Web index infrastructure is not the sort of thing that normally makes headlines. On Tuesday 25 August 2026 it did. Keenable came out of stealth with $26 million in seed funding led by Accel, and with a claim that it has already crawled and organised more than 100 billion documents — not for people, but for machines.
The pitch is narrow and unusually clear. Ranking the web has been a natural language processing problem for three decades, but the reader at the other end of it has always been a person. Every search engine you have ever used was designed around a human being with ten blue links and a short attention span. A large language model has neither. It can absorb far more text than a person will ever read, it cannot click, and it needs source documents it can quote. Keenable’s argument is that serving that consumer properly requires a different kind of web index, built from scratch, with different economics and different ranking assumptions.
“This actually creates a new flywheel that is different from what Google learned from human behavior,” co-founder and chief executive Andrey Styskin told TechCrunch. Styskin previously ran search, artificial intelligence and cloud at Yandex, then worked on web search infrastructure for AI applications such as Alexa at Amazon. His co-founder and chief scientist, Matthias Petri, is a German AI scientist who worked alongside him there.
This article breaks down what Keenable has actually built, what the funding buys, how an agent-facing web index differs from consumer search, who else is fighting for the same ground, what the company has pointedly not disclosed, and what any of it means if you are building or buying AI systems today.
Table of contents
- Why an Agent-Ready Web Index Just Raised $26 Million
- What Keenable’s Web Index Actually Is
- Why a Web Index for Machines Is Not a Search Engine
- The Web Index Numbers Keenable Has Published
- The Web Index Land Grab: Exa, Parallel, Tavily and Brave
- What Keenable Has Not Said About Its Web Index
- What an Agent-First Web Index Means for Your Business
- Web Index FAQ: Your Questions Answered
- References
Why an Agent-Ready Web Index Just Raised $26 Million
The round is small by 2026 standards and the ambition is not. Building a web index means crawling, storing and serving a meaningful fraction of the public internet, which is one of the most capital-intensive things a software company can attempt.
The round, the backers and the timing
Accel led the $26 million seed, with participation from Conviction Partners and a group of business angels. Accel partner Zhenya Loginov led the investment. Keenable exited stealth on the same day the round was reported, which is a deliberate choice: the product was already live and in production before the market knew the company existed.
Seed rounds of this size usually fund a product search. This one funds infrastructure that already exists. That inversion is the most interesting thing about the deal.
What Accel says the gap is
Loginov’s stated rationale is scarcity. AI companies have “very few options when it comes to web-scale search infrastructure,” he told TechCrunch, particularly because Google and Microsoft have “shut down their existing search APIs to avoid cannibalization.”
That is the whole thesis in one sentence. The two organisations that already own a planet-scale web index have commercial reasons not to rent it to the companies most likely to replace them. Everyone else has to build or buy.
| Detail | What was disclosed |
|---|---|
| Round | $26 million seed |
| Lead investor | Accel, deal led by partner Zhenya Loginov |
| Other participants | Conviction Partners, business angels |
| Valuation | Not disclosed |
| Documents indexed | More than 100 billion |
| Engineering team | 15 people across the US and Europe |
| Hiring plan | Roughly double headcount by end of 2026 |
What Keenable's Web Index Actually Is
Strip away the positioning and Keenable sells three things: a retrieval API, a set of client integrations, and the index underneath both.
One hundred billion documents behind one API
The core product is a search API. An agent sends a query, Keenable returns ranked results from its own web index, and a companion endpoint fetches the full content of a page so the model can quote it directly. That second step matters more than it sounds, because grounding an answer in a retrieved document is what separates a citation from a hallucination.
Keenable states the web index covers over 100 billion documents. That figure is the company’s own and has not been independently audited, which is normal for this category and worth remembering anyway.
The MCP server, the CLI and Web Query Language
Keenable ships an MCP server so that agent clients — Claude Code, Cursor and similar tools — can call the web index through the Model Context Protocol without bespoke glue code. A Rust command-line tool handles configuration across those clients, and the company’s public GitHub organisation carries dozens of repositories including integrations for automation platforms and a live retrieval benchmark.
An unreleased feature called Web Query Language is the more ambitious part. Keenable describes it as a way to answer a question by combining information from several web sources even when no single page contains the whole answer. That is synthesis over a web index rather than lookup against one.
Where the web index already runs
Keenable says its API is in production at several AI labs and inference providers, used both during training and at runtime. It also has a named partner: the Paris voice-AI company Gradium, which raised a $100 million seed backed by Nvidia in July 2026, has wired Keenable retrieval into its open-source voice-agent framework so that spoken answers can be grounded in live web results.
What it deliberately is not
Keenable is not building a chat product, an agent framework or a model. It sells the layer underneath all three, which keeps it out of competition with its own customers — a positioning choice that matters when the buyers are AI labs. The company is selling picks and shovels into a gold rush, and the picks happen to be a web index.
