Onton's New AI Trust Model Beats Google and Amazon at Product Accuracy

Ask an AI agent to find the best sofa under $2,000 and it will do exactly what it was built to do. It searches, ingests thousands of results, review aggregations, influencer posts, sponsored comparisons, and returns a confident recommendation. The problem is what sits underneath that confidence. Much of the data is synthetic, incentivized, or manipulated, and the agent has no reliable way to separate signal from noise. The consumer buys, the purchase disappoints, and nobody logs it as a system failure. It just looks like a bad sofa. That failure mode is the entire premise of Onton's new model release. The company, which built a neurosymbolic product discovery engine now serving more than two million monthly users, is debuting a trust and authenticity model built from scratch for the agentic web. It is not a smarter shopping agent. It is a model that decides what shopping agents should believe.And the timing is not incidental. The internet's information layer was designed for human evaluation. Reviews, forums, and social proof assumed a person on the other end applying intuition, context and skepticism. As agents absorb the discovery and decision process, those human filters vanish and what remains is a machine-speed system acutely vulnerable to machine-speed manipulation.AI-referred traffic to US retail sites has compounded relentlessly since Adobe began tracking it in October 2024.The scale of the shift is now measurable. Adobe Analytics, tracking over one trillion visits to US retail sites, reported AI-referred traffic up 693% year over year during the November to December 2025 holiday season, up 393% in Q1 2026 and up 138% in May 2026, for cumulative growth of 1,324% since October 2024. This is no longer an experimental channel. It is the fastest-growing front door in retail.The Conversion Flip That Changed the Stakes What makes trust the binding constraint, rather than a nice-to-have, is what AI-referred shoppers now do once they arrive. In March 2025, visitors coming from AI assistants converted 38% worse than traffic from paid search and email. By March 2026, they converted 42% better, a record, and by May 2026 the gap had widened to 54%. A 92-point swing in fourteen months: AI-referred traffic went from the worst-converting channel to the best.Read that swing carefully, because it encodes the whole problem Onton is attacking. Shoppers arriving from AI assistants convert better precisely because they arrive pre-decided. The research happened inside the chat. Which means the quality of the information the model consumed upstream now determines the purchase downstream, with no human skepticism in between. When the deliberation layer moves inside the model, whoever poisons the model's inputs effectively writes the purchase order.The Poisoned Well The corpus those agents are reading is in worse shape than most consumers realize. Researchers cited by the World Economic Forum estimate roughly 30% of online reviews are fake, costing US businesses around $152 billion a year. The platform enforcement numbers tell the same story from the other side: Google removed or blocked over 240 million policy-violating reviews in 2024, Amazon removed 275 million after spending over $500 million and staffing 8,000 people against the problem, Trustpilot pulled 4.5 million fake reviews, and Tripadvisor intercepted 2.7 million fraudulent submissions, including AI-generated photos of hotels that did not exist. Platform enforcement at nine-figure scale, and the fakes keep coming. Note the log axis.Regulators have moved too. The FTC's Consumer Reviews and Testimonials Rule took effect in October 2024, explicitly covering AI-generated fake reviews, with penalties of $51,744 per violation, and the agency issued its first enforcement warning letters in December 2025. But enforcement is a lagging control. It punishes manipulation after the fact. It does not give an AI agent, mid-search, a way to score whether the information in front of it deserves belief. That gap between what regulation can police and what agents need in real time is precisely where Onton has planted its model.What Onton Actually Built Onton's answer is architectural rather than cosmetic. The model was built from scratch, and the company says it evaluates not just what a product is, but whether the information surrounding it can be trusted, interpreting the attributes and intentions behind a user's preferences, including visual inputs, and treating them as meaningful data rather than potential noise. The benchmark claim is the headline. In head-to-head tests against the most dominant players in product discovery, including Google Shopping and Amazon, Onton says its model outperformed on accuracy across essentially every dimension tested, with the sharpest gap in the domain where accuracy has historically been hardest to achieve: interpreting the veracity of product information. The company has published a whitepaper backing the results, alongside new research examining how leading AI systems currently evaluate product-related information and where they fail once synthetic content and incentivized recommendations enter the pipeline. Its conclusion is blunt: existing systems are