Seltz has raised $12.5 million because AI agents have a boring but serious problem: they still have to search a web built for people.
Seltz is making a direct bet on one weak part of the agentic AI stack. Google and Bing were built for humans sitting in browsers, not for software agents making repeated calls inside a reasoning loop. If you plug an agent into old search infrastructure, you inherit ads, redirects, HTML clutter and results shaped for clicks. That is fine when a person can skim and ignore the junk. It is expensive when a model has to read it.
According to Fortune, Seltz has closed a $12.5 million seed round to build a search layer for AI agents. The company was founded in 2025, has offices in Turin and the US, and has nine people. Its pitch is not that it has wrapped an existing search API in cleaner packaging. Seltz says it has built its own crawler, index and retrieval models in Rust, with queries returning in under 200 milliseconds and results delivered as structured, metadata-enriched data.
That detail matters. An agent does not need a blue link page. It needs fresh, clean, machine-readable context it can pass into the next step without burning half its context window on navigation bars and malformed snippets. If your agent has to do dozens of lookups to answer one customer request, small delays stop being small. Bad retrieval becomes the tax you pay on every loop.
The last year of AI has made that problem harder to ignore. Reasoning models have improved. Orchestration frameworks are everywhere. You can find a tool for planning, memory, evaluation, workflow routing, you name it. But the agent still has to connect to the live web, and that is where a lot of impressive demos start to wobble. Retrieval-augmented generation is not glamorous infrastructure, but it is the pipe between the model and reality.
Seltz is trying to own that pipe. Its Web Knowledge API is designed around the machine as the first user, with structured results and contextual metadata prepared for model consumption. Rust is doing real work here, at least in the company’s telling, because low-latency retrieval is not a nice extra when your customer is running production inference systems. The promise is less cleanup, fewer wasted tokens and a tighter loop between query, retrieval and reasoning.
Frankly, that is a better seed-stage story than another thin agent wrapper with a dashboard on top. The agent market already has enough companies selling interfaces. The harder question is whether the underlying data layer can keep up when those agents leave demos and start handling live work. A nine-person team raising $12.5 million says investors think that question has a price tag.
The VivaTech detail helps the story, but only if you read it correctly. Seltz was selected for the AWS and NVIDIA Startup Village at VivaTech 2026 in Paris, held June 17 to 20, as one of seven European AI startups in that showcase. That does not prove the company will win the retrieval market. It does suggest the startup is being taken seriously by the cloud and chip companies that need agentic workloads to turn into real compute demand.
There is competition around the edges. Perplexity and You.com have pushed search toward AI-native answers from the consumer side. Seltz is coming from the infrastructure side, selling to the teams building agents rather than the people typing questions into a search box. That is a narrower market, but it may be a more durable one. Enterprise AI teams do not switch core retrieval infrastructure as casually as consumers switch search tabs.
The risk is obvious. Search is brutally hard, and building a crawler and index from scratch is not the same thing as making them reliable across the messy, changing web. Google and Microsoft did not become search giants because indexing pages was simple. Seltz has to prove that machine-first retrieval can be good enough, fast enough and fresh enough to justify being a separate layer in the stack.
Still, the shape of the bet is clear: nine people, a Rust-built crawler, a sub-200 millisecond retrieval target and $12.5 million to make AI agents less dependent on search tools built for a different era. If agents are going to do useful work on the live web, they need better inputs. Seltz is betting that the input layer is where the next fight starts.
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
