Onton Releases Ontology 1: A Neurosymbolic Search Model That is 2.7x More Accurate than the World’s Best E-commerce Search Engines
MarkTechPost Michal Sutter
A small SF startup's search AI beat Amazon and Google Shopping at understanding fuzzy product queries like 'pet-friendly couch.' It did this with 1% of their catalog — but you can only use it on Onton's own site, not download it.
Onton, a scrappy search company out of San Francisco, just published a benchmark claiming its new model, Ontology 1, beats both Amazon and Google Shopping at the kind of search query that trips up every e-commerce engine on earth: the vague, human, context-loaded one. Think "pet-friendly sectional" or "something to put on a weirdly deep windowsill." Across 90 such queries, scored by three separate AI judges (Claude Opus 4.8, Gemini 3.1 Pro, and GPT-5.5), Ontology 1 posted a mean precision@10 of 0.630. Google Shopping landed at 0.543, Amazon at 0.469. Onton won 52 of the 90 queries outright.
What makes this notable isn't just the score, it's the approach. Most product search still runs on a filter logic that hasn't changed much since the late 1990s: category, price, brand, material. There's no checkbox for "pet-friendly," and there's certainly no checkbox for "fits my weirdly shaped room." Ontology 1 skips the seller's self-reported tags entirely and instead builds an explicit, inspectable knowledge graph, reasoning through properties like fiber, weave, and construction to infer things like cleanability, rather than trusting a marketing claim that might be false or simply missing. When it hits a concept it doesn't have a model for, it works one out on the fly and reuses that reasoning on future queries. That's the neurosymbolic pitch: not a black box absorbing patterns, but a structured system you can actually trace.
It's not flawless, and Onton is upfront about that. On queries where Amazon's sheer category depth does the heavy lifting, like finding a reading lamp dim enough not to wake a partner at 3am, Amazon crushed Onton 0.9 to 0.4. On the deep-windowsill query, the gap was even starker: 0.67 versus 0.07. Onton chalks this up to catalog breadth, since it's indexing roughly 1% of what Amazon or Google carries and operating in a single vertical, home decor, with zero sponsored placements muddying results.
There's also a real asterisk on the headline number. Krippendorff's alpha across the three AI judges came in at 0.465, which is not a strong agreement score, meaning the exact precision figures are noisy and somewhat judge-dependent. To Onton's credit, all three judges still ranked the three engines in the identical order, which is the part that actually matters if you're trying to prove a directional claim rather than a precise one.
Underneath all this sits Ograph, a custom graph database Onton built to run the reasoning fast enough to matter, reportedly outpacing the widely used SuiteSparse:GraphBLAS library by roughly 100x per core, with a GPU variant running 43x faster still. None of this is available as a download, though. There's no open checkpoint, no public API, no pricing page. You can try it live at Onton.com, or, if you're a retailer or platform building agentic shopping tools, you can apply for partner access and wait for Onton to decide if you're a fit.
My take
I like the science here more than the rollout. A self-graded, three-LLM-judge benchmark with a Krippendorff's alpha under 0.5 is exactly the kind of number that should come with a footnote the size of the headline, and burying the caveat in paragraph six of a press-friendly post is a familiar move. But the underlying idea, decomposing vague human intent into checkable properties instead of trusting seller tags, is genuinely the right instinct, and it's a shame Onton is choosing closed, case-by-case gatekeeping over letting the wider search and agentic-commerce world stress-test it. If the method really generalizes past furniture, prove it in the open.
Read more about this at: MarkTechPost
