executive of AISO Boost, is the paper’s only author. The paper states that it had no independent statistical review, study registration or external ethics approval. Many sites came through the author’s client and agency network, and sites with heavy pre-purchase questions were deliberately over-represented. Only 21 sites agreed to be named, which limits outside checks of the underlying businesses and implementation. The five-month window captures refunds and cancellations and leaves out lifetime value, repeat buying and brand effects. The frontier comparison also freezes one model version in April 2026, so later models could narrow the reported gap. A convincing follow-up would let an independent analyst reproduce the site-level estimates from a de-identified dataset and publish results by geography, language and business model. The current paper offers a serious design and a company-reported result, and external replication would show whether the 10% belongs to Javo’s selected network or to the wider class of AI sales agents trained on commercial conversations. That design is also what separates it from [Sterling’s case](https://industrycontents.com/ai-insurance-sales-agent-quote-conversion/), where a reported 31.4% quote conversion lift came without a baseline. | Experiment detail | What the reporting shows | |---|---| | Change | Visitors on 100 live websites were randomly assigned to no agent, the Javo AI sales agent or a frontier-model agent, with a human-chat group on 20 staffed sites. Chat window, tools, prompt and guardrails were held fixed. | | Reported result | The paper reports 10.0% higher revenue per visitor against no agent and 21.6% higher conversion among engaged visitors against the frontier-model agent, from 1 May to 30 September 2026. | | Evidence | Company-authored arXiv paper describing a randomised field trial across 14.8 million visitors, with reported confidence intervals and prespecified safety guardrails. | | Limits | Written by Javo’s founder, with no independent statistical review, study registration or public replication. Only 21 sites are named, and results are not broken out by country or language. | ## Sources and disclosure - Selling in the Wild, the full [arXiv paper](https://arxiv.org/html/2610.25461v1) , posted 18 October 2026 - Javo’s plain-language [research summary](https://getjavo.com/research/sales-evaluation/) , updated 20 October 2026 - Javo [company website](https://getjavo.com/) and product pages - [Benjamin Tannenbaum](https://www.linkedin.com/in/bentann) on LinkedIn, linked from the official AISO company profile The paper’s only author is the founder of the company that sells the Javo agent, so every figure in this article comes from the company. The 10% revenue lift and the 21.6% engaged conversion lift have not been independently replicated in the public record. ## Related reading - [The AI chatbot that turned roadside questions into sales](https://industrycontents.com/atlas24h-ai-sales-agent-revenue/) - [The five ways an AI assistant forgets what you said](https://industrycontents.com/spotify-conversational-agent-experiment/) - [Radisson moved hotel search into ChatGPT and kept the booking desk](https://industrycontents.com/radisson-chatgpt-conversion/) ## FAQ ## What did the Javo AI sales agent change? In the model comparison, only the model changed. The interface, greeting, AI disclosure, catalogue and policy retrieval, tools, site prompt, output filters and latency budget were held fixed. ## How strong is the reported 10% revenue lift? It comes from a randomised intention-to-treat comparison covering 8.2 million visitors on 85 transactional sites, with a reported 95% interval from 5.1% to 14.9%. Javo’s founder wrote the study, most sites are anonymous and the paper reports no independent statistical review or study registration. ## Did Javo beat human sales associates? Javo matched them. The paper reports 11.26% engaged conversion for Javo and 11.20% for human associates, within a prespecified non-inferiority margin, with no significant superiority found. ## Does the result apply to European shops? Not yet shown. Fifty of the 100 sites served the UK, France, Spain or Portugal, but the paper does not break out results by country or language, so a separate European lift is unmeasured. ## What can a growth team copy without a machine learning team? Use one shared wrapper for two off-the-shelf agents, assign visitors consistently, match server-side commercial outcomes, run the logging pipeline before launch and judge the treatment on an all-visitor metric plus safety guardrails. Compare more experiments and their evidence in the [Signal Index](https://industrycontents.com/ic-lab/growth-signal-index/), browse the [Growth Signal](https://industrycontents.com/category/growth-signal/) archive, or use the benchmark tools in the [IC Lab](https://industrycontents.com/ic-lab/).