The database says the company exists. It cannot say how it knows.
Discerno researches a company or a person on demand and hands back the evidence trail: the register number and the court that holds it, the source behind every claim, the tier that confirmed an email address, and — separately — the places where the sources disagree with each other.
Four things, and all of them narrow
Company overviews and person profiles are done well by many tools. These four are where Discerno is actually better, and the product is built around them rather than around a feature list.
Registry identity, not name matching
Every company record carries its register number (IČO, FN, HRB), the register court and the registered seat — and an explicit note when the number could not be reconciled. A name is not an identity: European mid-caps share names routinely, and the wrong match is worse than no match because nobody notices it.
Lookups run against local registry templates — ORSR and RegisterUz for Slovakia, Firmenbuch and WKO for Austria — because a Slovak manufacturer is not findable through English-language sources at all.
A document that decides, not a summary
The seller dossier opens with the decision: go or no-go, why now, through whom, and what kills the deal. The research follows the thesis instead of a template, which is the difference between sixteen useful pages and sixty-five that contain no sentence about what to do with the company.
An output gate blocks the document rather than warning about it — when the share of unsourced paragraphs, the domain diversity or the internal consistency of figures falls below the bar, the file does not get handed over.
Every address names the tier that confirmed it
Six tiers, free ones first: pattern generation, MX and provider, Microsoft 365 enumeration, literal evidence on a live page, the same page as the web archive held it, and only then a paid verifier — and that one only when the caller explicitly asks for it.
A guessed pattern is returned marked as guessed and must never be read as verified. A bought-in finder once returned 43 “hits” on our own Austrian and Slovak data, of which 3 were actually checked; four of the rest were wrong because Göth had become “goth” and Jürgen “jurgen”. Both spellings are now generated and the tiers decide.
Contradictions are named, not smoothed over
When sources disagree — two revenue figures for the same year, two different managing directors — the disagreement comes back as its own field rather than being resolved into fluent prose. Most research tools hide the discrepancy by writing well.
That matters for anything that ends up in a decision: a number nobody flagged as disputed will be quoted as fact in a meeting three weeks later.
Where it is used today
Three shapes of work. The figures are from our own runs on our own infrastructure — there are no customer numbers here, because there are no external deployments yet.
Outbound to DACH and CEE mid-caps
Before. A salesperson spends 40–90 minutes before a first contact: registry, website, news, LinkedIn, then guessing whether the person still works there. Half of it is unverifiable when questioned later.
After. One call returns the dossier with the decision on page one, every claim carrying its source, and the contact address labelled with the tier that confirmed it.
Preparation per account: under an hour of manual work → one request
Enriching an existing company database
Before. A list of ten thousand companies with names and nothing reliable: no register number, no current managing director, no idea which entries are duplicates of the same firm under two spellings.
After. Batch enrichment against local registries with the register number as the join key, so duplicates collapse and the wrong-match rate becomes measurable instead of invisible.
Own run, 21–22 Aug 2026: 3,528 Austrian and Slovak companies in one day
Research as a step inside another system
Before. An analysis agent needs live web evidence next to formal filings, but every general-purpose research tool answers in prose that cannot be parsed or audited.
After. Structured output per intent, contradictions as their own field, sources as their own list — consumed as an API step rather than read as a report.
Peak observed on our own infrastructure: 3,875 search queries per hour
Why the quality gate blocks instead of warning
A warning at the top of a document is read once and then ignored, and the document travels onward without it. So the gate returns a hard state: when the share of paragraphs without a source is too high, when the sources cluster on too few domains, or when the document contradicts itself on a figure, the file is not handed over.
The same mechanism handles a mistake that is specific to doing this work for several clients at once. A dossier carries a confidentiality marking and an offering it is written against; the names of your other clients are passed in so the gate can block the document if one of them surfaced in it. That is not a hypothetical risk — it is a correction to something that happened.
The structural decision behind the length is related. Deriving the section list from the thesis rather than from a template cut one real dossier from sixteen sections to ten and from 187 KB to 70 KB without losing a single load-bearing fact — because the sections that disappeared were the ones the research had nothing to say about.
Good fit
- Pre-contact research on DACH and CEE companies, where English-language sources do not reach
- Enriching a company list against official registers with the register number as the key
- Decision documents for named accounts, where the first page has to carry an argument
- Research as an API step inside an agent, a CRM or an analysis pipeline
- Work where a claim without a source has to be visibly a claim without a source
Not what this is for
- Buying a ready-made contact database — this is a research engine, not a data vendor
- Consumer or SMB lead lists at volume, where coverage matters more than verifiability
- US-only prospecting, where the established providers already have better coverage
- Anything that needs an answer in seconds: a dossier takes five minutes or more
Frequently asked
How is this different from ZoomInfo, Cognism or Dealfront?+
Those sell coverage from a maintained database: you query a record that already exists. Discerno researches on demand and returns the evidence trail — which is slower and more expensive per company, and worth it exactly where their coverage is weakest. For a Slovak or Austrian mid-cap the established databases frequently hold a name, a guessed address and a managing director who left in 2019. We would not recommend replacing them for US prospecting at volume.
Is it a deep research tool like the ones in ChatGPT or Gemini?+
Related mechanism, different product. Those write an essay for a human to read. Discerno returns structured fields per intent — register identity, decision makers, trigger events, contradictions, sources — because the output is meant to be processed by a CRM, an agent or a gate, not only read. The prose report is optional and off by default for enrichment runs, because it costs an extra model call and nothing downstream reads it.
What does “the tier that confirmed it” mean for an email address?+
The response always names which step produced the verdict: a generated pattern, an MX and provider check, a Microsoft 365 enumeration, a literal match on a live page, the same page as the web archive held it, or a paid verifier. A generated pattern is marked as a guess. The practical consequence is that you can decide per campaign which tiers you are willing to send to, instead of treating one confidence score as truth.
Can it re-write a dossier without paying for the research again?+
Yes — the raw material is stored per company, and a re-run can reuse it. Rewriting after a change to the prompt or the structure then takes minutes rather than a full paid collection. Forcing a fresh collection is a separate flag, so the choice is explicit rather than accidental.
How do you know the quality is not drifting?+
There is an eval harness with a small golden set: deterministic metrics that cost nothing — recall of required anchors, source count, domain diversity, schema fill, latency, cost — plus an LLM judge on a fixed rubric with cached verdicts. It exists so that every change to the engine is a measured delta and so the engine can be put on the same ruler as the general-purpose deep research products.
Where does the web data come from?+
A metasearch layer over many engines with rotating residential egress, static extraction first and escalation to a rendering path only for sites that need JavaScript, and a cache that serves repeats without going out again. The cache is what makes batch work affordable: in a list of ten thousand companies the overlap between queries is large.
Is it available today?+
Not as a self-service product. The engine runs and is used internally — the throughput figures on this page are our own runs, not a customer’s — and the interface is being finished. If the four properties above are what you are missing, an early-access conversation is useful now, because the shape of the first external deployment is still open.
What about GDPR?+
The engine processes business contact data and published company information, which is ordinary B2B processing with the usual duties: a lawful basis, information obligations towards the people in the records, and retention limits. Inference runs on our own hardware inside the EEA, which removes the Chapter V transfer question but not the rest. We say this plainly because vendors in this category tend not to.
Early access, while the shape is still open
The engine runs; the self-service interface does not yet. If registry identity in DACH and CEE, a decision document rather than a summary, or a named contradiction is the thing you are missing, this is the useful moment to talk — the first external deployment is not designed yet.
Ask about early access