Every commercial decision is a bet on what is happening inside another company: who to call, what to say, which competitor to fear, which product to build, where the next $10M goes. Get this picture right and everything downstream inherits the accuracy. Get it wrong and every step inherits the error, then compounds it.
The quality of that picture sets the size of the world you can act in. A seller who knows ten accounts works ten. A seller who knows a thousand works a thousand. A founder who can see every company in a category at once picks a market nobody else has noticed.
01The old world
For decades, knowing what is happening inside a company came in two forms. Analysts were accurate and expensive. Databases were cheap and described a company by what it is (a sector code, a headcount band, a revenue estimate, an address) rather than by what is happening to it. I lived the first form for ten years, running go-to-market programs at BCG. Research teams read everything a company had ever said in public, looking for one usable sentence. If the budget allowed, we booked expert calls an hour at a time, hoping the former employee on the line remembered something useful. If the company was private, good luck. The picture was reasonably good. It took weeks, and it went stale almost immediately.
Around 2015 the industry decided the fix was intent data: watch the web, count the clicks, score the account on the count. That premise is collapsing. 68 percent of US Google searches ended without a click in the first four months of 2026, up from 60 percent in 2024.1 The buyer no longer visits the pages intent vendors scrape; the footprints are being made inside ChatGPT, where nobody is watching. Scores go stale, sellers stop trusting them, and enterprise win rates sit in the mid-teens, where they have sat for a decade.2 Sellers spend about 40 percent of their week selling.3 Much of the rest is rebuilding, badly, the picture the research teams used to build.
Buyers still research. What stopped happening is the visit that intent data is priced on. Source: SparkToro and Similarweb, June 2026.
02Five against a thousand
Take one company as a database sees it: a name, a sector code, a headcount band, a revenue estimate, a list of technologies inferred at arm's length, and an intent score that says someone there read an article. That is the picture most of the world’s sellers work from today.
A year ago we started building a different one. The first account we mapped had 147 technology suppliers serving it, and we could say what each one did there; the expert we had paid to talk to knew about five. That number now looks small. On one US operator the record holds 1,906 suppliers, graded on presence and trend, by business unit, with evidence by year back to 2018. On a large bank it holds 456. On a retailer it holds 334 from the public record and 50 more that record cannot see, because offshore delivery leaves no trace in it. Across the model, the same evidence read in the other direction produces a client list for any supplier: one IT services firm that has never named a client in thirty published engagements has 82 on the record, dated, with the delivery location.
Each mark is one supplier relationship, with the date it was read. The five on the left is what a paid expert call produced on the same account.
Instead of a flat technology stack, the record holds the movement inside it: a warehouse being sunset, the platform replacing it, the team funded to make the move, all named together in a single document the company wrote itself. Instead of an org chart scraped from titles, it holds seats verified in role on the day, with the reporting line graded separately because the two fail independently. And it holds the programs nobody outside the building was meant to know about, with the evidence attached.
The first thing anyone says when they see a record is a question: how did you get that. No single source produces it. Companies say very little about themselves and are compelled to say a great deal in places nobody reads. An NDA binds what a company says about itself. It does not reach the record that company is required to leave elsewhere, nor the account of the work written by the people who did it. We read those records, on a schedule, across more than sixty channels. Every line carries its source, its date and its confidence. We call it the Atlas methodology. Where two sources disagree, the record shows both. What one pass gets wrong, the next pass says.
A database has facts as they were published. What it lacks is the thing every line of ours has: a derived fact about that specific company, traceable to its evidence, describing what the company is doing rather than what it is. Five against a thousand is the gap between how the world knows companies today and how it is about to. It is harder to build, and it compounds. It is the thing we would have paid anything for, on every program we ever ran.
A supplier’s clients and a client’s suppliers are one dataset. It is why a firm that has never published a client list still has one here.
03Why now
What changed underneath is the price of reading. The cost of running a capable language model over a million tokens fell from $20 in November 2022 to $0.07 by October 2024, a 280-fold drop in two years.4 An annual report is roughly 100,000 tokens. Reading everything every large company on earth is compelled to disclose, every day, became a line item.
A 280-fold fall in two years. An annual report is roughly 100,000 tokens, so reading everything a large company is compelled to disclose became a line item. Source: Stanford HAI, AI Index 2025.
Reading was never the interesting part. A good salesperson knew what to read for, and that judgement is what we have spent a year writing down. Every offering (436 mapped by hand so far) has preceding conditions, states or events that create the need for it. Apply that rule at scale and a company stops being a row in a database. It becomes a living thing: what it is doing, what it is leaving, who it depends on, who depends on it, what it has promised and by when, with the evidence attached to every line.
04Ask the model
We believe every company in the world will have a living map: a current, structured, evidenced picture of what is happening to it, refreshed on the cadence of the fact, that any person or agent can query before making a commercial decision about it. We believe one company will hold the most comprehensive version of it. We intend that company to be Zylabs.
The numbers, stated plainly. More than 40 million companies are resolved and connected on the map, which is every company that sells to, buys from or competes with another, more than 250,000 of them are prioritized for reading, and more than 2,500 carry a deep record today. It already reads in both directions, from an account to its bench and from a supplier to its book. Bloomberg made every price in the world knowable and finance reorganized itself around the terminal. What is happening inside companies is the larger and less legible fact, and it has never had its terminal.
Areas are to scale. Anybody can buy a universe count. What counts is how many companies carry all six stages of the Atlas methodology.
Two phases. The first is every large enterprise. Sales teams come first, because they feel the pain first and pay for it first. Every account read, every conversation opened, every win and loss recorded sharpens the model, and that is the flywheel we are building with our first customers. By the end of this phase no enterprise seller starts a conversation blind, and the sales research department, the analyst pool and the expert-network call become things people remember doing.
The second is every company that sells to companies, and every use of the picture. A product leader sees which platforms an entire industry is walking away from a year before the earnings admit it. A competitor sees a rival’s churn list that was never published. A procurement head sees a supplier bench of a thousand firms whole for the first time. A founder sees the 300 companies about to have the problem she is solving, before they know they have it. And the thousand agents that will soon act on behalf of sellers, buyers, investors and analysts all begin their work the same way: by asking the model what is going on at this company. At that point the model is the product. The applications are how we proved it.
We do not know precisely what this looks like in ten years. We know what it looks like in five, and we know that when every commercial decision starts with “ask the model”, someone will own the model. We plan to be that someone, and we are looking for the people who want to build it with us.
1. SparkToro and Similarweb, zero-click search study, June 2026, as reported by Search Engine Land.
2. Forrester, Buyers’ Journey Survey 2025 and 2026 Buyer Insights report.
3. Salesforce, State of Sales 2026, 4,050 sellers across 22 countries.
4. Stanford Institute for Human-Centered AI, AI Index Report 2025.
Supplier, client and record counts are from Zylabs work, read between August and September 2026.