Onton publishes benchmarks behind Ontology 1's 2.7x ecommerce search claim

Why it matters

Onton is betting that structured product knowledge can beat ad-funded keyword search. The 2.7x claim would support that case, but the missing benchmark leaves its size unproven.

Onton co-founders Zach Hudson and Alex Gunnarson announced Ontology 1 on X on July 29th, claiming the AI model is at least 2.7 times more accurate than leading e-commerce search engines and can process queries that previous systems could not handle.

Onton published the Ontology 1 benchmarks and methodology with the launch and released the Subtext Decor 90 evaluation dataset on Hugging Face. The 2.7x figure remains a company-reported result, but the public materials allow researchers to inspect the evaluation design and underlying dataset. Onton did not specify whether Ontology 1 will be released as a standalone model, exposed through an API or used solely inside its shopping products.

The launch advances a technical thesis Hudson and Gunnarson have pursued since Onton operated under its original name, Deft: product search needs structured knowledge and reasoning rather than another chat interface placed over retailer catalogs.

Gunnarson began working on the problem in 2019 after spending roughly 30 hours searching for a gray, mid-century couch with wood trim priced below $800. Hudson had spent 11 years in product roles involving e-commerce before teaching himself to program and building Rcmmd.com, a product review service he described as Yelp for product reviews. The founders met while working on their separate projects, collaborated for several months and eventually combined their efforts.

Onton publishes its benchmark methodology

Onton's architecture combines language and image models with a knowledge graph and reasoning engine. Hudson has argued that conventional large language models can infer likely intent but struggle when product data is incomplete, inconsistent or written to rank in search results.

That distinction matters in searches such as pet-friendly furniture. Retailer copy may omit the phrase entirely, while the product's materials can still indicate whether it resists stains or scratches. Onton says its system can connect those attributes and return relevant products even when a shopper's words do not directly appear in a listing.

Onton previously described a separate product-search experiment in an October 17th, 2025 research post. That test covered about 1,000 products from 20 home-goods brands and 100 generated queries, with results manually checked for relevance. Onton reported that factual descriptions generated from product images improved retrieval by replacing vague, SEO-oriented catalog language with concrete details about color, shape and material.

That earlier experiment explains Onton's focus on catalog quality. The materials released with Ontology 1 now provide the benchmark methodology and evaluation dataset behind the new model's claimed advantage. Independent testing would still be needed to determine whether the reported gain holds across other catalogs, product categories and real-world query distributions.

Onton is pushing beyond furniture discovery

Onton publicly launched as Deft in late 2023 and adopted its current name on September 5th, 2024. Gunnarson coined the name from ontology, reflecting the founders' effort to break products into structured concepts and relationships that software can reason over.

The service began with furniture and home decor, where users can search with natural language or images, generate room concepts and arrange products on visual canvases. Onton's current pricing ranges from a free tier to paid plans for higher image-generation limits and additional design tools. That subscription structure supports Hudson's long-running argument that shopping search should earn money from users rather than sponsored placement.

Onton had more than 2 million monthly active users by November 2025, up from 50,000, according to figures Onton supplied to TechCrunch. The same report said Onton raised a $7.5 million seed round led by Footwork, with Liquid 2 Ventures, Parable and 43 participating. The round brought reported funding to approximately $10 million and was intended to support expansion into apparel and, later, consumer electronics.

Ontology 1 gives Hudson and Gunnarson a technical foundation for that expansion. Furniture searches depend heavily on style, dimensions, materials and visual similarity. Apparel adds fit, sizing and personal taste, while electronics introduce specifications and compatibility requirements. A model that accurately connects those relationships could let Onton reuse its search system across categories without rebuilding each experience around keyword filters.

The next test is whether Ontology 1's benchmarked advantage survives independent evaluation and use beyond the decor dataset. Onton has put a precise multiple on its advantage and published the methodology and data behind that result, giving researchers a basis for scrutinizing the claim.