Staan, the First European Search API, Opens Self-Service: A Deep Dive

On the afternoon of June 15, 2026, at the OVHcloud stand on the second day of VivaTech, Qwant opened self-service access to Staan, the search API it has been quietly running for itself and a small set of enterprise customers for the past year. The announcement post on X reads, in French first and English below, with the kind of careful escalation that this corner of the AI infrastructure stack has been waiting for. The post gathered 60,400 views, 454 likes, 82 reposts, and 321 bookmarks in the first day, an unusually strong signal for a developer-tools announcement from a consumer search engine that most English-language readers had, until now, mostly ignored.

@Qwant_FR - 15:14 · 15 juin 2026

🇪🇺 Staan, la première API de recherche européenne est désormais disponible en libre-service 🇪🇺

Staan est l'API de recherche 100% européenne co-construite avec @ecosia à travers la joint-venture européenne European Search Perspective. Staan connecte les applications d'IA au web en les alimentant avec des millions de documents et de contenus web dans un cadre juridique respectueux des données et des utilisateurs.

À partir d'aujourd'hui, Staan permet à n'importe quel développeur d'obtenir sa clé API et de commencer à alimenter ses assistants, copilotes et agents IA en quelques minutes.

Pendant que les solutions américaines restreignent leurs API de recherche et leurs modèles de LLM, une seconde voie européenne et souveraine s'organise. ✊

Switch to Europe. Now.

The interesting sentence is the one that the marketing copy has spent the last few weeks softening. "Pendant que les solutions américaines restreignent leurs API de recherche et leurs modèles de LLM" is a description of a market shift, not a slogan, and it does real work in explaining why this launch lands the way it does. The American retreat from programmatic web access is not a hypothetical, and it is not a Twitter narrative. It is the result of two deliberate business decisions in 2025 that, between them, removed the two largest sources of search results that AI applications had been quietly building on.

What this actually is, and what it is not

The framing correction is worth making up front, because most English-language coverage will describe Staan as a European search engine launch, which it is not. Staan is a real-time web search API, in the same product category as Tavily, Exa, Brave Search, SerpApi, and the search layer inside Microsoft Azure AI Foundry. It does not have a consumer front end. It does not compete with Google for end users. It does not, in any direct sense, compete with Qwant the consumer engine either, because Qwant is one of its customers, not its adversary. What Staan competes for is the layer in the AI stack that connects a language model to the live web, the part that decides whether a citation in a chatbot answer points to a real, fresh source or to a hallucinated URL that does not exist.

That layer matters more in 2026 than it did in 2024, and the reason has very little to do with Staan itself. The layer matters because the supply of search APIs that AI applications can call, behind the scenes and at scale, has contracted. Microsoft retired the Bing Search API in August 2025, eight years after the original Bing Search API first shipped, and redirected developers to a different product, Grounding with Bing Search, that ships as part of Azure AI Foundry rather than as a standalone HTTP endpoint. Google never sold a comparable standalone web search API at scale, and the limited programmatic access it does offer is constrained by quota, partner status, and Google's own product surface area. The third major US source, the open-web index used by academic and hobbyist tools, has been litigated into increasingly narrow terms by the New York Times v. OpenAI line of cases, the UMG v. Anthropic settlement, and a steady drumbeat of publisher-side enforcement actions that have made the safest answer 'do not scrape unless you have a deal.' The result is a vacuum at exactly the layer that AI applications need to function, and the vacuum has been widening for a year.

The Qwant announcement lands into that vacuum with timing that is hard to call accidental, and with a positioning that is unusually direct. Staan is not the first European search engine; Qwant, Ecosia, Mojeek, and a handful of smaller players have been at that for years. It is the first time a non-US company has exposed a production-grade web search index to the global AI developer market at the price points and the latency envelope that the AI developer market has been trained to expect. The list of paying customers that Qwant has named so far, CamoCopy, OVHcloud, Infomaniak, xPrivo, Pappers, and the two founding consumer engines, is short but not unimpressive. OVHcloud and Infomaniak in particular are the two largest European cloud and hosting providers that are not AWS, not Azure, and not Google Cloud, which is the right customer set for a product whose main pitch is 'we are not them.'

The pipeline, in the order a query sees it

The most useful way to understand Staan is to walk through the path a single query takes, because the order is the product. A search API for AI applications is not the same product as a search API for human users, and the differences show up in the middle of the pipeline, not at the edges.

