Skip to content
ATDbbeta

Verification

ATDb shows a small chip next to every company, person, event, and news entity that tells you, at a glance, how confident you should be in the facts on that page. We call it the verification seal.

This page explains what the seal means, how we score facts, and what you can do when something looks wrong.

The seal — what each state means

You'll see one of three states on every entity header.

"Verified" (no source count)

The page's headline facts (parent company, founding year, M&A status, etc.) match a structured source we trust. We've cross-checked the most important fields against an editorial outlet, a regulator filing, a press wire, or a primary corporate source.

This is the default healthy state for established entities.

"Verified · N sources"

Same as above, but we've also recorded N distinct origin sources backing the headline facts. "Distinct origin" means independent publishers — wire-syndicated reprints of the same press release count as one origin, not N. You'll see this chip on tier-S companies and well-covered events, where multiple independent outlets cover the same announcement.

The higher N gets, the harder it is for a single bad source to drag a fact off-true.

"Unverified"

We don't yet have at least two independent sources backing the headline facts. The data may still be correct — most of it was originally enriched by AI before our verification system was built — but we haven't gotten our own structured sources lined up yet.

When you click the chip, you land on the entity's /verify/[slug] page, which lists the specific facts that need confirmation and lets you submit a source.

How we decide what's "verified"

For each entity, we score the headline facts against three lenses:

  1. Source class — different kinds of sources earn different weight. A regulator filing (e.g., SEC.gov, FTC.gov) carries more weight than a trade-press article; a trade-press article carries more weight than a press wire (PRNewswire, BusinessWire); a press wire carries more weight than a tertiary reference (Wikipedia, which itself cites other sources). The full classes are editorial, press_wire, primary_regulator, primary_corporate, and tertiary_reference.
  2. Stance — when we find a source about a fact, an automated classifier reads the source's framing. A source that says "X acquired Y" is a "support" stance. A source that says "X is reportedly exploring acquiring Y" is "rumored." Only sources that align with the canonical claim count toward verification.
  3. Independence — we filter out syndication. If five publications all reprint the same press release verbatim, that's one origin, not five. We score distinct origin domains, not raw article counts.

The seal updates automatically as enrichment runs find new sources or as you (the community) submit corrections.

Pinned facts

Some facts on ATDb are pinned by an editor. These show up the same as any other verified fact, but they're protected from being overwritten by future AI enrichment runs. Pinning happens when we manually research and confirm something — for example, when a community-submitted correction is approved.

Pinned facts always win against AI guesses, but they don't bypass the source-count threshold. A pinned fact still needs two independent origins before earning the "Verified · N sources" chip.

Hallucination scoring

When an enrichment AI proposes a new fact (or rewrites an existing one), an automated scorer runs to check whether the proposed claim is supported by the source the AI cited. This catches the most common AI failure mode — the AI confidently asserts something the source doesn't actually say.

Facts that score low for grounding get a "Help confirm" CTA on the entity page, asking you to look at the source yourself and tell us whether the fact matches what the source actually says.

Hallucination scoring is invisible to you in normal browsing — it just decides which facts get the gentle "Help confirm" CTA versus shipping fully verified.

The "Help confirm" CTA

You'll occasionally see a small "Help confirm" pill next to a specific fact on an entity page. Clicking it opens /verify/[entity-slug] with that fact highlighted.

From the verify page, you can:

  • Suggest an edit — propose a corrected value via the existing edit flow
  • Send us a source — paste a URL that supports (or disputes) the current value

We're building out the in-page confirm flow over the next few releases. For now, both routes go through normal community-correction admin review.

What gets verified

The seal scores headline facts — the kind of thing you'd see in the first paragraph of a Wikipedia article:

  • For companies: parent company, founding year, operational status, headquarters location, primary category
  • For people: current role, current company, notable past roles
  • For events: event type (acquisition / merger / IPO / etc.), the parties involved, the announcement date
  • For news: the source publication and publication date

Body-text claims (the overview prose) are NOT individually scored against the seal yet — they may still contain stale info even on an otherwise "Verified" entity. The seal is a header-row signal, not a guarantee that every word of the page is true.

Reporting something wrong

If you see something that looks wrong on any page:

  1. Click "Report a field" in the title row — drops a structured correction into the admin review queue
  2. Click the verification chip — opens the /verify page where you can suggest an edit or send a source
  3. Click "Spotted an error?" in the footer — opens a free-text feedback form

All three routes go through admin review. Submissions don't auto-apply to data.

Why we built this

ATDb's headline promise is "the AdTech database you can actually trust". AI enrichment is fast but error-prone; manual editorial review is accurate but slow. The verification seal is the bridge: every fact tells you, on the entity page itself, how solid the underlying evidence is — so you can spot the soft data before quoting it.

The system is new and improving fast. We'd rather show "Unverified" honestly than fake confidence we haven't earned.

See also: Glossary — definitions for source class, origin, fact_sources, Help confirm, and other verification terms.