Onton says its from-scratch model, Ontology 1, scored higher than Google Shopping and Amazon on product data accuracy in internal benchmarks, positioning the tool as a trust layer for AI shopping agents that the San Francisco company says the industry currently lacks.

Onton, the product discovery company behind a search engine for consumer goods, today launched Ontology 1, a model built to evaluate whether product information reaching AI shopping agents can be trusted. According to Onton, the model was constructed from scratch and, in head-to-head benchmarks, outperformed Google Shopping and Amazon on accuracy across nearly every dimension the company tested. The announcement lands as AI systems take on a growing share of the research and recommendation work that human shoppers once did themselves, a shift that has already reshaped how retailers think about product data across the wider retail media ecosystem.

The company frames the release as a response to a structural problem rather than a routine feature update. Purchasing decisions, Onton argues, increasingly pass through AI agents that research options, compare listings, and recommend products with little human oversight along the way. Those agents pull from a mix of aggregated reviews, influencer content, and sponsored comparisons, and according to Onton, a meaningful share of that material is synthetic, incentivized, or otherwise manipulated. Without a mechanism to separate credible signals from manufactured ones, the company says, agents recommend anyway, and the resulting purchase looks like an ordinary bad decision rather than a systemic failure.

A model built for a web designed for humans

Onton's central argument rests on a distinction between how the internet was built and how it is now being used. Reviews, forums, and social proof, the company says, were designed for human evaluation, where a person could apply skepticism, context, and intuition before trusting a recommendation. As AI agents absorb more of that evaluative work, those human filters do not transfer with them. What remains, according to the company, is an information environment that is unusually exposed to manipulation because the systems parsing it were never built to detect it.

Alex Gunnarson, co-founder of Onton, described the company's focus as distinct from the broader push toward more capable agents. "Everyone is focused on building smarter agents," Gunnarson said. "We're focused on a different question: what should those agents trust? As commerce moves from clicks to conversations and eventually autonomous actions, the quality of information underneath those decisions becomes critical infrastructure."

Ontology 1 attempts to address that gap by evaluating not only what a product is but whether the information describing it holds up to scrutiny. The model interprets the attributes and intentions behind a shopper's stated preferences, including visual inputs, and treats them as usable data rather than noise to be filtered out. Onton says the practical result is an agent capable of grounding its purchasing recommendations in information that has actually been assessed for trustworthiness, rather than passed along unexamined.

The company also says the model shows early signs of generalizing beyond the ecommerce context it was built for, and that it learned largely without human-labeled supervision. Onton describes this as a research finding that extends past product discovery, though the company has not published the technical benchmarks, peer-reviewed data, or independent audit that would allow outside researchers to evaluate that generalization claim directly.

Benchmarking against Google Shopping and Amazon

Onton's comparison singles out Google Shopping and Amazon by name as the incumbent players in product discovery against which Ontology 1 was tested. The company says its model outperformed both on accuracy across essentially every dimension evaluated, with the widest gap appearing in a category Onton describes as historically the hardest to get right: assessing the veracity of the product information itself, rather than simply matching a query to a listing.

No independent party has reviewed these results as of publication. The comparison originates from Onton's own whitepaper and press materials, and the company has not disclosed the specific accuracy percentages, the benchmark dataset composition, or the methodology an outside lab would need to reproduce the test. That distinction matters for a marketing audience accustomed to vendor claims arriving alongside third-party verification, particularly in a year when PPC Land has tracked a steady stream of research questioning how reliable AI-generated and AI-mediated content actually is once it reaches consumers.

The timing situates Onton's claim within a broader pattern that PPC Land has documented extensively. Yelp's April 2026 study found that only 15 percent of users trust AI search results outright, with 80 percent of restaurant search respondents in that research preferring results featuring authentic human content over AI-generated alternatives. A separate Raptive study from July 2025 found that suspected AI-generated content reduces reader trust by nearly half and produces a 14 percent decline in both purchase consideration and willingness to pay a premium for advertised products. Together, these findings describe exactly the trust deficit Onton says its model is designed to close, even though Onton's own benchmark results have not undergone the kind of independent scrutiny that produced those trust figures.

The wider push toward agentic commerce infrastructure

Onton's launch arrives roughly six months into a period of rapid infrastructure building around AI-mediated shopping. Google introduced the Universal Commerce Protocol on January 11, 2026, in partnership with Shopify, Etsy, Wayfair, Target, and Walmart, establishing open standards for AI agents to discover products and complete transactions across retail platforms. Google later unveiled a dedicated shopping ad format for AI Mode on February 11, 2026, positioning retailers to appear within AI-generated conversations at what the company called key moments of discovery. By April 2026, Amazon, Meta, Microsoft, Salesforce and Stripe had joined the Universal Commerce Protocol's Tech Council, a notable shift for Amazon in particular, which had previously blocked AI bots from OpenAI, Anthropic, Meta, Google, and Huawei before pursuing its own proprietary agentic shopping tools.

