How did you get that?
The question every first meeting asks. No single source produces the answer, so it takes six stages run in order, a standard that refuses most of what it finds, and a map that makes every company legible through the companies around it.
The Atlas methodology turns one company into one record in six stages: resolve the entity, read what the company and its people leave behind, triangulate across channels of different types, derive what no source states, grade every line, and connect the record to every other record. The methodology is published in full on this page and the channel list behind it is not.
Six stages, run in order, on a schedule.
This is the research method a strategy firm runs by hand, on one account, over six weeks. We built it to run on a hundred thousand accounts, on a schedule.
We separate the legal entities, trading names, subsidiaries, acquisitions and the units that buy on their own budgets before anything else runs. A picture that fails here cannot recover later, because a group read as a single account produces a wallet estimate wrong by an order of magnitude and a stakeholder map in which nothing has an owner.
A company controls what it chooses to publish about itself, and it controls neither the record it is required to leave elsewhere nor the account of the work written by the people who did it. We read both. Primary sources are what the company and its people write themselves, which covers job specifications, engineering writing, conference talks, tender documents and practitioner accounts. Secondary sources are what the company is compelled to file and what third parties are compelled to record about it. Each channel runs on its own cadence and carries a stated blind spot that is read before the channel is called.
The same question is put to channels that fail in different ways, because a statutory filing, a job specification and an account written by the person who did the work are each wrong about different things. Where they point the same way the reading is stronger. Where they conflict, the record carries both readings with their dates rather than resolving them into a single value that neither supports. The record also states which classes of evidence were silent on the account, because a channel that cannot see a kind of company returns nothing and nothing reads exactly like a clean account.
The model computes what no channel reports. Supplier presence and trend come from evidence density over time. A wallet band comes from top-down and bottom-up benchmarks reconciled against each other rather than averaged. Team size comes from record counts. Program phase comes from the language of the documents that name it. The gap between what a company says and what its hiring does falls out of the two read side by side. This stage is what separates a record from a search result, and it is the stage whose machinery we do not publish.
Every line carries the method that produced it and the date it was read, shown wherever the line is shown. Precedence between the four methods is fixed, so a stronger method is never overwritten by a weaker one however recent the weaker reading is, and a modeled value is labeled modeled everywhere it appears.
We never read a company on its own. Every supplier, client, subsidiary and partner connected to it is read as well, because each of them leaves evidence about the company that it never publishes itself. A supplier's clients and a client's suppliers are the same evidence read in two directions, which is why a services firm that has never published a client list still has one here, dated, with the delivery location attached.
Every data point carries one of four confidence bands.
The band says how the data point was produced and how much weight it can bear. Precedence between the bands is fixed, so a stronger method is never overwritten by a weaker one however recent the weaker reading is.
Each field carries its own interval, so a disclosed figure moves when the company next reports while a program, a seat or a supplier placement is re-read on a cadence of days to weeks. A seat counts as current only while the role has no end date, and that single check is what keeps departed people out of a published stakeholder map.
Enough to judge the record. Not enough to rebuild it.
An NDA binds what a company says about itself. It does not reach the record that company is required to leave elsewhere. Finding those records, reading them at scale, deriving what none of them states and showing you how much evidence each line rests on is the entire business, so the method is public and the machinery is not.
- The Atlas methodology, stage by stage, as set out on this page.
- The classes of evidence we read, and the stated blind spot of each class.
- The grade and the date on every line of every record.
- The volume read for a given account, and which classes were silent on it.
- Every correction, written into the record, with the claims that rested on it withdrawn.
- The channel list, and which channel produced which line.
- The detection patterns that turn a document into a graded fact.
- The derivation model, the field dictionary and the offering library.
- The entity resolution graph that makes one company one row.
We never read a company on its own.
Every supplier, client, subsidiary and partner connected to a company is read as well, because each of them leaves evidence about the company that it never publishes itself. A supplier’s client list, a partner’s case study and a subsidiary’s filing all describe the account from the outside, and reading them together is what turns a thin profile into a complete picture.
Accounts sit on the inner shell, suppliers on the outer; an edge is a dated engagement. Hover a node to isolate its relationships.
The same evidence reads in both directions, which is why a firm that has never published a client list still has one here. More than a million relationships, of which roughly a quarter connect a supplier to more than one account.
Every instrument is blind to something.
A channel that cannot see a kind of company returns nothing for that company, and nothing reads exactly like a clean account. So each channel carries a stated blind spot, the blind spot is read before the channel is called, and every record says on its surface which classes of evidence were silent on it.
Every bar starts at first evidence and ends at most recent. Nothing here is a judgement about a supplier: a bar that stops in 2023 means the account stopped disclosing that firm, which the record states and does not interpret. Two firms are left unnamed because their windows have closed.
Common questions
We publish the six stages, the classes of evidence we read and the stated blind spot of each class, the grade and the date on every line, the volume read for an account and which classes were silent on it. The channel list, the detection patterns and the derivation model stay ours, because any provider can buy a database and the work that separates one from the Atlas is knowing what to read for and what to compute from it.
We are wrong less often because we write down less and we say which method produced each line, so you can see which parts of a record are load-bearing and which are not. Where two channels conflict, both readings stay on the page with their dates rather than being resolved into one. Corrections are written into the record by name, and every claim that rested on a withdrawn line is withdrawn with it.
Each field carries its own interval. A disclosed figure moves when the company next reports, while a program, a seat or a supplier placement is re-read on a cadence of days to weeks. The date on every line is the date it was last read.
Private and non-US companies leave a thinner trail and the record says so. Stage two carries more of the weight, the coverage note states which classes were silent, and confidence on the affected lines drops accordingly. A thin record labeled thin is useful, and a thin record presented as complete is not.
Yes, and the fastest way to find out is to name an account you know cold on a demo. Two claims were withdrawn from the retailer record on this site between its first and second pass, both stated in the record itself.