AI Search Engine: Onton's Ontology 1 Beat Google and Amazon With 1% of Their Catalogue — No Keywords Required

For nearly thirty years, product search has worked the same way: type a couple of keywords, land on a category page, then narrow the results with filters for size, price, brand and material. That model breaks the moment your intent does not map onto a checkbox. There is no filter for "pet-friendly but still looks expensive", and no dropdown for "furniture that might belong in this room".

On 29 July 2026 a small San Francisco company called Onton published Ontology 1, a neurosymbolic AI search engine model built from scratch, and the coverage caught fire this week after independent write-ups landed on 2–3 August. The headline claim is the sort of thing that usually deserves a raised eyebrow: on a benchmark of complex, intent-heavy shopping queries, Ontology 1 beat both Google Shopping and Amazon while having indexed roughly 1% of their catalogue.

This piece does three things. It explains what neurosymbolic search actually means in plain language, it walks through the benchmark carefully — including the number several outlets have reported incorrectly — and it explains what the result implies if you sell products online, build AI agents, or simply want to know where search is heading after keywords.

What Ontology 1 Actually Is

Onton describes Ontology 1 as "a neurosymbolic model that answers complex, intent-heavy searches more accurately than today's leading engines, and improves itself through continuous self-driven learning." Every word in that sentence is doing work, so it is worth unpacking each one.

Neurosymbolic, in plain language

Modern search is dominated by two techniques. Keyword retrieval matches strings and is unbeatable when you know exactly what you want ("Nikon Z6 III body"). Vector retrieval embeds queries and products into a shared numeric space and finds nearest neighbours, which handles paraphrasing well but blurs precise constraints — it is notoriously bad at negation, counting and hard attributes.

A neurosymbolic system adds a third layer: an explicit, structured representation of the world — categories, attributes, relationships, rules — that the neural model reads from and writes to. The neural half interprets messy human language and images. The symbolic half enforces the facts. Ask for "a rug that isn't wool, under $400, that works in a north-facing room" and the symbolic layer can genuinely exclude wool and genuinely apply the price bound, while the neural layer reasons about what "north-facing" implies for colour and warmth.

This is the same architectural instinct we described in our coverage of Astra's machine-checked mathematical proofs: let a model be creative, then let a formal system verify the claim. Different domain, identical philosophy.

Continuous self-driven learning

The second claim is that the model improves itself without a human labelling queue. In practice this means the system generates its own hypotheses about product attributes and relationships, tests them against retrieval outcomes, and promotes the ones that hold. That is what makes the 1%-of-catalogue figure interesting rather than embarrassing: Onton's argument is that a smaller, better-understood index beats a vast, shallowly-modelled one for questions that require inference.

The Benchmark: Subtext-Decor-90

Onton published its methodology alongside the model, which is more than most vendors do. The benchmark is called Subtext-Decor-90, and here is how it works.

• 90 text queries, curated to include rich aesthetic descriptors, negation, cultural references and emotional framing — the cases that defeat keyword and vector retrieval.
• Three engines compared: Onton, Google Shopping and Amazon.
• Three independent multimodal judges: Claude Opus 4.8, Gemini 3.1 Pro and GPT-5.5, each shown screenshots of all three result sets in a single call.
• Metric: precision at 10 (P@10) on the first ten visible result cards, averaged across judges, with 10,000-resample bootstrap confidence intervals across queries.

The results:

• Onton: P@10 of 0.630, 95% CI [0.571, 0.688] — 52 outright query wins.
• Google Shopping: P@10 of 0.543, CI [0.490, 0.596] — 19 outright wins.
• Amazon: P@10 of 0.469, CI [0.417, 0.521] — 16 outright wins.

The Number Being Misreported

Several outlets have run with a headline along the lines of "2.7x more accurate than the world's best e-commerce search engines". That framing does not survive contact with the data, and it is worth being precise about why.

Ontology 1's accuracy advantage on the published metric is 0.630 versus 0.543 — roughly a 16% relative improvement over Google Shopping, and about 34% over Amazon. Respectable, well outside the confidence intervals for Amazon, and clearly meaningful for a product built by a small team. It is not 2.7x.

Where does 2.7x come from? Outright query wins: 52 versus 19 is 2.7x. That is a legitimate statistic about how often one engine produced the best result set for a query, but it is a win rate, not an accuracy multiplier. Conflating the two inflates a good result into an implausible one, and implausible numbers are exactly what makes buyers dismiss a category.

Three caveats Onton discloses

To its credit, the company flags its own limits. First, precision, not recall — recall is unanswerable without full access to Amazon's and Google's catalogues, so nobody knows what all three engines missed. Second, LLM judges: three frontier models agreeing is a reasonable proxy for human preference on aesthetic queries, but it is a proxy, and models share biases. Third, image and multimodal queries were excluded from the 90 because a one-to-one comparison is impossible — Amazon lacks comparable multimodal search entirely.

That last exclusion cuts both ways. It removes the scenario where Onton claims its widest lead, but it also means the published benchmark understates the differentiator the company is actually selling.