The first stage is the search pipeline itself. A query comes in over HTTPS, hits a load balancer, and is routed to a search cluster that runs a stack that the FAQ describes, in a useful understatement, as 'a search pipeline.' The underlying engine is not public, but the behaviour is: a query returns a list of up to ten results, each with a title, a URL, a snippet, and a few metadata fields. The result count is fixed at ten per call, with a higher limit on the roadmap. The company does not currently expose a paginated cursor, which means that a workflow that wants more than ten documents has to issue multiple parallel calls, not multiple sequential calls, and that is a real constraint for any agent that wants to summarise the top fifty results for a query.

The second stage is the one that justifies the price difference between the two lower tiers. Web Search returns the raw search results and stops there. Web Search for AI takes those same results, fetches the actual page content for each one, extracts targeted excerpts, and reranks the excerpts by relevance to the original query. The default is three extra snippets per result, configurable from one to ten, and the reranker is the part that turns a list of ten URLs into a list of ten passages that an LLM can actually use. The pipeline is described in the FAQ as a fetch, then an extraction, then a reranking pass, which is the same shape that Tavily and Exa use internally, and the description lines up with what one would expect from a 2025-era AI search product.

The third stage is the AI Answer tier, which is a different product entirely. Instead of returning a list of results and letting the caller pass them to whatever LLM the caller is already using, AI Answer takes the same query, runs the search pipeline, runs the fetch, extraction, and reranking pass, and then calls a language model to produce a single answer with inline citations. The language model is Mistral Small, hosted on the same OVHcloud infrastructure, and the citations are JSON objects that the caller can render, audit, or strip before showing the answer to an end user. The whole thing is one HTTP call, and the answer comes back in either markdown or plain text, with optional related queries and an optional conversational context for follow-up turns. The choice of Mistral Small is a small but real data point in the European sovereignty story: the same Mistral, in a different week, was at the centre of a fake launch that fooled the timeline for 24 hours, and the fact that Staan quietly relies on the real Mistral Small for production answer generation is the kind of detail that makes the 'European stack' pitch more than marketing.

The latency budget is the easiest part to verify. The marketing page says sub-800ms P95 for the Web Search tier, and the FAQ confirms a target rather than a contractual SLA. For the AI Answer tier, latency is dominated by the Mistral Small call, and the company has not published a P95 number. For comparison, Brave Search's published P95 is around 1.2 seconds for the AI tier, and Tavily's is around 1.5 to 3 seconds depending on the operation, so Staan is in the right neighbourhood for the cheaper tier and unmeasured for the expensive one. The 800ms figure, if it holds up under independent benchmarking, is the most consequential number on the page, because it puts Staan inside the budget of synchronous chat experiences without requiring a streaming answer to feel responsive.

The product matrix, in one table

Two details that the marketing page buries are worth surfacing. First, domain whitelisting and blacklisting are supported on the Web Search and Web Search for AI tiers, but the current console lets you restrict to one domain only, with multi-domain whitelisting scheduled for Q3 2026. Second, credit card payment is not on the self-service console yet, which means the 'self-service' label is qualified: developers can sign up, get an API key, and start calling the API, but the billing layer still appears to be a sales conversation. The company has flagged both as roadmap items, and the FAQ describes them as such.

The third detail, the one that is going to come up in every enterprise procurement review, is the absence of a published SLA, the absence of a SecNumCloud qualification, and the absence of a public certification roadmap. Staan is hosted on OVHcloud, which is its own kind of answer for the European public-sector and regulated-industry buyers who care about the difference between 'hosted in Europe' and 'qualified for European public-sector workloads.' The qualification is a separate problem, and the FAQ is candid that Staan is not SecNumCloud-qualified at this time.

The 'American backup,' and what it does to the sovereignty story

The most interesting part of the Staan story, and the part that the marketing copy has been working the hardest to soften, is the 'American backup mechanism' that the FAQ discloses on its own initiative. The language in the FAQ is, in full:

It is an American backup mechanism that steps in to supplement certain results only when necessary. This coverage evolves empirically, in line with the growth of our own index. Queries transmitted to the fallback solution are not retained by it. No data is stored in the United States or reused by the fallback solution. The fallback can be fully disabled on request, with the understanding that the quality of some results may be diminished as a result.