not equipped for the information environment they are being asked to navigate.Two properties in the release deserve particular attention from the research community. First, Onton says the model learns entirely on its own. Second, it shows signs of being generalizable beyond e-commerce. If both hold up under external scrutiny, this stops being a shopping story and becomes a trust-infrastructure story, because a self-supervised credibility model that transfers across domains is exactly the missing primitive for agentic systems reading news, financial data, or health information, not just sofa listings.The institutional forecasts diverge on scope but converge on direction: agents will mediate a large share of commerce by 2030.The market Onton is building toward is enormous on any methodology. Morgan Stanley projects agentic shoppers will capture $190 billion to $385 billion of US e-commerce by 2030, roughly 10% to 20% of the market. Bain & Company puts the range at $300 billion to $500 billion, or 15% to 25% of US online retail. McKinsey's global framing, which includes the surrounding logistics and payments stack, reaches $3 trillion to $5 trillion. The definitions differ, but the consensus band is unambiguous: within four years, a fifth or more of online buying will run through agents, and every one of those transactions inherits the trust problem Onton is targeting.From Deft to the Trust Layer: The Company Behind the Model Onton has earned the right to make an infrastructure argument faster than most seed-stage companies. Founded by Zach Hudson and Alex Gunnarson and publicly launched in late 2023 under its original name Deft, the company grew from 50,000 to more than 2 million monthly active users, rebranded to Onton, and closed a $7.5 million seed round led by Footwork in November 2025 with participation from Liquid 2, Parable Ventures, and 43, bringing total funding to roughly $10 million. The company reports conversion rates three times the industry benchmark, with over 20% of users active weekly. 40x user growth in under three years, culminating in today's trust model debutThe through-line matters. Onton's stated mission has always been compressing the average 79-day shopping journey to under one day, and its neurosymbolic architecture, which pairs symbolic reasoning with neural models over a continuously refined knowledge graph, was built for exactly the kind of structured, verifiable interpretation that a trust model requires. Today's launch is less a pivot than a thesis reaching its logical conclusion: you cannot compress a decision journey without first deciding what information deserves to survive the compression.Why This Is Infrastructure, Not a Feature Onton's framing, that a trust layer for agentic commerce is a precondition rather than a feature, is the strongest strategic claim in the release, and the market data increasingly supports it. AI-referred shoppers are not merely more numerous. They are the highest-value cohort in e-commerce: 48% more time on page, 13% more pages per visit, 12% higher engagement, 37% more revenue per visit. Every point of trust failure now lands on the best customers a retailer has.That premium is exactly why manipulation economics are shifting. When the highest-converting channel is mediated by models, the return on gaming those models exceeds the return on gaming humans, and the attack surface scales with adoption. Without a credibility layer, every AI-powered purchasing decision sits on a foundation that can be gamed. With one, agents can actually deliver on their promise: better decisions, faster, on information people can rely on. Onton is making the model available today for users on Onton.com and on a case-by-case basis for partners building on the agentic web.Final Thoughts The most consequential line in this announcement is not the benchmark win over Google Shopping and Amazon, impressive as that is. It is the claim that the model learns on its own and generalizes beyond commerce. Product discovery is the ideal proving ground for a trust model, because the feedback loop is brutal and fast: bad information produces bad purchases at scale. But the same architecture, if it transfers, addresses the defining weakness of the agentic era across every domain agents touch. The industry has spent two years in an intelligence race. Onton is betting the next phase is a credibility race, and the structural logic favors that bet. Intelligence without trustworthy inputs simply produces confident errors faster. As agents take custody of a projected $300 billion-plus in US purchasing by 2030, the companies that verify will matter as much as the companies that generate. Onton just claimed the first-mover position in that layer, with a whitepaper on the table and a two-million-user platform already running on it. The burden now shifts to the rest of the ecosystem to test the benchmarks, and to competitors to answer a question they have mostly deferred: not how smart is your agent, but what does it believe, and why.Don’t forget to like and share the story!Vested Interest Disclosure: HackerNoon has reviewed the report for quality, but the claims herein belong to the author. All market data is independently sourced and hyperlinked. Do your own research. #DYOR.