How this was built

#What was done

Eight parallel research passes, run on 29 August 2026, each against a different slice of the question: the AI data and expert-labour market; recruiting and talent marketplaces; creator, UGC and clipping ad operations; adjacent verticals (red-teaming, robotics, voice, expert networks, forward-deployed engineering, back office, design, compute, community); the buyer — what funded startups and AI labs actually spend; middleman economics, theory and history; the legal, regulatory and operational risk surface; and the capital and outcomes layer — who raised what, at what mark, and what the exits paid.

Those passes produced roughly 54,000 words of research notes, kept verbatim in research/. Every figure in them carries a source or a tag: [EST] for a third-party estimate, [WEAK] for a claim from a source that could not be corroborated, [UNVERIFIED] for a claim that could not be confirmed at all, and [GN] where the citation resolves to a news-aggregation search rather than to the article itself.

A writing pass turned the notes into the fifty-odd pages of this atlas. Its job was compression and argument, not new research — every number on a page traces to a line in research/. Where the notes disagreed with themselves, the writing pass recorded the disagreement rather than choosing; those became the thirty-three items in What we could not establish.

A verification pass then read the pages back against the notes, checking figures, restating multiples on a consistent basis, and separating gross from net wherever both existed. It found more contradictions than the writing pass did, which is the expected result and the reason it was run.

#The tooling limit, and what it cost

The web-search budget was exhausted partway through. In several passes it was gone before the first query. From that point the research worked by direct URL fetch, publisher APIs, the SEC EDGAR API and CourtListener's RECAP API rather than by search. General search engines were also returning bot-degraded results.

That is a specific bias, not a general one, and it is worth stating precisely because it determines what this atlas is good at.

It biased sourcing toward:

• SEC registrants. Anything with a 10-K is well covered here. This is why What the public market pays for labour, What a rake can actually be and GMV is not revenue are the strongest pages in the atlas — Upwork, Robert Half, C.H. Robinson, Palantir, Uber, Etsy, Airbnb, DoorDash and Appen all file, and their filings say exactly what they mean by "revenue".
• Regulations and primary legal text. EUR-Lex, the Federal Register and IRS guidance pages fetch cleanly, which is why The law is about to arrive carries high confidence on a subject where most commentary does not.
• Court dockets. RECAP gave real cases, courts, filing dates and nature-of-suit codes — but not the complaints.
• Company pricing pages and blogs. Prolific's 42.8%, Deel's per-head pricing, HubSpot's startup discounts, Superside's tiers. Companies publish prices accurately because they want to be paid.
• Analyst profiles — Sacra, Contrary, Dealroom — which are frequently the only source that exists for a private company.

It biased sourcing away from:

• Forbes, The Information, Bloomberg, Reuters, WSJ and Business Insider, which either 403 automated fetching or sit behind paywalls.
• Anything that requires a search engine to find rather than a URL to fetch — which is most reporting about small private companies.
• Founder communities, niche podcasts and forum threads, which is why the $1M–$20M bootstrapped band is a hole rather than a section.

#The conventions used

#Confidence marks

Every page carries a confidence value in its frontmatter, and it describes the evidence, not the opinion. A page can be confidently wrong about what a market will do while being high-confidence about the facts it reasons from. The four values:

• high — the load-bearing figures come from filings, regulations, dockets or published prices. What the public market pays for labour, The law is about to arrive, One customer is a binary event, Appen.
• medium — the load-bearing figures come from reputable press, leaks or analyst profiles, and mostly corroborate each other. Expert data for frontier labs, Expert networks, Robotics teleoperation and physical-world data.
• low — the load-bearing figures come from vendor content, single sources or inference, and the page says which. Paid creators, clipping and UGC ad ops, Voice, speech and low-resource language data, Design, video and content production, Gray Swan AI, Distyl AI.
• contested — sources actively disagree on something the page depends on.

A low page is not a bad page. It is a page that has been honest about what is holding it up.

#The :::gap and :::flag conventions

Two callout types exist specifically to make holes visible rather than smooth them over.

:::gap marks a hole in the record — something the atlas would need to know and does not. It appears wherever an argument would be materially different if the missing fact turned out one way rather than another. Every :::gap on a vertical page corresponds to an entry in What we could not establish.

:::flag marks something stated but not verified — a claim the atlas is repeating with attribution rather than asserting. It is used where the source is real but thin, and where a reader might otherwise assume corroboration that does not exist.

The rule behind both: a visible hole beats a confident guess. The atlas would rather say "no source discloses the buy-side price in this vertical" than produce a plausible number, because a plausible number is indistinguishable from a real one once it has been repeated twice.

#Gross versus net, every time

Every revenue figure in the atlas is marked gross or net, without exception, because this is the central confusion in the sector and correcting it is part of why the atlas exists. A multiple computed on gross billings and one computed on net revenue differ by five to ten times with no change whatsoever in the underlying business. GMV is not revenue argues it; the The capital register is where it does damage.

#Wikilinks and red links

Cross-references use a double-bracketed slug, which renders the target page's title; a pipe after the slug substitutes your own wording. Links to pages that do not exist render red, and that is deliberate: a red link is a visible statement that the atlas refers to something it has not yet written, which is more useful than silently omitting the reference. What red links must never be is invented — every slug used comes from the page manifest, so a red link means "not written yet", not "guessed at".

#Scores

Vertical pages carry six scores from one to five, where higher is always better for the operator: budget depth, supply difficulty (hard is good), spread, hold, what AI does to it, and speed to first invoice. They are judgements, and an unargued score is worse than no score — so each one is justified in the body of the page it appears on. The definitions live in the authoring guide.

#How to extend this

The atlas is a directory of markdown files. To add a page:

1. Create a markdown file in content/. The filename is the slug — content/expert-networks.md becomes expert-networks.
2. Give it the frontmatter block specified in AUTHORING.md: title, kind, section, order, summary, updated, confidence, reading, tags. Vertical pages add stance, verdict, buyer, seller, spread, size and scores. Company pages add vertical, founded, hq, raised, valuation, revenue and status.
3. Add the slug to the page manifest in PAGES.md, so that other pages can link to it without guessing.
4. Follow the skeleton. Vertical pages use a fixed eight-section structure in a fixed order, so that a reader jumping between two verticals finds the same question in the same place. Consistency is the product.

The writing rules that matter most: every figure carries an inline source link in the sentence that uses it; if the number is a leak, a press claim or a company's own blog, the sentence says so; gross and net are distinguished every single time; and if the research notes say something is unverified, the page says it is unverified.

#The standing warning