Know who wrote it,
before you trust it
An AI answer is only as good as what was retrieved for it. We index the people and organisations who originate knowledge, so retrieval can be judged by who stands behind it, not by what ranks.
Retrieval is the expensive layer, and it is the one nobody has fixed
Models are commoditising. Context is not. Every serious AI product now spends a large part of its budget and almost all of its latency on the same step: finding the right material to put in front of the model. Get that wrong and no amount of model capability saves the answer. This is the layer we work on.
A web search ranks documents. It finds pages about a subject and says nothing about who produced them, which leaves the hardest judgment to whoever reads the results. We start from the other end: who has authority on this subject, and what have they published. Same retrieval step, a different thing being ranked. PageRank was the last serious attempt at the question, and it was a good proxy for a web of documents built by people who cared enough to create hyperlinks. It was never designed for a web of creators, platforms and machine-written text, where social metrics measure popularity rather than expertise and a verification badge is something you can buy.
AI has made this urgent rather than merely interesting. Producing a convincing, plausible-sounding piece of writing or analysis now costs almost nothing, so the constraint has moved.
The constraint is no longer production. It is verification.
There is a feedback loop underneath it. Models train on web data, generate content that ends up on the web, and that content is scraped back into the next training run, so a model's sense of what a coral reef biologist sounds like drifts toward what AI thinks one sounds like. Retrieval compounds it: a RAG pipeline surfaces whatever ranks best, so an OSINT analyst's meticulous thread carries the same weight as a hallucinated summary of it. Every layer of the stack inherits the authority problem from the web below it.
We started with a different question
Not "what pages rank for this topic?" but "who actually has authority on this topic?" Those two questions sound similar, but they lead to completely different architectures. One starts from documents. The other starts from people.
The Authority Index is a graph of millions of ranked human authorities, continuously updated and queryable across hundreds of topic domains: mRNA researchers and maritime lawyers, OSINT analysts and coral reef biologists. We built it by mapping cross-platform signals that are genuinely hard to fake: sustained output over time, peer citation, consistent topical specificity, and presence across several independent platforms. A single viral post proves nothing; a pattern of substantive engagement in one domain, sustained over years, is much harder to manufacture.
Authority is a spectrum, not a binary credential, and we treat it that way: someone can be a tier-one authority on wastewater treatment policy and a complete non-expert on the electoral system even if they post about both. We publish the results, not the recipe. The principle is simple: we measure real expertise, not reputation proxies.
That decision has a consequence worth stating plainly, because it is the whole shape of the company. We do not try to hold the entire web. We hold the few per cent of it that originates what the rest repeats, the long tail of real practitioners included. It is a chosen slice rather than a sample, which is why it still answers the kind of question you would otherwise take to a search engine, and why holding it costs a fraction of what holding everything costs.
The author is the ranking signal, not the publication
Ask a search API what the latest is on a subject and it hands you the outlets: the BBC, the FT, a trade publication. Usually a reasonable answer, and it rests on an assumption doing more work than it looks: you are trusting the institution's name as a stand-in for the work, and inheriting whatever its editorial process did or did not catch.
We can do something a document index cannot: resolve the person who wrote the thing and look up how much authority they hold on that specific subject, so the byline becomes a ranking input rather than metadata. An outlet's reputation stays one of those inputs. It is no longer the only one.
The same move works far outside the news. A paper is weighted by its researchers rather than the journal's name. Whether a skincare product does what it claims is answered by the practitioners who used it, not the company selling it. The right build for a game is whatever the pros playing it at the top level already said. In each case the useful source is a person, and the reason nobody retrieves them well is that document indexes do not know who people are.
Going to the source is also how you hear things first. A full-web crawler has to decide how often to revisit a billion pages, so any given page comes round on a cycle measured in days; our surface is far smaller and we know whose it is, so the accounts that matter are watched continuously. The larger reason has nothing to do with crawl budgets: the authorities on a subject publish before the coverage of them exists. The researcher posts the result, the analyst posts the thread, and the article about it follows hours or days later, if it is written at all.
Going to the source is not only a better citation. It is an earlier one.
One graph, many applications, one consistent answer
Every AI system that retrieves, generates or trains on web content has to choose what to put in front of the model, and has to be able to ask one question about each candidate: who is behind this, and do they know this subject? That capability, as a fast, composable API, is the context layer. It sits below the application, not inside it.
A RAG pipeline narrows its retrieval pool to ranked authorities before the first token is generated; a training run weights its corpus by authority rather than raw engagement; a newsroom checks a source in seconds. One graph, many applications, one consistent answer to the question of who knows what.
Trust is one of the things this buys, and it is worth being precise about why it matters more than it used to. Citing an institution and citing the work are not the same act; the first is a bet on a process you cannot see. That bet has been a good one for a long time. In an era where plausible text is free to produce and reputations are cheaper to borrow, a system that can name the person and weigh them on the subject is making far fewer assumptions than one that reads a masthead.
The same way TLS made "is this connection secure?" answerable across the entire web, this makes "who is behind this, and do they know this subject?" answerable across every system that ingests or produces information. The value compounds with adoption: every application that queries the index adds to the signal we use to improve it.
This is not a feature. It is infrastructure.
Trust infrastructure should not live inside the systems it evaluates
We did not choose to build this in Europe by accident. The hyperscaler ecosystem, however capable, has a structural conflict of interest here. The same companies whose business models depend on scraping, indexing, and monetizing web content are not the right companies to build the system that evaluates that content's authority. That conflict shapes incentives in ways that are subtle but real: what gets indexed, how it gets ranked, what counts as authoritative.
So the index runs on European infrastructure throughout, which is not a GDPR checkbox but a statement about where this layer should live: outside the blast radius of any single hyperscaler, and accountable to European courts. Sovereignty matters in almost every product category, and it matters most in the one deciding whose knowledge gets amplified and whose does not.
There is an institutional head start here as well. GDPR established that data about people has obligations attached to it; the AI Act is establishing that systems trained on that data have obligations too. We are not building to comply with regulation. We are building what regulation is pointing toward.
As automatic as HTTPS, as foundational as DNS
We want "is this from a verified authority?" to become as automatic as "is this HTTPS?" Every developer building a retrieval system, every journalist verifying a source, every model training run curating a corpus: they should all be able to assume that authority verification is available, fast, and accurate. The way no serious web developer today debates whether to implement transport security, no serious AI developer in ten years should have to build their own authority verification from scratch.
When every RAG pipeline can filter for ranked expertise, the floor for AI response quality rises. When every newsroom can check a source in seconds, the cost of disinformation increases. When every training set is weighted by genuine domain knowledge, the loop that degrades models over time gets interrupted. None of that happens instantly: the infrastructure has to exist before the applications that depend on it are built, which is why we started early and intend to be the layer the rest of the ecosystem builds on.
The people behind
the Authority Index
We're a tight-knit team with deep roots in public and social media data. Many of us have been working together for 5 to 10+ years. We know each other's strengths, we move fast, and we've built in this space long before it was fashionable.
Thomas Slabbers
Duco Janssen
Joran Cornelisse
Peter van Kampen
Want to work with us?
We're always interested in exceptional people who care about signal over noise.