Google and Microsoft Are Shutting Down Their Search APIs — A $26M Startup Is Filling the Gap for AI Agents

(Keenable Learn) /rename Keenable Learn 8/25/26 10:35am

A $26 million seed round announced this week signals how fast the market for web search infrastructure is shifting: Accel-backed startup Keenable, co-founded by a former Yandex search engineering lead and a German AI researcher, is building a 100-billion-document web index designed specifically for AI agents rather than human browsers. The timing is pointed — Microsoft formally decommissioned its Bing Search APIs on August 11, 2025, and Google has been pulling back its own third-party search access in parallel, leaving thousands of AI applications scrambling for a replacement.

What Changed

The Microsoft shutdown was abrupt by enterprise standards. An official Microsoft Lifecycle notice confirmed that Bing Search APIs would be "decommissioned completely" on August 11, 2025, with no new customer signups permitted in the months leading up to that date. Bing Work Search went first, shutting down on March 31, 2025, followed by the broader Bing Search API and Bing Custom Search stack a few months later. Microsoft's suggested path forward — "Grounding with Bing Search" inside Azure AI Agents — is a replacement for Microsoft's own ecosystem, not a neutral data pipe. For developers building products outside that ecosystem, it is effectively a dead end.

Google has followed a similar trajectory, tightening access to its search index for third-party API consumers. Neither company has publicly framed these decisions as hostile to developers — instead, both have positioned the changes as a consolidation of search into their own AI products. But the practical effect, documented by multiple sources, is the same: the two companies that indexed most of the open web are no longer willing to sell that index access to competitors or third parties.

Into that gap, Keenable arrives with a credible founding story. The ex-Yandex background matters here — Yandex built and maintained a genuinely independent, large-scale search index for years, and the engineering discipline behind that work is directly applicable to what Keenable is attempting. The $26 million seed round from Accel, a firm with a track record in developer infrastructure, gives the company enough runway to reach the scale where a purpose-built index starts to justify its own cost structure.

According to TechCrunch's reporting, Keenable's API is already live with several undisclosed AI labs and inference providers — which means this is not a pre-revenue announcement but an early commercial deployment.

How It Works

The founding insight, reportedly anchored in Cloudflare traffic data, is that AI crawlers now represent a significant and growing share of total web traffic. The web is increasingly being read by machines, not people, and the query patterns machines issue — short, high-frequency, context-dependent bursts from agent loops — look nothing like the long-tail queries humans type into a search box. Existing indexes were built around the human pattern. Keenable's argument is that an index built around the machine pattern will be meaningfully better and cheaper to serve at scale.

The practical implication is what the company calls a WebQueryLanguage, still forthcoming, which is designed to let an agent retrieve fragments across multiple sources when no single page has the complete answer. This is a genuinely different problem from traditional search. A human typing into a search box is satisfied by a ranked list of links. An agent running inside a reasoning loop needs structured, composable excerpts it can feed into context — and it needs them fast enough that the latency doesn't compound across dozens of retrieval steps in a single task. The cost dimension is just as important: at web scale, the only way to make per-query costs manageable for a startup is to narrow the retrieval space efficiently before the expensive ranking step kicks in.

Runtime grounding — the practice of pulling live web content at inference time rather than relying solely on a model's training data — is increasingly treated as a baseline requirement for production-quality agents. A model trained on data from six months ago that cannot reach the current web will hallucinate or confidently state outdated facts. Keenable is betting that the demand for a clean, agent-optimized path to current web content will only increase as more developers run into that limitation.

What It Means for Developers

The honest picture here is that there is no single drop-in replacement for what Bing and Google provided. A comparison published by Parallel on August 23, 2026 — one of the more rigorous vendor-neutral analyses I have seen this year — splits the current market into three distinct categories: SERP APIs (which return links you still have to scrape), AI-native search APIs (which return pre-extracted content), and native LLM-provider tools (which are model-specific and non-portable). Choosing the wrong category is more expensive than choosing the wrong vendor within a category, because the architectural assumptions differ.

Within that landscape, Parallel's analysis names itself as the top-ranked option on the Artificial Analysis Search Index — though vendor self-ranking should always be treated skeptically. It reports response times of around 200 milliseconds on its Turbo tier and under one second on its Fast tier, at $1 per 1,000 requests. SerpAPI covers the broadest range of engines, including Google Scholar, Patents, and Shopping. OpenAI and Anthropic have both shipped native web search tools that avoid a separate retrieval pipeline entirely but lock developers into those specific model providers.

Keenable, Exa, and Brave occupy the agent-native search niche, which is the most directly relevant to teams building autonomous agents that need to read and synthesize web content rather than just retrieve links. Our read is that Keenable's 100-billion-document scale ambition distinguishes it from Exa — which has carved out a strong reputation for semantic search over a curated subset of the web — but the scale claim is only meaningful if the relevance and latency characteristics hold at query time. That remains to be independently verified.

For developers currently depending on any Bing Custom Search implementation: those pipelines are already broken as of August 2025. The LinkedIn discussion thread from Dhruv Ohri, circulating widely after the announcement, catalogued the affected product types clearly: custom GPT-based tools that relied on Bing for citations, real-time summarization bots, browser knowledge tools, and vertical search wrappers. Microsoft's Azure AI Agents path is viable if you are building inside the Azure ecosystem; it is a walled replacement, not a neutral pipe. If your stack is model-agnostic and you want to keep it that way, Keenable and its peers are now the infrastructure layer worth evaluating.

Sources

techcrunch.com Best Web Search APIs for AI Agents in 2026: Honest Comparison | Parallel Bing Search APIs Retiring on August 11, 2025 - Microsoft Lifecycle | Microsoft Learn Microsoft is killing Bing Search APIs for AI alternatives Bing Search API shut down: what it means for AI startups | Dhruv Ohri posted on the topic | LinkedIn