Google indexed the open web for humans: type a query, scan ten blue links, click through. That model is quietly aging out. The fastest-growing consumers of the web today are AI agents — and they don't read, click, or scroll the way people do.
That's the bet behind Keenable, a startup that came out of stealth this week with a $26 million seed round led by Accel (Conviction Partners and angels also joined). Founder Andrey Styskin spent nearly two decades at Yandex running search, AI, and cloud, then built web infrastructure for Amazon's Alexa; co-founder Matthias Petri did the same at Amazon AGI. Their stated goal: become "the next Google for AI agents."
Not a wrapper — a re-built stack
Keenable isn't wrapping Bing or Google APIs with a model. It rebuilds crawl, index, retrieval, and ranking from scratch, exposing them through a REST API, an MCP server, and a CLI. The index already covers more than 100 billion documents, with p95 latency under 250ms in US-East and pricing that drops to roughly $1 per thousand requests for high-volume customers (standard tier is $4; Exa's base price is around $7). The company says its API is already in production at several AI labs and inference providers, used both during training and at runtime, and it recently partnered with voice-AI company Gradium for live retrieval.
Why re-crawl the whole web? Because agents use it differently. A human search is one query and one click; a research agent runs a loop — search, fetch, read, spot a gap, re-query, cross-verify, then act — often tens to a hundred times for a single task. Live retrieval is exactly the class of infrastructure we benchmarked in our DeepSeek Harness live-search review. Styskin's point: if your index structures aren't tuned for the task, serving the entire internet gets painfully expensive, so the real engineering problem is narrowing the search space per query as fast as possible. The optimization targets flip from click-through rate to recall, freshness, latency, machine-readability, token efficiency, and provenance — ultimately, cost per completed task.
Google's flywheel doesn't spin for agents
Google's moat is a flywheel built on human behavior: what people click, how they rank pages, what satisfies them. Agents don't click. They read whole pages and execute — which is also why agent traffic has almost no ad monetization, a problem we examined in the zero-eyeballs ad dilemma for AI agents. The feedback loop Google learned from is largely irrelevant to a model that needs structured, verifiable context instead of a ranked list of links.
This is a textbook innovator's dilemma — and the incumbents are clearing the runway themselves. Both Google and Microsoft have been restricting or retiring their public search APIs to avoid cannibalizing their own products, leaving a vacuum at exactly the layer agents need. "It's extremely hard to convince people to leave Google for search," Styskin admits, "but on agentic queries, the giant is beatable."
Who owns the agent-to-web pipe
The economics sharpen the stakes. Even at $1 per thousand requests, a complex task that runs 100 searches costs only a tenth of a dollar in search alone — but production adds fetch, content processing, and reasoning tokens on top. The unit that matters is shifting from cost-per-search to search-cost-per-finished-task. That's why every provider in this lane — Alibaba's OpenSearch, Tencent's agent-search MCP, and the startups — is obsessing over token efficiency and query rewriting rather than just ranking.
Keenable isn't alone. Exa already runs a 100-billion-document vector index, raised $250 million at a $2.2 billion valuation, and counts Cursor, AWS, and Groq among its customers; Chinese entrants like Bocha and Xinliu are fighting for the same API slot. What they're all contesting is one position: the default interface between AI agents and the open internet — a layer that barely existed a year ago.
What this means if you build with agents
- Benchmark by task, not by request: compute search cost per completed workflow, and weigh token efficiency alongside latency when comparing providers.
- Prefer MCP-native services: your agents can swap search backends without rewriting orchestration logic. More agent infrastructure: our AI agent tool directory.
- Track freshness: agent workloads need up-to-date data far more than consumer search did — a stale index silently degrades agent output quality.
- Watch the API-shutdown trend: as Google and Microsoft pull back public search APIs, agent-native indexes turn from a nice-to-have into a strategic dependency.
Further reading: The Next Search Engine Is Built for Agents, Not Humans — a deep dive into the three structural changes behind agent search: optimization targets, cost units and value exchange.
