Skip to content

AI data market

Vertical or horizontal AI data business? Price the next category first

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

A vertical AI data business goes deep in one industry's records and buyers; a horizontal one sells one data type or service across many industries. Investors should price the cost structure, not the label: verticals reuse rights templates, reviewers and schemas inside a category but carry its concentration, while horizontals spread that risk and pay a fresh opening cost for every category they enter.

Value drivers

Vertical horizontal AI data

This page covers vertical or horizontal AI data business? Price the next category first.Illustrative

Key takeaways

  • The vertical-horizontal difference is where reusable work stops: inside one industry, or at generic tooling.
  • Normalize every metric to the category, the unit where rights templates, reviewers and schemas start and stop paying off.
  • In the illustrative ladder, breadth costs 2.6 times the reviewer-weeks for equal output and cuts one-category exposure from 50 percent to 8.3 percent.
  • Filings show depth (Verisk), the cost of an adjacent category (CoStar) and category risk surviving breadth (Dun & Bradstreet).
  • A horizontal platform whose category opening costs do not fall is a services business with a broader sales list.

What makes a vertical AI data business different from a horizontal one?

Where its reusable work stops. A vertical AI data business licenses or builds data inside one industry, so the rights templates, sector rules, schemas and domain reviewers it pays for once keep paying off on the next dataset in that industry. A horizontal business sells one kind of data or service across many industries, so its reusable layer is tooling and process, and each new industry brings rights, rules and reviewers it has not paid for yet.

The investor question follows from that boundary. A vertical earns more per unit of effort inside its category and rides that category's fortunes; a horizontal spreads demand across categories and pays an entry price each time it widens. Neither shape wins in general. A fair comparison normalizes every metric to the category, because that is the unit where reusable work starts and ends.

This article builds that comparison: four roles that separate the shapes, three public companies that show depth, the cost of crossing into a neighboring category and concentration inside breadth, an illustrative category ladder with per-category metrics, the services test, where SourceX sits by its own description, the risks, and the questions to ask.

Who holds the rights, does the work, builds the tools and keeps the value in each shape?

Split every data business into four roles before comparing shapes, because labels such as platform or marketplace hide who actually performs each one. The table sets the two shapes side by side, role by role, with the evidence that would confirm each reading.

Who holds the rights, does the work, builds the tools and keeps the value in each shape?
RoleVertical shapeHorizontal shapeEvidence to ask for
RightsContributors and licensors in one industry, often on repeat termsMany industries, each with its own contract forms and noticesShare of datasets licensed under an existing template versus a new one
LaborDomain reviewers reused across datasets in the categoryGeneralist teams retrained or rehired for each categoryReviewer-weeks per dataset, first versus repeat, by category
ToolingSector schemas and taxonomies that compoundGeneric pipelines that work everywhere and differentiate nowhereWhich tools carried over unchanged into the latest category
Recurring valueRenewals tied to one industry's budgets and rulesRepeat buyers across industries, each a smaller shareRevenue share of the largest category and of the largest buyer

What depth looks like in a filing: an insurance data company

Verisk is a clear public example of a vertical data business, and its own documents show both the moat and the ceiling. Its fiscal 2025 10-K and 2025 annual report describe statistical agent services that a substantial number of U.S. property and casualty insurers use to report to regulators, a role it holds as a licensed or appointed statistical agent in all 50 states, Puerto Rico and the District of Columbia, built on proprietary and contributory data about insurers' premiums and losses. The 10-K adds that in 2025 its clients included all of the top 100 U.S. property and casualty insurers for the lines it serves.

Depth was a choice it kept making. Verisk announced in February 2023 that it had closed the sale of Wood Mackenzie, its energy data business, and its 2025 annual report records the sale of its marketing solutions business on December 31, 2025. A company whose clients already include every top-100 insurer in its lines grows by selling more inside the category or by entering another one; these divestitures point to the first.

The same 10-K names the vertical's characteristic risk: competition from insurers that develop their own technology and actuarial methods. In a vertical, the customers are often the parties that hold the underlying data and could keep it. For an AI data business the parallel is direct, since the companies whose records make a vertical dataset valuable can license them elsewhere or use them in house.

What crossing into a neighboring category costs: a real estate information company

Years of reported investment, not a sales push. CoStar built its business on commercial real estate information and has been entering residential real estate through Homes.com. Its 2026 outlook release, filed with the SEC, tracks that move as a net investment, which it defines as revenue minus directly attributable and allocated costs, plans to cut that net investment every year through 2030, and expects Homes.com to reach positive adjusted EBITDA in 2030. Its second-quarter 2026 release reports the residential segment's adjusted EBITDA turning positive for the first time.