Why a Web Index for Machines Is Not a Search Engine
The technical case rests on the observation that human search and machine search optimise for different things, and that retrofitting one into the other wastes most of the advantage.
People skim, models read
Consumer search evolved to put the single best answer in position one, because a person will not read page two. Search engines were, in Styskin’s framing, built and optimised for people. An AI system has the opposite constraint: it can read and process much larger portions of information, and it performs better when it can ground its responses with source documents.
So the shape of a good result set changes. Ten links is the wrong output. Two hundred passages with clean text, provenance and enough coverage to reason across is the right one, and that is what an agent-facing web index is tuned to return.
A different flywheel
Google’s ranking quality compounds from decades of human click behaviour. Keenable cannot buy that, and Styskin’s claim is that it does not need to, because the feedback loop from AI usage is a different signal entirely. Whether a machine-generated flywheel can reach the same quality bar is the open technical question in this whole category, and nobody has answered it yet.
Narrowing the search space fast
The hard part is cost. “If you do not fine-tune your index structures for a specific task, the cost of serving and scanning the whole internet is enormous because of the volume,” Styskin said. “That’s why you need to innovate on how you can narrow the search space based on your query very fast. This is what we are bringing to the table.”
Asked what building the web index had cost, he was blunter: “Don’t ask — it is painfully expensive.” That is a founder naming his own biggest risk, which is more useful than a spin-free press release.
The Web Index Numbers Keenable Has Published
Very little about Keenable’s economics is public, but the pricing page is, and prices reveal more about a business than most funding announcements do.
| Tier | Price per 1,000 requests | Who it is for |
|---|---|---|
| Free | $0, capped at 100,000 requests a month | Evaluation and prototypes |
| Pay as you go | $4 | Production apps with variable volume |
| Enterprise | $1, at 100+ requests per second | Labs and inference providers at scale |
What the pricing implies
A four-to-one gap between the pay-as-you-go rate and the committed rate is a volume business telling you where it wants to live. At $4 per 1,000 requests, a million agent queries a month costs $4,000. At the enterprise rate, the same million costs $1,000.
A free allowance of 100,000 requests a month is also unusually generous for infrastructure this expensive to run, and it only makes sense if the company expects developers to build against the web index and then grow into a contract.
Latency, and who measured it
Keenable’s partner Gradium states that Keenable returns web retrieval in under 200 milliseconds. That is a partner-reported number rather than an independent benchmark, and it is the only latency figure with a name attached to it.
The distinction matters for a voice product. A spoken agent that pauses for a second before answering feels broken, so sub-200-millisecond retrieval is the difference between a usable voice assistant and a demo.
The team behind the index
Fifteen engineers across the United States and Europe built and now run this, with a plan to roughly double that by the end of 2026. For context, a full web index at Google scale is maintained by an organisation orders of magnitude larger.
That comparison cuts both ways. It is either evidence that modern infrastructure makes a small team viable, or evidence that 100 billion documents is a very different thing from the whole web.
The Web Index Land Grab: Exa, Parallel, Tavily and Brave
Keenable is not early. It is arriving into a category that has already seen large rounds and at least one exit, and its $26 million is the smallest number in the room.
Who else sells retrieval to agents
Exa raised a $250 million Series C at a $2.2 billion valuation in May 2026, led by Andreessen Horowitz, and says it serves more than 5,000 customers. Parallel Web Systems, founded by former Twitter chief executive Parag Agrawal, raised a $100 million Series B at a $2 billion valuation led by Sequoia in April 2026, five months after a $100 million Series A at $740 million.
Tavily took the other exit. Nebius agreed to buy it in February 2026 for $275 million in initial cash, rising to as much as $400 million on performance milestones, after Tavily had raised roughly $25 million in total. Brave sells a search API as an alternative data source, and Firecrawl and Linkup compete on the crawling and factuality edges of the same problem.
| Company | Latest capital event | Valuation or price | Backer |
|---|---|---|---|
| Keenable | $26m seed, Aug 2026 | Not disclosed | Accel |
| Exa | $250m Series C, May 2026 | $2.2 billion | Andreessen Horowitz |
| Parallel Web Systems | $100m Series B, Apr 2026 | $2 billion | Sequoia |
| Tavily | Acquired by Nebius, Feb 2026 | $275m, up to $400m | Nebius |
Set against those four capital events, the scale of Keenable’s raise is easier to read.
Money is not the moat
Being outspent ten to one is survivable if the underlying asset is better, and Keenable’s bet is that owning the crawl and the ranking beats renting someone else’s results. Several competitors in this space are wrappers over existing search providers; a first-party web index is a genuinely different cost base, and a genuinely harder thing to fund.
The Yandex thread running through the sector
There is a neat coincidence worth noting. Styskin ran search at Yandex. Nebius, which bought Tavily, was itself spun out of Yandex in 2024. The people who built one of the few non-Google planet-scale search systems are now spread across the companies trying to build the next one, which tells you how narrow the talent pool for this work really is.
