From raw data to data product

A data product is an answer a buyer pays for, built from data by work the buyer does not have to do: work that produces something the data does not state on its own, and work that puts the result in a form the buyer can act on. Raw data is what a business holds before any of that work is done.

In short. Raw data is cheap because the buyer still has to do the work, and the buyer captures what the work produces. A data product is valuable because the seller has done the work and can price the result. The distance between the two is the capability, built over years, that turns one into the other. That capability, more than the data, is what buyers of data businesses pay for.

What is raw data worth?

Less than its owner usually thinks. The data a business holds before any work is done on it, the substrate, records what happened: transactions, visits, documents, images, messages. On its own it answers nobody's question. A buyer who takes it has to clean it, make sense of it and connect it to the decision it cares about, and the buyer prices that work into what it is willing to pay.

The first page of this guide gives the sharpest example in the register. In August 2026 the corporate data of Spirit Airlines, about 100 million emails, 500 million internal chat messages and 30 million lines of code, sold in a bankruptcy auction for $10 million. The volume was enormous. What the buyer received was the records themselves, de-identified, and nothing built from them.

Raw data also has three commercial weaknesses that a product does not. It is easy to copy, so once delivered it is out of the seller's control. It is hard to price, because its value depends on what the buyer does with it, and only the buyer knows that. And it is generic: the same file goes to every buyer, so it competes with every other file that roughly answers the same need.

What turns data into a product?

Two kinds of work, and the difference between them matters more than it first appears.

The first is enrichment: work on the content of the data that produces something the data does not itself state. Validating and structuring it, matching records, modelling, inferring, classifying. The second is product work: putting the result in a form a buyer acts on, such as a benchmark, a score, a placement or an answer inside a tool the buyer already uses, and delivering it at the cadence the decision needs. A business needs both to sell anything. Only the first creates new value from the data. The second decides how much of that value reaches a buyer, and on what terms.

Enrichment

Octus, the credit-intelligence business LSEG agreed to buy in October 2026 for $3.535 billion, sources hard-to-find credit documents, covenants and deal terms, and turns them into structured, searchable data; it is expected to earn over $300 million of revenue in 2026. OneMedNet's imaging deal for AI training, worth more than $11.5 million, rests on the same kind of work: de-identifying medical images, matching them to the buyer's specification and tracing where each one came from.

Modelling goes further, and produces what no record states. Circana's Liquid Mix, launched in September 2026, runs a marketing-mix model on Circana's own point-of-sale data and a household panel covering 290 million households, and sells advertisers channel-by-channel return on their spend. The modelling code is Google's, released free in February 2025; the value is in running it on Circana's own data. Simbe's quarterly retail benchmark starts from more than 45 billion shelf photographs. No photograph states an out-of-stock rate; the work infers one, and the first release showed out-of-stock rates rising from a 3.8% baseline to 8.3% in the 14 days around Easter and Passover.

Product work

Product work is real commercial work, and much of the price is set by it. Xero's Industry Benchmarks, launched in June 2026, score each small business against its industry and country cohort on nine financial metrics, labelled leading, lagging or in line, inside the analytics product its customers already use. PhonePe launched PulsePro in July 2026, a paid enterprise tier built on its aggregated payment data, above a free statistics portal it has run since 2021; the paid tier is organised into three products, for category growth, retail networks and individual postal codes.

Product work is not enrichment. Reformatting, repackaging or redisplaying data adds nothing the data did not already say, which is why a competitor holding the same data can copy it. Enrichment is harder to copy, because the output depends on both the data and the capability that worked on it.

What is enrichability?

The book behind this site defines it precisely: enrichability is the capacity of a data substrate to yield commercial value when a capability stack operates on its content to produce outputs the substrate does not itself express. Put simply, the data has the value in it, and the capability gets it out. Three things have to be true: the substrate is data with a structure the capability can work on; the capability operates on its content; and the output is something the data did not already state. Data that has only been reformatted or repackaged does not qualify.

Enrichment produces something new in three ways.

From one dataset alone. The work brings out what the data implies but does not state. A transaction history yields a segmentation of customers; a set of shelf photographs yields an out-of-stock rate; a clinical record yields a risk score. The signal was always in the data. The capability makes it visible.

Across two datasets with no shared key. The same customer, company or place is recognised in both datasets by probability, from what the records say about it, rather than by an identifier both of them carry. The answer exists in neither dataset alone, only in their combination.

Across datasets with no match at all. The capability infers properties of people, companies or places it has never observed, from similar ones observed elsewhere or from patterns a model has learned. The answer is not in the data at all; it is inferred from entities like it.