This is, to put it plainly, an admission that the index is not yet at parity with the US incumbents, and a candid description of the workaround. The fallback is a US-based index, the partner is not named, and the company is open about the fact that the data does leave Europe for those queries, even if it does not get stored on the other side. The position is defensible: no company with a two-year-old index is at parity with Google on long-tail coverage, and the worst outcome for a self-service launch is the answer 'I asked it something niche and got nothing.' The position is also, in the sovereignty framing that the announcement post leans on, a meaningful caveat. A 'souveraine' API that, on long-tail queries, falls back to a US engine, is sovereign most of the time and not sovereign on the queries where the index has not yet caught up. Buyers whose threat model is full data residency will turn the fallback off and accept the coverage gap. Buyers whose threat model is 'I want a European option for most things, and I am fine with the US filling in the rest' will leave it on.

The candid disclosure is, in its own way, a sign of the company's seriousness. Tavily, Exa, and Brave Search all describe their coverage in confident terms without naming the upstream sources they fall back to when their own index is thin. Staan is naming the fallback as a fallback, documenting the privacy properties of the fallback, and offering to disable it on request. That is a posture that enterprise procurement teams can build on, and it is also a posture that competitors will have to match or explain why they have not.

The American retreat, in two specific decisions

The Qwant announcement post frames the launch as a response to 'American solutions restricting their search APIs and LLM models.' The framing is correct, and the two specific decisions it is pointing to are both worth describing in detail, because the gap they have opened is exactly the gap that Staan is built to fill.

The first is Microsoft's August 2025 retirement of the Bing Search API, the long-running standalone endpoint that AI developers had been calling since 2017. The deprecation was announced in May 2025 and completed in August, and the migration path that Microsoft pointed to was Grounding with Bing Search, a feature of Azure AI Foundry that bundles the search layer with the rest of the Azure AI stack, including content safety, prompt caching, and Azure-specific identity. The effect, in practice, was that the cheapest way to get a fresh web signal into a custom AI workflow, which had been a $7 per 1,000 calls HTTP call, was now a $30 to $50 per 1,000 calls bundled service that also required an Azure account. For AI developers who were not already on Azure, the migration was a real cost. For AI developers who were on Azure, the new offering was a real product, but it was no longer a drop-in search API. Microsoft did not break anything, but it raised the floor on what it costs to call Bing from an AI surface, and the timing happened to coincide with the moment that AI surfaces were getting the most usage.

The second is the steady contraction of Google's programmatic web access. Google has never sold a standalone web search API at the scale that Bing did, and the access it has offered, through the Custom Search JSON API, the Programmable Search Engine, and a handful of partner-only deals, has been constrained by quota, by partner status, and by Google's preference for keeping the search layer close to its own product surfaces. In 2025, the constraint tightened further as Google shifted partner-only access toward the Gemini API surface, where grounding is bundled with the rest of the Gemini offering. The effect was the same as Microsoft's, but smaller in absolute terms, and the relevant comparison is that a Google-grounded AI workflow now requires both a Gemini relationship and a partner-only deal, which is a higher bar than the Custom Search API that a hobbyist could call from a laptop.

The vacuum that those two decisions opened is what Staan, and Tavily, and Exa, and Brave Search, and a handful of smaller players, are all competing to fill. Staan is the only one of the group whose pitch is 'we are not American, and our index is not a wrapper around an American index.' The pitch is mostly right, with the caveat above. The timing also coincides, not by accident, with the broader contraction in US-hosted AI infrastructure that has played out over the last three weeks, from the June 12 Fable 5 export control order that suspended Anthropic's strongest commercial model for almost every non-US user, to the deeper legal mechanics of why the order targeted a class of person rather than a class of capability. Staan is not a response to Fable, but the two stories land in the same week because they share the same underlying pressure: the supply of US-hosted AI infrastructure that non-US builders can rely on is shrinking, and the alternatives are suddenly worth their integration cost. The index that powers Staan is the same index that powers Qwant, Ecosia, and Lilo, and the EUSP governance is 50/50 Qwant and Ecosia with the operational headquarters in Paris. The pipeline is hosted on OVHcloud, the Answer service runs on Mistral Small, the SDK is in TypeScript, and the data is in France.