Why "No Keywords" Is the Real Story

The behavioural shift underneath this launch matters more than any single benchmark. Onton cites Bloomreach data indicating that, in 2025, 57% of shoppers used AI to help them shop and 41% now search using natural language rather than keywords. A 2026 Klaviyo study found consumers abandoning keywords in favour of full phrases and questions, with daily AI users routinely issuing queries of eight or more words — sometimes entire paragraphs.

Eight-word queries with negation and emotional framing are a different retrieval problem from two-word queries. Filters cannot express them. Keyword indexes cannot match them. Vector search returns plausible-looking near-misses. This is the gap Ontology 1 targets, and it is why the company built multimodal surfaces — Onton "Canvases", a moodboard-style interface where a picture of your room plus a few words is the query.

Onton's own example is instructive: that query hides judgements about style, scale, palette, what is already in the room and what is missing. No filter touches any of it. This is the same capability curve we tracked in our piece on embodied reasoning models planning in physical spaces — models are getting good at inferring context from images rather than being told it.

The Second Claim: A Trust Layer for AI Shopping Agents

Alongside the accuracy story, Onton positions Ontology 1 as a verification layer: a model built to evaluate whether the product information reaching an AI shopping agent can be trusted. The company argues the industry currently lacks one, and says its internal benchmarks show better product-data accuracy than Google Shopping and Amazon on that axis too.

This is the more strategically interesting claim, because it addresses the failure mode that will define agentic commerce. When a human shops, star ratings, review counts and marketing copy are noisy but survivable — people discount them intuitively. When an autonomous agent shops, that noise becomes input to a purchase decision made without a human reading the page. Wrong dimensions, stale stock, mismatched variants and review manipulation all become financial errors rather than mild annoyances.

Verified product attributes are exactly the kind of capability that gets consumed through a protocol rather than a website. Our guide to MCP servers exposing tools to agents covers the plumbing; a trustworthy product-data endpoint is a natural thing to plug into it.

What This Means If You Sell Online

Ontology 1 itself is not something most businesses will buy this quarter. The transferable lessons are about how discovery is changing, and there are four worth acting on.

1. Structured attributes are becoming your ranking signal

A model that reasons symbolically can only reason over facts it has. Materials, dimensions, care instructions, compatibility, room suitability, weight, country of origin — every attribute you leave in a paragraph of marketing prose is an attribute a neurosymbolic engine cannot use to match you to an intent-heavy query. Machine-readable product data is no longer a feed-management chore; it is discoverability.

2. Long-tail intent is where the traffic is moving

If shoppers are typing eight-word emotional descriptions, your category pages built for two-word keywords are answering a question nobody is asking any more. The practical response is content that names the intent explicitly — use cases, constraints, negations, room and lifestyle framing — rather than another page optimised for a head term. Our AI writing and content hub covers how to do that without producing thin pages.

3. Assume an agent will read your product page

Treat every page as having two audiences: a human and a machine acting for a human. That means schema.org product markup, unambiguous variant data, honest availability and clear returns terms in structured form. The site an agent can parse confidently is the site an agent recommends.

4. Judge new search vendors on your queries, not their benchmark

Subtext-Decor-90 is a decor-heavy benchmark curated by the vendor. That is fine — it is transparent and reproducible — but it is not your catalogue. Build a 50-query set from your own logs, weighted toward the messy, low-converting ones, and score any vendor against it. Our AI for business hub has the evaluation checklist.

The Competitive Picture

It is tempting to read this as a small startup beating two of the largest companies in the world. A more useful reading: Onton is beating them on a slice of the problem they are not optimised for. Amazon's search is tuned for conversion on transactional intent at enormous scale. Google Shopping is tuned for breadth across an index nobody else can assemble. Neither is architected around 90 aesthetic, negation-heavy decor queries.

That does not make the result unimportant. It makes it a classic wedge. The open question is durability: what happens when a hyperscaler applies the same symbolic-plus-neural approach to a catalogue a hundred times larger, and how much of Onton's advantage comes from architecture versus from a carefully curated 1% index. That question will not be answered by another benchmark post — it will be answered by whether shoppers change where they start their search.

Verification Notes

Everything above is drawn from primary and secondary sources you can check yourself: Onton's Ontology 1 research post and its benchmarks write-up with methodology and data, MarkTechPost's coverage, and PPC Land's report on the product-trust angle. For the structured-data work in point three, the canonical reference is schema.org's Product vocabulary. Where a claim is Onton's own and not independently verified — the self-improvement mechanism, the product-trust benchmark — we have said so explicitly.

The Bottom Line

Ontology 1 is a genuinely interesting result reported with an inflated number. The accuracy advantage is about 16% over Google Shopping on the published metric, not 2.7x; the 2.7x figure describes outright query wins. Both are real, and the honest version is still the most credible evidence yet that neurosymbolic retrieval handles intent-heavy natural-language search better than keyword or vector approaches alone.

The strategic signal is bigger than one benchmark. Shoppers have stopped typing keywords, agents are starting to shop on their behalf, and the engines built for a filter-and-category world are the ones with the most to lose. If you sell anything online, the work this quarter is unglamorous and concrete: make your product data structurally true, then test discovery with the messy sentences your customers actually type. Follow the next moves in our generative AI news and AI tool reviews hubs.