How is enrichability different from linkability?

Linkability is the older mechanism: joining two records through a shared key that identifies the same entity in both, such as a loyalty card number, a device identifier or a company registration number. For most of the history of data businesses it was the binding constraint. Companies competed on whose databases could be joined to whose, and whoever owned both sides of a join kept the value.

Linkability still works, and many data businesses run on it alongside enrichment: a retail media network joins purchases to a loyalty card, then models and targets on top of the join. Its limit is the key. No shared key, no join. Enrichment needs no shared key, and in its most consequential forms it does not need the two datasets to describe the same entities at all. Linkability matched what was there. Enrichability creates what was not.

The shift matters because the capability to work on content without keys has spent two decades getting cheap. A dataset that could once be used only for what it recorded, or for what it could be joined to, can now anchor uses its owner never planned. The same is true of the data a business licenses out.

Why does the capability matter more than the data?

Because the capability is what a competitor cannot copy quickly. Data can often be bought, licensed or collected again. A capability stack, the people, methods, tools and client knowledge that turn data into an answer, is built over years, in a sequence that cannot be skipped: you cannot model what you have not validated, or package what you have not modelled.

That is why buyers of data businesses pay for the capability. The two acquisitions in the register that disclose both price and revenue, S&P Global's $1.8 billion for With Intelligence and LSEG's agreed $3.535 billion for Octus, come to about 14 and 12 times revenue. S&P Global paid for proprietary data, and LSEG has agreed to, but both priced it as a running business: With Intelligence serving about 3,000 clients, with annual contract value growing in the high teens; Octus serving 750 customers and more than 40,000 users. The data came with the capability that keeps those clients. Chegg, the education company, made the same point from the other side in May 2026, when it announced it would license to AI labs not only a decade of step-by-step solutions but also its calibrated network of subject experts, and stressed the calibration, quality assurance and auditability of their work: the capability, offered alongside the data.

Parts of the capability are also getting cheaper. Circana's model runs on free code, and some work that once needed a dedicated data business can now be done by a small team in months. The client knowledge, and the sequence in which the stack is built, still take years. For data holders this cuts both ways: more of them can do the work on their own data, and more buyers can do it on data they license, and keep what it produces. Who captures the value takes that question up.

Is a data product the same as "data as a product"?

No. The terms sound alike and mean different things. A data product, in the commercial sense used in this guide, is something a buyer pays for: a benchmark, a feed, a score, an answer. "Data as a product" is an internal management practice, popularised by the data mesh approach, in which teams treat the datasets they share inside a company as products, with owners, documentation and quality standards. The practice helps a company use its own data well, and it can be a step toward a commercial data product, but it is not one.

Data-as-a-service is a third idea: a way of delivering data, as a continuous feed or query service on subscription, that can sit underneath a data product or under plain licensing.

Questions people ask

What is a data product? An answer a buyer pays for, built from data by work the buyer does not have to do: a benchmark, a score, a forecast, a placement, an insight delivered inside the buyer's own tools. PhonePe's PulsePro, launched in 2026, and Octus's structured credit data are examples.

What is the difference between raw data and a data product? Raw data records what happened; a data product answers a question someone pays to have answered. The difference is the work in between: enrichment, which produces something the data does not state, and product work, which puts it in a form the buyer acts on. Whoever does the work captures most of the value.

What is enrichability? The capacity of data to yield commercial value when a capability works on its content to produce something the data does not itself state: from one dataset alone, across datasets matched without a shared key, or by inference from similar entities elsewhere. Reformatting or repackaging does not count.

What is the difference between enrichability and linkability? Linkability joins two records through a shared key identifying the same entity in both, such as a card number or a device identifier. Enrichability needs no shared key: it matches entities by probability, or infers properties from similar ones, and produces outputs neither dataset states.

Why do data businesses sell for high multiples? Because buyers pay for the capability that turns data into answers clients renew, which takes years to build and cannot be copied quickly. The two data-business acquisitions in the register that disclose both price and revenue came to about 14 and 12 times expected annual revenue (the second, LSEG's agreed purchase of Octus, is due to complete in 2027).

How do you build a data product? Start from the buyer's decision rather than from the data: what the buyer needs to decide, what it is worth to get right, and what work turns your data into that answer. The Data Monetisation Readiness Assessment on Monetaise is a structured first pass.

---

Next in the guide: Who captures the value, why the party that does the work keeps the money, and what the rake is. Every deal cited here is in the register, each with its reading.