The comparison that matters, in one paragraph

On raw web search, the three credible US-hosted incumbents are Tavily (search plus extraction, oriented around RAG, around $1 per 1,000 basic credits, more for advanced features), Exa (neural and embedding-based search, around $5 per 1,000 neural results), and Brave Search (Google-quality web index, $3 per 1,000 calls, $5 for the AI tier). The two credible US-hosted scraper-and-render services are SerpApi (Google SERP scraping, around $50 per 1,000 for the standard plan) and Linkup (which sits in the same category as Tavily with a more European-leaning team and a slightly different feature mix). Staan is cheaper on raw search (€1 vs $1 to $5 per 1,000), is the only one of the group with European data residency by default, and is the only one whose index is independently owned and operated by a non-US company. Staan is also the only one in this group whose index has a publicly verifiable consumption footprint: it is the index behind the Qwant, Ecosia, and Lilo search engines, which together handle a meaningful slice of EU consumer search, on the order of a low single-digit percentage of EU queries by some industry estimates. The trade-offs are the smaller language coverage (FR, EN, DE, no plans shared for others), the no-SLA posture at launch, the credit-card payment gap on the self-service console, and the American fallback on long-tail queries that the index has not yet covered.

The TypeScript angle, for the readers who got here from a Claude Code tab

For a TypeScript or JavaScript team that is wiring an AI assistant or an agent loop to a fresh web signal, the practical path to Staan is short. The Answer service has a first-party SDK at github.com/Qwant/answer-node, the surface covers streaming and non-streamed responses, the response format is JSON, and the licence is permissive. The Web Search and Web Search for AI tiers are single HTTP endpoints that return JSON, and the same fetch call a team would use to talk to Tavily or Brave Search works against Staan with no special handling. The credit card gap means that the first €1 to €2 of usage has to be coordinated with a sales conversation, which is a real friction for a one-person side project but not for a team that is already doing a vendor onboarding. The bigger practical question is whether the index is good enough for the workflow the team has in mind, and the right way to answer that question is to run a benchmark on the actual use case, not to read the marketing page.

The second practical question is whether to disable the American fallback, and the right answer depends on what the team is doing. For an internal RAG pipeline over public web content, the fallback is fine to leave on, and the privacy posture is the one the FAQ describes. For a pipeline that is under a formal data processing agreement with a European enterprise customer, the fallback should be turned off, the coverage gap should be measured, and the team should plan to either live with the gap or fill it with a second source. For a pipeline that is shipping to a public-sector buyer in a country with residency requirements, the fallback is a non-starter, and the procurement conversation needs to be with EUSP directly, not with the self-service console.

What to watch

For the next two weeks, the signals that matter are: whether the EUSP team publishes a public API status page and the first batch of independent benchmarks; whether a non-European AI coding tool (a Claude Code competitor, a Cursor rival, a Continue fork) ships Staan as a backend; whether Qwant publishes a self-service credit card flow and a public SDK for the Web Search tier; whether Mistral publishes an update on Mistral Small that has a measurable latency impact on the Answer service; whether the 'no SLA, no SecNumCloud' gap is closed or explicitly punted to a future roadmap; whether a US-based competitor publishes a comparable 'European data residency' tier in response; and whether the American fallback gets a name, a public status page of its own, or a sunset plan as the EUSP index grows. The first 72 hours will tell us whether the launch is a real product or a marketing event. The first 30 days will tell us whether the index is good enough to keep the launchers who try it.

For a TypeScript or JavaScript team that is already routing an AI surface to Tavily, Exa, Brave, or a Bing-shaped vacuum, the practical move is to spend an afternoon swapping in Staan on a side workflow, measure the answer quality on the queries that matter, and decide whether the €1 to €2 per 1,000 calls is worth the European data residency and the missing credit card flow. For a team that is not yet routing to any search API, the move is to start with Web Search at the €1 tier, run a few hundred real queries, and only step up to Web Search for AI or AI Answer once the basic coverage is good enough to justify the price.

The story is not over. The interesting question is not whether a European company can build a search API; the Qwant and Ecosia teams have been doing that for years. The interesting question is whether a European company can build a search API that an AI agent, running on a US-hosted language model, will pick up by default. The answer, in the first 24 hours of self-service, looks more like 'yes, for the queries that matter' than 'no, but it's a nice try.' That is a different answer than the one that would have been true in 2024, and it is a different answer than the one the American API retreat would suggest if the retreat had been the only story.