A nine-person startup has closed a $12.5 million seed round to replace legacy search APIs with infrastructure purpose-built for AI agents, signaling where investors think the agentic stack is most fragile.
The pitch for Seltz is deceptively simple: Google and Bing were built for humans, not machines, and bolting an AI agent onto those APIs means inheriting every ad, every redirect, every piece of noise that makes those results useful in a browser and useless in a reasoning pipeline. Founded in 2025 with offices in Turin and the US, the nine-person company has spent the last year writing its own crawler, its own index, and its own retrieval models, all in Rust, from scratch. Queries come back in under 200 milliseconds with structured, metadata-enriched results stripped of the clutter. That's the product. The $12.5 million seed round, reported exclusively by Fortune, is the market's answer to whether it matters.
It does. The reason is where agentic AI is right now. Twelve months ago, the conversation was about reasoning models and orchestration frameworks. Today those pieces are more or less solved, or at least commoditized enough that the real bottleneck has shifted downstream to retrieval. An AI agent that can plan and execute multi-step tasks is only as good as the live web data it can pull in. Retrieval-augmented generation, the RAG layer, is the plumbing that connects the agent to the world. Most teams building production agentic workflows have discovered the hard way that existing search APIs were never designed for this use case: they return HTML-heavy responses at latencies that compound badly across multi-turn agent loops, with no structured metadata and no reliable signal about what's actually fresh.
Seltz's answer is to treat the machine as the primary user, not an afterthought. Its Web Knowledge API delivers results that are already structured and contextually enriched, optimized for what a language model needs to reason over rather than what a human needs to scan visually. The Rust stack is not an aesthetic choice. Rust gives Seltz the latency profile and memory safety that matter when you're serving high-volume, low-tolerance inference pipelines. The team has spent years building web search at scale; the crawler and index they wrote reflect that. This isn't a wrapper around an existing search service with a cleaner JSON schema on top. It's a ground-up rebuild with a different primary user in mind from line one.
That distinction matters more than it might sound. When you retrofit agents onto Google's API, you get results that are ranked for click-through rates, polluted with ads, and structured for rendering in a browser. Strip the HTML and you lose context. Keep it and you're sending noise through your context window at scale. Seltz returns results that are already clean, already enriched, already ready for the retrieval step, which means less preprocessing, fewer tokens consumed on garbage, and tighter latency across the whole agent loop. For teams running production workflows where a single user request might trigger dozens of web lookups, that compounds fast.
The company's selection for the AWS and NVIDIA Startup Village at VivaTech 2026 in Paris, held June 17 to 20, is worth noting for reasons beyond the obvious PR value. Seven European AI startups were chosen for that showcase, and the criteria lean heavily toward production-readiness. AWS and NVIDIA are not picking demos. They're picking companies whose infrastructure is already running on their stack at meaningful scale. For a nine-person team, that's a credible signal that Seltz is past the prototype stage.
Frankly, the seed size is the more interesting signal for anyone watching where smart money thinks the agentic stack actually breaks. $12.5 million at seed is not a modest bet. It reflects a view that purpose-built retrieval infrastructure is not a feature someone else will ship next quarter, it's a category, and getting there first with a genuinely differentiated technical stack is worth backing early and heavily. The investors apparently looked at the RAG layer and concluded it's structurally underserved. The companies building on top of it, the agent frameworks, the frontier labs, the enterprise workflow tools, they all hit the same ceiling when retrieval doesn't perform. That's the gap Seltz is filling.
The market Seltz is entering is heating up fast. As agentic workflows move from demos into production, the demand for reliable, low-latency, machine-readable web data is going from nice-to-have to critical path. Perplexity, You.com, and others have chipped at the edges of this problem from the consumer search direction. Seltz is coming at it from the infrastructure side, targeting the teams building the agents themselves rather than the end users. That's a narrower but arguably more defensible position, because enterprise AI teams buying infrastructure are stickier customers than consumers switching search engines.
Nine people, $12.5 million, and a crawler they wrote themselves in Rust. The agentic web is going to need a lot of retrieval infrastructure. Seltz is betting it should own the layer no one else bothered to rebuild.
Also read: Meta and Microsoft just pre-bought the AI future, and the landlords are the ones with real leverage • The British Army just proved AI can compress 72 hours of war planning into one, and the race to replicate it has begun • A San Jose legal-tech startup just sued the US government over losing Anthropic's Fable 5 and it won't be the last