Amazon's own investment in agentic shopping has already produced measurable commercial results. The company's Rufus assistant drove 12 billion dollars in sales during 2025, with more than 300 million customers using the tool over the course of the year, according to statements Amazon included in its earnings materials. That scale illustrates why questions about the trustworthiness of the data feeding these systems carry commercial weight well beyond any single vendor's product launch.

A parallel infrastructure effort has emerged around the Model Context Protocol, the open standard Anthropic developed and later donated to the Linux Foundation. The Ad Context Protocol, built on that standard, launched on October 15, 2025, with founding members including PubMatic, Scope3, Swivel, Triton Digital, Optable, and Yahoo, aiming to let AI agents negotiate and execute media buys directly with publisher systems. The IAB Tech Lab has pursued a parallel initiative, naming its own umbrella framework Agentic Advertising Management Protocols on February 26, 2026. Onton's announcement does not reference either protocol directly, positioning its model instead as a trust layer that could sit underneath whichever transactional infrastructure eventually wins broader adoption.

Why data accuracy has already become a live problem

Google's own product data infrastructure offers a useful point of comparison for why Onton's accuracy claims matter to marketers specifically. PPC Land's coverage of Google's Shopping Graph documented more than 60 billion product listings, with over 2 billion receiving updates every hour, feeding into both AI Mode and traditional Shopping surfaces. Separately, an independent analysis found that Google Merchant Center's automatic product import was failing to meet its own stated 24-hour refresh cycle, with one documented case showing a nine-day gap between the platform's declared update frequency and the actual timestamps on product data. That gap illustrates a distinct but related problem: even without deliberate manipulation, product data feeding AI shopping surfaces can be stale or incomplete, a condition that would presumably also challenge any trust-scoring model layered on top of it.

The broader manipulation problem Onton describes has also been documented directly. A Wall Street Journal investigation published January 30, 2026 found that businesses pay to influence ChatGPT and other AI chatbot recommendations through tactics known as generative engine optimization, with AI chatbot referrals accounting for as much as 44 percent of traffic for some optimization firms, up from nearly zero a year earlier. A separate controlled experiment, published July 15, 2025, and covered by PPC Land, demonstrated that AI responses across ChatGPT, Perplexity, Google Gemini, Claude, and DeepSeek could be manipulated within days using strategically placed content on low-authority domains with no prior connection to the test topic. These findings describe precisely the vulnerability Onton says its model is built to close, though they also underline how difficult that problem has proven for other well-resourced organizations to solve.

What the announcement leaves open

Onton says Ontology 1 is available today directly to consumers through the company's own product, and separately, on a case-by-case basis, to partners building agentic commerce products who need what the company describes as a trustworthy foundation for product recommendation. The company positions the release as a step toward what it calls a broader trust layer for the internet, though it has not disclosed pricing for partner access, a public API, or a timeline for expanding availability beyond its own platform.

Zach Hudson, the company's other co-founder, framed the stakes in platform terms rather than product terms. "We believe the next major internet platform will not be defined solely by who has the best model," Hudson said. "It will be defined by who can provide the most trustworthy foundation for those models to operate on."

That framing places Onton in competition not with a single feature at Google or Amazon, but with the broader infrastructure layer both companies are already building through the Universal Commerce Protocol and their respective agentic shopping tools. Whether a startup-scale model can establish itself as neutral infrastructure underneath platforms controlled by two of the largest companies in commerce remains an open question the announcement itself does not answer. Onton's benchmark claims, absent independent verification, function for now as a marketing assertion rather than a settled technical result, and the marketing community will likely want to see reproducible testing before treating the accuracy gap as established fact.

For advertisers and publishers, the relevance sits less in Onton's specific product than in what its launch confirms about the direction of the underlying problem. As PPC Land's coverage throughout 2026 has shown, product data accuracy, AI-mediated discovery, and synthetic content detection have moved from academic concerns to active commercial battlegrounds, with Google, Amazon, and now smaller entrants like Onton all making competing claims about who can be trusted to get it right.

Timeline

Summary

Who: Onton, a San Francisco-based product discovery company, and its co-founders Alex Gunnarson and Zach Hudson, alongside the incumbent platforms named in its comparison, Google Shopping and Amazon.

What: Onton launched Ontology 1, a trust and authenticity model designed to evaluate whether product information reaching AI shopping agents is genuine or manipulated. The company says the model outperformed Google Shopping and Amazon on accuracy in internal benchmarks, with the largest gap appearing in assessing the veracity of product information itself.

When: The announcement was made today, July 29, 2026, according to the company's press release.

Where: Onton is headquartered in San Francisco, California, and the model is available today through the company's own platform at Onton.com, with case-by-case access offered to partners building agentic commerce infrastructure.

Why: The launch responds to a documented trust gap in AI-mediated shopping, where research from Yelp, Raptive, and independent investigations covered by PPC Land has repeatedly shown that consumers distrust AI-generated recommendations and that AI systems remain vulnerable to manipulated or synthetic content. As agentic commerce infrastructure from Google, Amazon, and industry protocols like the Universal Commerce Protocol and the Ad Context Protocol continues to expand, the question of which information those systems can trust has become a commercial and competitive issue rather than a purely technical one.