Homes.com is a consumer marketplace, not a dataset business, so the comparison is about shape rather than model. What carries over is the accounting of expansion: the adjacent category was planned and reported as its own investment line with a dated path to break-even, not folded into incremental sales. A vertical AI data company proposing a second industry should be able to show the same plan, in reviewer-weeks, months and new rights templates if not in dollars.

Does breadth remove concentration? A cross-industry business data company

Not entirely. Dun & Bradstreet sells business data across industries, the textbook horizontal shape, and its 2024 10-K describes a diversified client base in which no single client is a large share of revenue. The same filing still warns that a substantial portion of revenue comes from companies in financial services. Breadth spread the customer risk; it did not remove the category risk.

For AI data businesses the lesson cuts both ways. A horizontal supplier can serve many industries and still lean on one class of buyer, such as model developers, or on one category of supply that happens to be easiest to source. Ask for revenue by buyer category and by supply category separately, because concentration can hide on either side of the ledger.

The category ladder: one illustrative model, normalized per category

Normalize every metric to the category and the two shapes become comparable. The model gives two illustrative firms the same output, 12 datasets a year, and the same unit costs; only breadth differs. The vertical firm delivers all 12 in one category, and the horizontal firm delivers 2 in each of 6. Work is counted in reviewer-weeks, never money, and every input is an editable assumption, not a measurement of any company.

Inputs (illustrative): opening a category takes 30 reviewer-weeks for rights templates, sector rules, schema mapping and recruiting domain reviewers; the first dataset in a category takes 10; each later dataset in an opened category takes 6. The totals follow: the vertical firm spends 106 reviewer-weeks, 8.8 per dataset, and the horizontal firm 276, or 23.0 per dataset, about 2.6 times as much for the same output.

What the horizontal firm buys with that extra work is resilience. If demand in its largest category halves, the vertical firm loses 6 of its 12 datasets and the horizontal firm loses 1. Whether that insurance is worth its price depends on how stable demand is in the specific categories involved, which is a question about evidence, not about labels.

The per-category metrics, computed from those inputs and rounded once, half up, to one decimal:

The category ladder: one illustrative model, normalized per category
Metric (per year)Vertical: 1 categoryHorizontal: 6 categoriesWhat it tells an investor
Reviewer-weeks in total106276The price of breadth at equal output
Reviewer-weeks per dataset8.823.0Unit effort once opening costs are spread
Datasets reusing an opened category91.7%50.0%How much of the work compounds
Volume in the largest category100.0%16.7%Exposure to one industry's budgets and rules
Datasets lost if that category halves6 (50.0%)1 (8.3%)What the extra effort insures against
Next dataset: new category vs opened one40 vs 6 reviewer-weeks40 vs 6 reviewer-weeksWhy widening is a decision, not a default
Two stacked bars compare reviewer-weeks for 12 datasets: the vertical firm's 106, mostly repeat work, and the horizontal firm's 276, mostly category opening work, beside gauges of volume lost if the largest category halves, 50 percent versus 8.3 percent.
The category ladder. Illustrative inputs; values are model outputs, not company data.

When does a horizontal platform turn out to be repeated custom work?

When its reuse share stays low as it grows. A horizontal firm that calls itself a platform should show each new category costing less to open than the last, because templates, reviewers and tools carry over. If every category still needs fresh rights work, new sector reviewers and a new schema, the firm is selling the same project again and again, and its economics are a services firm's whatever the label says.

The test fits the ladder's terms: list opening cost per category in the order the categories were entered. Falling opening costs support the platform reading; flat ones point to services. A vertical can fail the same test from the other side, if its repeat datasets cost nearly as much as its first because every buyer wants bespoke labeling.

The strongest counterargument is that some horizontal value is real tooling. De-identification, delivery and contract machinery do travel across industries. Give credit for it where the opening cost actually falls, and nowhere else.

Where SourceX sits between vertical and horizontal, by its own description

Horizontal on supply, with named focus areas. SourceX licenses business data directly from companies and licenses it onward to AI developers in its own name, covering sourcing and qualification, rights review, contracting, preparation, delivery and payment. Its eligibility is set by company size and geography rather than industry: U.S., Canadian and EU companies with 50 or more full-time employees at their peak, and companies in several other markets when the offered data is at least 90 percent English. Its investor page names six focus areas: software, BPO and contact centers, professional services, finance, healthcare administration and logistics.

That puts the ladder's questions squarely on it. Datasets are sourced against a buyer's request rather than held in stock, so the reuse question is whether templates and reviewers built for one request carry into the next request in the same focus area. Every supplying company still goes through its own rights review and approves the scope and terms before anything is delivered. SourceX publishes no category-level figures; company and commercial metrics go to investors directly, and its Rule 506(c) raise accepts only verified accredited investors as purchasers.

