What Is Ontology 1 and Why Is It Turning Heads in AI Search?
A San Francisco-based startup called Onton has released a new AI search model — named Ontology 1 — that claims to outperform both Google Shopping and Amazon on product search accuracy. The model is described as a neurosymbolic AI search model: a hybrid architecture that merges neural network pattern recognition with symbolic reasoning, the kind of structured, rule-based logic that traditional AI systems have long relied on. The result, according to Onton, is a system that doesn't just match keywords but genuinely understands the intent behind complex, conversational product queries.
On a benchmark of 90 queries scored by three independent large language model (LLM) judges, Ontology 1 achieved a mean precision@10 score of 0.630. For context, Google Shopping scored 0.543 and Amazon scored 0.469 on the same benchmark — making Ontology 1 significantly more precise in returning the most relevant results within the top ten listings. What makes this particularly striking is that Onton achieved these results while indexing roughly 1% of the product catalogues that its larger rivals operate at scale. This suggests the advantage lies in the model's reasoning architecture, not in sheer data volume.
For developers, IT decision makers, and digital product teams, this is more than a benchmark curiosity. It raises real questions about whether the dominant platforms — Google and Amazon — have optimised their search primarily for monetisation rather than relevance, and whether leaner, purpose-built alternatives can now credibly challenge them on the metric that actually matters to users: finding the right product, fast.
Neurosymbolic vs. Pure Neural: Why the Architecture Distinction Matters
To understand why Ontology 1's benchmark results are significant, it helps to understand the architectural difference between neurosymbolic systems and the dominant neural transformer models that power most modern AI search engines.
Most contemporary search engines — including the AI-enhanced versions deployed by Google and Amazon — rely heavily on large neural language models. These models are powerful at recognising patterns in language and matching semantic similarity, but they can struggle with compound, multi-attribute product queries. Ask a standard neural search engine to find "a waterproof hiking jacket under £150 that packs small and is available in women's sizing above UK 16," and it may return results that are semantically adjacent but miss critical constraints. Neural models learn from patterns in training data; they don't inherently reason through logical constraints.
Symbolic AI, by contrast, operates through explicit rules and logical relationships — the kind of structured representation found in knowledge graphs, ontologies (hence the model's name), and formal logic systems. The weakness of pure symbolic systems is that they are brittle and difficult to scale to the messy ambiguity of natural language. Neurosymbolic AI attempts to get the best of both approaches: the linguistic flexibility of neural networks combined with the precision of symbolic reasoning.
Research into neurosymbolic AI has been advancing steadily through academic institutions and labs. Work published through venues like arXiv's AI research repository has shown that hybrid architectures can outperform pure neural models on tasks requiring compositional reasoning — exactly the kind of multi-constraint thinking that complex product search demands. IBM Research has also been a notable proponent of neurosymbolic approaches, publishing extensively on the topic and arguing that symbolic grounding is essential for AI systems that need to be reliable and interpretable.
"The gap between what users actually ask for and what e-commerce search engines return has been a persistent, largely ignored problem — because the dominant platforms profit from sponsored placements regardless of organic relevance."
— Industry analyst perspective on e-commerce search optimisationThis is precisely the tension Onton appears to be exploiting. If their benchmark holds up to independent scrutiny, Ontology 1 represents a meaningful step toward search systems that prioritise relevance over revenue maximisation.
Breaking Down the Benchmark: Precision@10 and What the Numbers Really Mean
The metric used in Onton's evaluation — precision@10 — is a standard information retrieval measure. It calculates what proportion of the top 10 results returned by a search engine are genuinely relevant to the query. A score of 1.0 would mean every result in the top 10 is relevant; a score of 0.5 means half of them are. Onton's Ontology 1 scored 0.630, meaning roughly 6.3 out of every 10 results were judged relevant. Google Shopping managed 5.43 out of 10, and Amazon reached just 4.69 out of 10.
The evaluation methodology is also worth noting. Rather than using human raters — which can be expensive, inconsistent, and slow — Onton used three independent LLM judges to score query-result relevance. This approach has precedents in AI research and is increasingly used to scale evaluation tasks, though it introduces its own dependencies on the quality and bias of the judging models. The use of three independent judges helps mitigate individual model biases, as agreement across models provides stronger signal than a single judge.
The 90-query benchmark size is modest — a dataset of this scale is useful for initial validation but may not capture the full distribution of real-world queries, which can number in the millions across diverse categories, languages, and user demographics. Independent replication on larger, more diverse query sets will be essential before these results can be treated as definitive. That said, the margin of difference between Ontology 1 and its rivals is large enough that even with some degradation at scale, a meaningfully superior precision rate may persist.
What Ontology 1 Means for Developers Building E-Commerce and Search Tools
For developers and IT architects evaluating search infrastructure options, Onton's release introduces a genuinely interesting alternative to the status quo. The dominant approach for custom product search has typically involved either building on top of established platforms (Elasticsearch, Algolia, Typesense) or relying on retrieval-augmented generation (RAG) pipelines that bolt LLMs onto existing search indices. Both approaches have well-documented limitations when it comes to multi-constraint, conversational queries.
| Search Approach | Strength | Limitation |
|---|---|---|
| Keyword/BM25 (e.g. Elasticsearch) | Fast, transparent, predictable | Poor semantic understanding |
| Vector/Semantic Search (e.g. Pinecone) | Strong semantic matching | Struggles with logical constraints |
| RAG + LLM pipelines | Conversational, context-aware | Latency, cost, hallucination risk |
| Neurosymbolic (Ontology 1) | Logical precision + semantic flexibility | Early stage; scale not yet proven |
The multimodal aspect of Ontology 1 is also notable. Modern product search increasingly involves not just text queries but image inputs — a user photographing a piece of furniture they want to find a similar version of, or uploading a screenshot of a fashion item. Building multimodal search that handles these inputs alongside text, while maintaining logical reasoning about product attributes, is a non-trivial engineering challenge. If Onton has genuinely solved this in a compact model that can run against a small index, the implications for small and medium-sized e-commerce operators are significant.
According to research from Gartner on AI in enterprise search, one of the primary blockers for businesses adopting advanced AI search is the cost and complexity of infrastructure. A model that achieves superior precision at a fraction of the data overhead could substantially lower the barrier to entry for organisations that cannot compete with the catalogues and compute budgets of Amazon or Google.
The Digital Sovereignty Angle: Could Leaner AI Search Reduce Platform Dependency?
There is a dimension to Onton's release that extends beyond raw benchmark performance, and it matters particularly to organisations concerned with data sovereignty, GDPR compliance, and reducing dependency on dominant US-based platforms. When retailers integrate with Google Shopping or Amazon's search infrastructure, they implicitly hand over significant data about their users' search behaviour, purchasing intent, and product preferences — data that then feeds into the platforms' own advertising and competitive intelligence systems.
The European Commission has been increasingly vocal about the risks of this kind of structural dependency, particularly under the Digital Markets Act (DMA), which designates large platforms as "gatekeepers" and restricts certain data practices. According to the European Commission's Digital Markets Act documentation, retailers using gatekeeper platforms for discovery and search may face restrictions on how their data is used and shared —
Originally reported by MarkTechPost. Summarised and curated by European Purpose.
