The ways data makes money

Data makes money by three routes, improving the business, wrapping data into what it already sells, or selling it, and through about nine commercial models, from licences and benchmarks to media networks, AI deals and the sale of the data business itself.

In short. The route says where the money shows up: in margin, in the price of an existing product, or in a new revenue line. The model says how the buyer pays: for access, for an answer, for a placement, for a use, or for the whole business. Most companies pick a model before they know what their buyer is paying for, which is why so many data products are priced like files and sold like services.

What are the three routes?

The three routes, set out by Barbara Wixom, Cynthia Beath and Leslie Owens, are the first cut of the field, and the first page of this guide covers them. In brief: to improve is to use data inside the business and take the money as lower cost or better performance; to wrap is to build data into an existing product so customers pay more, buy more or stay longer; to sell is to put a price on the data, or on something made from it, for a buyer outside.

The routes tell you where the money lands. They do not tell you how the buyer pays, and that is where most of the commercial design happens. A bank can wrap its payment data into a treasury portal at no separate price, or sell the same insight as a subscription, or turn it into advertising placements. Same data, two routes, three different businesses. The models below are the forms the money actually takes.

What are the main data monetisation models?

1. Data licensing

The buyer pays for access to a dataset, for a defined use and a defined period. It is the oldest model, and healthcare, where data is scarce and regulated, uses it heavily: the register holds 13 healthcare data licences. In April 2026 TriNetX licensed Regeneron access to de-identified health data on about 300 million people, 170 million of them in the US, with the exclusive opportunity to connect Regeneron's own genomic cohorts to TriNetX's network; Regeneron also agreed to invest up to $200 million in TriNetX. In August 2026 Thermo Fisher's clinical research business took enterprise access to HealthVerity's claims dataset of more than 270 million de-identified US patient records, plus over 70 linked sources, for trial design and recruitment.

The scale of the model shows up in disclosures. Tempus AI reported about $200 million of data licensing contract value in a single quarter of 2026, its third straight quarter above $100 million, and named five drugmakers each holding licensing agreements worth more than $100 million.

A licence is simple to sell and hard to price. The buyer learns what the data is worth by using it; the licensor sees only the fee.

2. Data products and benchmarks

The buyer pays for an answer built from the data, not for the data. Benchmarks, indices, scores and market intelligence all sit here. PhonePe's PulsePro, launched in July 2026, sells category, retail-network and postal-code-level intelligence built from aggregated payment flows, above a free portal the company has run since 2021. Simbe, which photographs store shelves for retailers, launched a quarterly retail benchmark in April 2026 from more than 45 billion shelf images; its first release showed out-of-stock rates rising from a 3.8% baseline to 8.3% around Easter. Simbe has not disclosed who pays for the index or at what price; many data products start this way, as a public benchmark that advertises the paid one.

A data product travels further than a licence because the buyer never receives the raw data, so the seller can sell the same answer again. The product is only as good as the question it answers, and a benchmark nobody acts on is a press release.

3. Data built into an existing product

The wrap route as a model: the data is a feature, and the money arrives as price, retention or volume on the product the customer already buys. Xero's Industry Benchmarks, launched in June 2026, score each small business against its industry and country cohort on nine financial metrics, inside the Xero Analytics product. Bank of America's payments benchmarks, launched in September 2026, sit inside its corporate treasury portal, and the bank named no separate price for them.

The model is invisible in the accounts, which is both its strength and its risk. It earns the most of any model in many businesses, and it is the one most exposed to under-capture, because nobody can point to the revenue line it protects.

4. Access inside other people's tools

A newer model: the data is delivered as answers inside a tool the buyer already works in, increasingly an AI assistant. In September 2026 PitchBook became a data partner for OpenAI's assistant for financial services, with its firmographic records on companies, investors and funds queryable inside it and its answers citing back to PitchBook. PitchBook runs parallel partnerships with several other AI developers, so this is distribution, not exclusivity.

The model reaches users the data owner could never reach alone. It also moves the customer relationship to whoever owns the tool, which is why the citation back to the source matters more than it looks in this kind of deal.

5. Media networks

The buyer pays for a placement, and the data is what makes the placement worth paying for. Retailers, banks, restaurants, fuel stations and landlords have all launched advertising businesses on their customer traffic. Simon Property Group launched a media network across more than 200 US shopping, dining and entertainment destinations in September 2026, selling placements priced against its own visit and commerce data while the data stays with Simon. Walmart reported advertising revenue growth of 41% year on year in the fourth quarter of its fiscal 2026.

Media networks are the most common model in the register, 21 of its 112 deals, and the most crowded. Their weak point is proof: an advertiser buys the next campaign only if it believes the last one worked. In September 2026 DoorDash brought in Circana's independent household panel to validate its own advertising results; in one study, a beverage brand recorded $0.59 of incremental sales per household through DoorDash, three times Circana's category benchmark. A network that cannot show that kind of number, from a source the advertiser trusts, ends up selling cheap reach.

6. AI training and AI access licences

The buyer is an AI developer, and it pays for the right to train on data, or to reference it in generated answers. CuriosityStream reported $14.1 million of licensing revenue in one quarter of 2026, driven by AI developers. Chegg, the education company, announced in May 2026 that it would license a decade of step-by-step STEM solutions and its calibrated expert network to frontier AI labs. Universal Music licensed its catalogue to ElevenLabs in September 2026 for artist tools and a fan remix platform.

