Accel backs Keenable with $26M to build a search engine for AI models

Keenable, a startup that has built a web search index the company says contains more than 100 billion documents, raised a $26 million seed round led by Accel to sell AI labs a retrieval layer built for machines rather than human browsers. The round also included Conviction Partners and business angels, with Accel partner Zhenya Loginov leading the deal. Keenable's API is already in production with "several AI labs and inference providers," the company says, and it has a public partnership with voice-AI startup Gradium for live information retrieval.

Google and Microsoft have been narrowing public access to the search APIs that AI labs have leaned on to ground their models, in moves that Keenable and its lead investor describe as an effort to keep that capacity for the search giants' own bundled AI products. Accel partner Zhenya Loginov framed the gap in the announcement as one of very few web-scale options left for model developers that do not come from the two search incumbents. The mechanism is a kind of unbundling: a layer of the internet that used to be wholesale, the live web search used to ground a model's answer in current events, is being carved off the consumer search product and resold to a different customer.

The customer is the model's retrieval step, not a person. Every time an LLM is asked a current-events question, the answer is supposed to be backed by fresh, citable web evidence, and the index and ranking that produces that evidence is now a separate product. Keenable is selling into that retrieval step with what it calls fine-tuned index structures for AI-scale serving, designed to narrow the search space quickly per query so the model pays only for the documents it actually needs. The technical neighborhood—recency ranking and freshness for narrow-set queries—is the same problem the company says its index is built to solve, and a paper on that topic covers the relevant approach.

The differentiator Keenable is selling is cost at the per-query level, not just the size of the index. The pricing page on keenable.ai ranks the service below serper, parallel-turbo, perplexity, exa, tavily, and parallel on a dollars-per-1,000-queries chart, and frames itself as the "lowest public price." That chart is company positioning rather than a verified third-party benchmark, and Keenable has not published independent measurements, but the broader argument is that an index designed from the start to serve model retrieval queries can undercut general-purpose APIs that were originally priced for human search traffic.

Styskin spent roughly 20 years running search, AI, and cloud infrastructure at Yandex, the default search engine across Russia and most of the former Soviet Union, and later at Amazon. The kind of full-stack web-index engineering needed to operate a 100-billion-document production index is concentrated in a small number of teams worldwide, Loginov said in the announcement, and Styskin's team is one of them. That pedigree is what Accel is underwriting.

Three open questions sit on top of the thesis. The 100-billion-document figure is a Keenable self-claim, not an independent count. The customer set beyond Gradium is undisclosed, and "several AI labs and inference providers" is the company's own framing rather than a named list. The core bet, that AI labs will route around the search giants to a specialized supplier rather than negotiate direct enterprise deals with Google or Microsoft, has not been tested at scale outside of search-adjacent workloads. Watch, in that order: which model labs Keenable names publicly, whether the index ships a third-party audit of its size and freshness, and whether Google and Microsoft respond with their own unbundled, AI-targeted retrieval product rather than continued throttling.