Coverage methodology

What this site publishes

This site publishes research on pre-IPO AI companies: who is covered, how each name is classified, and how well each fact about it has been verified against a source. It publishes no price, valuation or reference mark. Those sections of the methodology are listed at the end as not yet published, rather than filled with placeholders.

Every fact shown carries its outcome, the kind of source that settled it, a link to that source, the date of the check and who made it. Where a check was made and no source settled the fact, the gap is shown as a gap. Nothing is hidden to make the coverage look more complete than it is.

Who is covered

The inclusion rule is version 2 of the taxonomy and is still proposed. It is printed here so that readers can see the rule the current classification follows; it is ratified separately before this site is made public, and the wording may change.

A company is listed when AI is a named, evidenced part of the business: Core, AI is the majority of revenue or valuation narrative; Material, AI is a disclosed, material growth driver below that threshold; Entrant, AI capability acquired, launched with disclosed traction, or the company has reorganised around AI, in the last 24 months. Cost-side AI use alone does not qualify; Entrant status expires after 24 months.

Routes

A company qualifies through one route. The route decides the tier.

Hard gates

All three must hold, whatever the route.

1. Named, not implied. The AI activity must be on the revenue or narrative side and citable to a primary artifact: a product announcement, a round release, a closed acquisition, investor-day framing, a filing. Cost-side use of AI does not qualify.
2. Evidence grade. A company statement or filed document outranks trade press, which outranks speculation. Third-party speculation about an "AI strategy" places a name on the Watchlist, not in the covered set.
3. Expiry. Entrant status lapses after 24 months unless the company is promoted to Material or Core. Without an expiry the universe drifts into "private technology", which is the failure the rule exists to prevent.

The classification a company carries is the research's own finding. Each assignment records its evidence grade: primary source, trade press, aggregator or unverified. That grade is never raised by a later verification; the two are shown side by side.

How a company is classified

Two axes are used for navigation and two tags record how AI entered the business.

Primary vertical

One per company. There are eight, and their addresses are fixed.

• 1. Frontier & foundation models · 8 published, 18 covered
• 2. AI compute & infrastructure · 10 published, 21 covered
• 3. Data & developer platforms · 6 published, 11 covered
• 4. Applied AI: enterprise & vertical software · 3 published, 8 covered
• 5. Consumer AI · 2 published, 7 covered
• 6. Physical AI & robotics · 4 published, 20 covered
• 7. Defense & dual-use AI · 1 published, 2 covered
• 8. Fintech, commerce & identity · 5 published, 5 covered

Stack layer

One or more per company: model, infra, data, application, hardware.

Posture

How AI entered the business: AI-native; product-led, an AI product line of the company's own; demand-led, AI is the demand driver rather than the product, as with energy for data centres; M&A-led, capability acquired; partnership-led.

Tier

Core, Material and Entrant, as the routes above define them, plus Watchlist for a name held for review that has not yet qualified under any route. A Watchlist name carries no route.

Restriction flags

A flag records a structural constraint on a name: why a fact about it may be hard to settle, or why holding or transferring exposure to it is restricted. A flag is shown on the dossier with its evidence and source. It explains a gap; it never hides one.

• Transfer restriction (ROFR or consent rights)
• Large strategic holder
• Sovereign or state capital
• US Entity List
• VIE structure
• NDRC review of foreign M&A
• Listing-course vetting
• Export controls
• Domicile or ownership cap

What counts as verified

Verification is a dated act, recorded for each fact about a company separately from the research that found it. One record is one check: the fact checked, the outcome, the kind of source, the source itself, the day and who checked.

Outcomes

• Confirmed. A source settled the fact. A confirmation always names its source.
• Refuted. A source contradicted the fact. A refuted fact removes the company from every published page until the correction has been made and checked again.
• Unresolved. The check was made and no source settled it. This is a result, not an absence of one, and it does not block publication: for a private company, facts such as domicile are often unverifiable from outside, and the honest answer is to publish the gap with its provenance rather than to publish nothing.

Sources

A check records what kind of source it relied on: a filing, a company registry, an exchange notice, a company statement, a venue page or press.

Press alone cannot confirm a lifecycle state, a listing venue, a headquarters or a domicile. Those facts have primary sources, and the weak spots in any research of this kind are exactly the places where press repeated a claim nobody checked. Press can still confirm an identity or a classification, and it can refute anything.

The publication gate

A company appears on this site only when its identity and its lifecycle state are confirmed against a primary source and no open check has refuted any fact about it. An unresolved fact does not keep a name out; a refuted one does.

Every check so far has been made by an automated process and is labelled as such on the dossier. A named human reviewer signs off on published claims before this site is made public; until then the reviewer field of this document is empty.

Known data gaps

These counts are read from the database when the site is built, so they move with every verification decision. Covered is every company in the coverage universe, published or not; published is the set that passes the gate.

A refuted fact removes its company from the published set, so the published columns have no refuted count. Published names with an unresolved fact.

Domicile is the weakest fact. It was seeded as a copy of the headquarters country, and the refutations so far show that assumption is wrong for companies incorporated offshore. Each company's own gaps are listed on its dossier, and the directory can be filtered to the names that have one.

Corrections

A refuted fact is corrected in three steps: the refuting check, the correction to the record, and a new check confirming the corrected value. The company is out of the published set from the first step until the third. The refutation stays in the record as the reason for the correction; nothing is edited in place.

A classification is never edited either. A changed view is a new assignment that supersedes the old one, with its own evidence and date, and a changed taxonomy is a new version alongside the old.

The reference mark

Not part of this release. Each of these sections will be published with the first mark, with its own version and reviewer.

Inputs and their sources

Not yet published

Confidence tiers

Not yet published

Gates and exclusions

Not yet published

Changelog

Not yet published

Track record

Not yet published

Versions

• 0.1, published 2026-10-23. First version: the coverage surface only, under taxonomy v2 as proposed.