The models inside this model differ sharply in what they leave the seller. OneMedNet agreed in July 2026 to supply de-identified medical imaging for AI training for more than $11.5 million: a one-off fee, with no share in what the trained model later earns. The News/Media Alliance's agreement with Bria, announced in March 2026, is open to its roughly 2,200 member publishers on an opt-in basis, and pays those who join a recurring revenue share based on how much each publisher's content is used. Nine Entertainment's July 2026 agreement with Microsoft lets an AI assistant reference Nine's articles, including paywalled text, in generated answers that link back to Nine; the payment terms were not disclosed. A one-off fee, a share of usage and a deal built around sending readers back are three different businesses with the same buyer.

7. Rights to a specific use

Instead of licensing data broadly, the seller grants rights to one use, in one market, often exclusively. In September 2026 Catalist Sports granted Kalshi exclusive US prediction-market rights to live streams and low-latency data covering more than 65,000 tennis matches a year, plus other sports; sportsbook and Canadian uses stayed outside the grant. The same content can then be sold again, to a different buyer, for a different use.

The narrower the use, the more often the same data can be sold. This model rewards sellers who know exactly what each buyer needs the data for.

8. Marketplaces and exchanges

A venue stands between many data owners and many buyers, and its operator usually earns a fee or a share of what flows through it. In financial data, the Pyth Data Marketplace added proprietary publishers including Euronext FX, Fidelity Investments and Tradeweb in April 2026. Exchanges can also be run by their members: in real estate, three regional listing services in California, Nevada and Oregon launched a multi-state exchange in September 2026 to share real-time listing data across about 30,000 agents, brokers and other members.

The venue owns none of the data it sells, and in a working exchange it captures the value anyway, because nobody else is positioned to.

9. Selling the data business

The final model is to sell the business that does the work. In October 2026 LSEG agreed to buy Octus, a credit-intelligence business expected to earn over $300 million of revenue in 2026, for $3.535 billion. In May 2026 Publicis agreed to buy LiveRamp, the identity and data collaboration business, at a 29.8% premium to its share price. In October 2026 Fintech, a US beverage-alcohol platform, bought Ground Signal for its on-premise market intelligence built over more than a decade.

Buyers in these deals pay for a capability that turns data into an answer clients renew, and for the clients who already renew it. A data asset with no capability around it sells for much less, as the first page of this guide shows with the Spirit Airlines auction.

How do you choose a model?

Three questions decide most of it, and each points to a page later in this guide.

What is the buyer paying for? Access, an answer, a placement, a use or a business. A buyer who wants an answer will not pay licence prices for a raw file, and a buyer who needs to combine your data with its own will not accept a finished benchmark. The model follows the buyer's need, not the seller's convenience.

Does the money recur, and does it grow with the value created? A one-off sale is paid once. A flat annual licence recurs but stays flat however much the buyer earns from it. A share of usage or of advertising spend grows with how much the buyer uses the data; a price tied to the buyer's outcome grows with the value itself. The business keeps a share of the value as it is created, a rake, only in the last group. Who captures the value develops this.

Who does the work? Every model above needs someone to turn the data into what the buyer pays for. If the seller does it, the seller can price the result. If the buyer or a partner does it, they will capture most of the value, whatever the contract says about ownership.

Which models are growing fastest?

Three models show strong disclosed growth in the register. Media networks: Walmart's advertising revenue grew 41% year on year in the fourth quarter of its fiscal 2026, and DoorDash's advertising business, with Wolt's, passed a $1 billion annual run rate in 2024. Healthcare data licensing: Tempus's licensing business grew about 36% in a single quarter. AI licences: CuriosityStream's licensing revenue grew 48% year on year in the second quarter of 2026. The figures are measured differently and are not a ranking.

All three are growing for the same reason. Each has a buyer with a budget that did not exist, or did not reach the data owner, ten years ago: brands moving spend toward measurable placements, drugmakers paying for evidence outside clinical trials, AI developers paying for training material and for answers.

Questions people ask

What are the types of data monetisation? Three routes (improve, wrap, sell) and about nine commercial models: data licensing, data products and benchmarks, data built into existing products, access inside other people's tools, media networks, AI training and access licences, rights to a specific use, marketplaces and exchanges, and the sale of the data business itself.

What is the difference between selling data and licensing data? Selling transfers the data outright, usually for a single price. Licensing grants the right to use it, for a defined purpose and period, and the seller keeps ownership. The difference matters less than it seems: once a buyer has used the data, especially to train a model, it cannot be taken back.

What is data-as-a-service? Delivering data as a continuous, updated feed or query service rather than as a one-time file, usually on a subscription. It is a delivery method that can sit under several of the models above, licensing and access inside other tools most often.

What is a data marketplace? A venue where many data owners list data and many buyers license it, with the operator taking a share. Financial market data, real estate listings and advertising audiences all have established marketplaces.

Which data monetisation model is most profitable? It depends on who does the work and whether the price grows with the value; the disclosures are measured too differently to rank. They range from a one-off fee with no further share (OneMedNet's imaging deal) to a company-wide gross margin of 72.8% at a media company whose growth came from AI licensing (CuriosityStream), and an expected adjusted EBITDA margin of about 40% at a credit-intelligence business LSEG agreed to buy for about 12 times its expected 2026 revenue (Octus, completion expected in 2027).

---

Next in the guide: From raw data to data product, what turns data into something a buyer pays for, and what enrichability is. Every deal cited here is in the register, each with its reading.