Keenable raised $26M to sell what Microsoft stopped giving away

Price the round against its own price list. At the volume rate of $1 per 1,000 requests ^[7], the $26 million seed ^[1] buys about 26 billion requests ^[1], or roughly one lookup for every four documents in the corpus Keenable says it holds ^[2]. Crawling and serving are separate budgets, so the comparison is rough, but it puts a floor under the answer Styskin gave TechCrunch when asked what the index costs to run: "Don't ask - it is painfully expensive" ^[8].

The discount threshold does the customer segmentation. A buyer sustaining the qualifying 100 requests per second ^[7] is issuing 8.64 million requests a day, about $8,640 a day and roughly $3.2 million a year at that rate ^[7]. That is not an agent startup's line item. It is an AI lab's, which is what Keenable says is already running against the index in training and at runtime ^[3]. Everyone below that line pays four times as much per request ^[5].

The demand side is the more interesting mechanism. When Microsoft retired the Bing Search APIs in August 2025 and Google restricted access to Custom Search JSON ^[15], retrieval stopped being a checkbox on a developer account and became a metered dependency with a rate card. But the same platforms are supplying the replacement themselves: Web Search on Amazon Bedrock lets a developer get web results without onboarding an outside vendor at all ^[16]. The source material is candid that this is Keenable's problem as much as its opening, since large accounts may prefer search packaged by the cloud provider that already holds their models and their security review ^[17].

Which is why the roadmap matters more than the latency chart. Time Machine, in early access, searches historical versions of pages and pulls content from old snapshots ^[11], and the planned WebQueryLanguage is meant to answer across sources when no single page holds the answer ^[12]. Neither can be resold by a company that does not keep its own crawl archive. The July integration with Gradium's voice-agent framework points the same way ^[13]: a text agent can hide a slow fetch behind a progress indicator, a voice agent hands the user silence ^[14]. At the advertised p95 of 250 milliseconds in US East ^[5], four sequential lookups consume a full second ^[4], which is the real design constraint on that product, not the per-request price.

All of this is being built by 15 engineers, with a plan to reach about 30 by the end of 2026 alongside a go-to-market function [c9, d6]. Accel partner Zhenya Loginov led the deal, and neither the valuation nor the closing date was disclosed ^[10]. What he has funded is the proposition that a web index is a product in its own right rather than a feature of somebody else's model platform, sold to buyers whose agents issue repeated queries and read many documents before answering once ^[18].