What Keenable Has Not Said About Its Web Index
A same-day funding story is not scrutiny. Several of the questions that decide whether this business works are simply unanswered.
No named customers
“Several AI labs and inference providers” is not a customer list. Nobody outside the company can confirm the volume, the retention, or whether those users are paying production customers or evaluating for free.
Nothing on robots.txt, licensing or publisher opt-outs
This is the loudest silence. A company that crawls the open internet at 100 billion documents sits directly in the middle of the fight between AI firms and publishers, and neither the funding coverage nor Keenable’s own site addresses how it handles robots.txt directives, publisher opt-outs or content licensing.
That is not evidence of bad behaviour. It is an unanswered question that any enterprise buyer will have to raise before routing production traffic through a third-party web index, and one that could become a commercial constraint rather than a legal one.
Freshness is unaddressed
Retrieval is only useful if the page you get back reflects the page as it is now. Recrawl frequency, index staleness and how quickly new content becomes visible are the operational numbers that separate a good web index from a stale archive, and none have been published.
Coverage is not the same as size
One hundred billion documents is a large number in isolation and a meaningless one without shape. Which languages, which regions, which content types, how much of it is duplicate or spam, and how deep the crawl goes on any given domain all decide whether a web index is useful for your queries. A general figure tells you the storage bill, not the coverage.
No independent scrutiny yet
The company launched publicly the same day the funding was reported, so there is no third-party benchmark, no comparative evaluation and no critical coverage. Absence of criticism at this stage means nothing either way.
What an Agent-First Web Index Means for Your Business
Most organisations will never buy retrieval infrastructure directly. It still shapes what your AI tools can do and what your own content is worth.
If you are building agents
You now have real alternatives to scraping the open web yourself, and the pricing is public enough to model. Before committing, run your own evaluation on queries from your domain rather than trusting a vendor benchmark, and check whether the provider supports the Model Context Protocol so you are not writing throwaway integration code.
If you publish content
Your pages are training and grounding material for systems that may never send you a visitor. Understanding how machines parse, retrieve and cite your site is becoming a distinct discipline from classic ranking work — which is exactly what answer engine optimisation and generative engine optimisation exist to address.
If you buy AI services
Ask your vendors where their grounding data comes from. A retrieval-augmented assistant is only as current and as defensible as the index behind it, and “we use a third-party search API” is a materially different answer from “we crawl it ourselves” when a compliance team asks.
The questions worth asking
Which index grounds this product? How fresh is it? What happens to answer quality if that provider changes its pricing or its terms? If you are scoping autonomous AI agents for real work, retrieval quality is usually the constraint that decides whether the project succeeds, not the choice of model. Our AI models and tools hub tracks how these components keep shifting.
Do not re-architect around a seed-stage vendor
Keenable is fifteen engineers with a product that went public today. Evaluate it, benchmark it, keep your retrieval layer swappable, and do not build a dependency you cannot unwind if the company is acquired or repriced.
Web Index FAQ: Your Questions Answered
What is Keenable?
Keenable is an AI infrastructure company that has built its own web index and sells access to it through a search API, an MCP server and a command-line tool. It exited stealth on 25 August 2026.
How much did it raise, and who led the round?
Keenable raised a $26 million seed round led by Accel, with the deal led by Accel partner Zhenya Loginov. Conviction Partners and business angels also took part. The valuation was not disclosed.
How big is the web index?
The company states it covers more than 100 billion documents. That figure is self-reported and has not been independently verified.
What does it cost?
Keenable’s published pricing is a free tier of 100,000 requests a month, $4 per 1,000 requests on pay as you go, and $1 per 1,000 requests on an enterprise tier at 100 or more requests per second.
How is this different from Google Search?
Google ranks results for a person choosing one link. An agent-facing web index returns large volumes of clean, quotable source text for a model to read and cite, and is tuned for machine consumption rather than human browsing.
Who are the main competitors?
Exa, Parallel Web Systems, Tavily under Nebius ownership, Brave’s search API, Firecrawl and Linkup all sell retrieval to AI systems. Exa and Parallel are both valued at over $2 billion.
Should my business care about this yet?
Only indirectly. Treat it as a signal that grounding infrastructure is consolidating into a paid layer, and start asking your AI vendors which web index sits behind the answers they give you.
References
Accel-backed Keenable is indexing the web for AI agents — TechCrunch
Keenable — web search infrastructure for AI
Keenable API and MCP documentation
Keenable open-source repositories on GitHub
Gradium and Keenable partnership announcement — Gradium
Nebius announces agreement to acquire Tavily — Nebius
Announcing our Series C — Exa
Parallel Web Systems hits $2B valuation — TechCrunch
Parallel raises at $2 billion valuation to scale web infrastructure for agents — PR Newswire
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.