Six ways the vertical-versus-horizontal reading could be wrong

Six developments could flip the comparison. Track each against a dated source where one exists:

  • Buyer concentration can outweigh category spread. If a few model developers account for most purchases of AI training data, a horizontal supplier's mix of industries protects less than it appears to.
  • Customers can become competitors. Verisk's 10-K names insurers that build their own technology as competition; in AI data, the companies holding vertical records can license them directly or keep them for their own models.
  • Rules can reprice a category after entry. Sector rules for health, financial or insurance records can raise the opening cost of a category a firm has already entered, or close it.
  • Synthetic and openly licensed data can erode the reuse premium. If a category's tasks can be learned from generated or public data, a vertical's compounding templates lose value.
  • Expansion can run longer than planned. CoStar reported its residential entry as a multi-year net investment with break-even targeted years out; a second category rarely opens on the first one's unit costs.
  • The ladder's inputs are assumptions. Opening and repeat costs vary by industry, and a category with a ready rights template may open at a fraction of the illustrative 30 reviewer-weeks.

Diligence questions, and the method behind the category ladder

Method: the four-role comparison and the category ladder were drawn up for this article on 2026-10-27. Company examples come from Verisk's fiscal 2025 10-K and 2025 annual report and its February 2023 release on the Wood Mackenzie sale, CoStar's 2026 outlook and second-quarter 2026 releases filed with the SEC, and Dun & Bradstreet's 2024 10-K, read through search excerpts of those documents. The ladder's inputs are illustrative; a short script reproduces every metric from them, with fractions kept exact and rounding applied once, half up.

Not measured: actual opening and repeat costs at any company, SourceX included; revenue by category; buyer counts; and whether any filer's experience transfers to AI training data. Those require company records, which an investor should request directly, from SourceX through sourcex.si/investors as from anyone else.

Questions to put to any vertical or horizontal data company, SourceX included:

  • What did your last three categories cost to open, in reviewer-weeks or months, in the order you entered them?
  • What share of last year's datasets reused an existing rights template, and what share needed a new one?
  • What share of revenue came from your largest category and from your largest buyer, and how did each move over two years?
  • Which tools and reviewers carried unchanged into your most recent category, and which were rebuilt?
  • If demand in your largest category halved, which contracts would renew and which would lapse?
  • Which customers in your main category could license or build the same data themselves?

Frequently asked questions

What is a vertical AI data business?

A company that licenses or builds data inside one industry, such as insurance or real estate, so its rights templates, sector rules, schemas and domain reviewers are reused across datasets in that industry. Its strength is compounding depth; its exposure is that industry's budgets, rules and customers.

What is a horizontal AI data business?

A company that sells one kind of data or service across many industries, such as annotation, contact data or general business records. Its reusable layer is tooling and process, so demand is spread out, but each new industry brings rights, rules and reviewers it must pay for again.

Is a vertical or a horizontal data business the better investment?

Neither in general. A vertical spends less effort per dataset once a category is open but carries that category's concentration; a horizontal pays more to widen and gains resilience. The answer depends on per-category evidence: opening costs in order of entry, reuse shares and revenue by category and buyer.

How do you compare data companies that serve different numbers of industries?

Normalize every metric to the category: reviewer-weeks per dataset, the share of datasets that reuse an opened category, the share of volume in the largest category, and the cost of the next dataset in a new category against one in an existing category.

Does SourceX publish category-level metrics?

No. SourceX shares company and commercial metrics with investors directly rather than in articles. Investor materials can be requested through sourcex.si/investors, and any offering is open only to verified accredited investors.

Sources

  • Verisk's fiscal 2025 10-K describes statistical agent services used by a substantial number of U.S. property and casualty insurers, a 2025 client base including all of the top 100 such insurers for its lines, and competition from insurers' in-house technology. Source
  • Verisk's 2025 annual report describes its role as a licensed or appointed statistical agent across all states, its proprietary and contributory premium and loss data, and the December 31, 2025 sale of its marketing solutions business. Source
  • Verisk announced the closing of its sale of Wood Mackenzie, its energy data business, in February 2023. Source
  • CoStar's 2026 outlook release defines Homes.com net investment, plans annual reductions through 2030 and targets positive adjusted EBITDA for Homes.com in 2030. Source
  • CoStar's second-quarter 2026 release reports its residential segment's adjusted EBITDA turning positive for the first time. Source
  • Dun & Bradstreet's 2024 10-K describes a diversified client base with no single large client and a substantial portion of revenue from financial services companies. Source
  • SourceX's investor page says it licenses business data directly from companies and onward to AI developers in its own name, names six focus areas and shares metrics with investors directly. Source

Request SourceX investor materials

SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.

See if your company qualifies