Data valuation means putting a number on what a dataset is worth. It sounds like it should be straightforward, but a review of the research by the Bennett Institute at the University of Cambridge concluded there is still no consensus method for doing it (Diane Coyle and Annabel Manley, Bennett Institute, University of Cambridge, July 2022). What the methods do give you is a way to reason about price before a buyer names one.
It also helps to know that most company data has no book value. Under the international accounting standard for intangible assets, internally generated items such as customer lists are not recognised as assets (IFRS Foundation). Your support tickets and project history are worth something only once someone wants to pay for them.
The short version. Use the cost method to set the lowest price you'd accept, market comparisons to see a realistic range, and the income method only if you'll be paid for refreshes over time. Then let competing offers tell you where in that range you are.
The three standard methods
The research literature divides most approaches into cost-based, market-based and income-based methods (Diane Coyle and Annabel Manley, Bennett Institute, University of Cambridge, July 2022). Here is how each applies to a dataset you might license.
| Method | The question | How to apply it | Where it misleads |
|---|---|---|---|
| Cost | What did it cost to create, or what would it cost to recreate? | Add up the staff time, tools and storage behind the records, or estimate what a buyer would pay contractors to produce similar examples. | Cost isn't value. Old records cost little to keep, and the work behind them was paid for to run the business, not to sell data. |
| Market | What have similar datasets sold for? | Compare with reported deals and marketplace listings for comparable data, adjusting for size, quality and exclusivity. | Few prices are public, and the famous deals are platform-scale content licenses that don't resemble a company's internal records. |
| Income | What will it earn? | Estimate the license fees over the term, including refreshes, and discount them for timing and risk. | Depends entirely on assumptions. With no track record, the forecast can say whatever you want it to. |
Cost: start with your walk-away price
The most useful cost figure isn't what the records cost to create. It's what the deal will cost you: legal review, the time of the people who scope and approve it, the cleanup, and the review of samples before delivery. That total is your floor. An offer below it loses money however good it sounds.
Replacement cost, meaning what a buyer would spend to have contractors produce similar examples, is worth thinking about too, because it is the buyer's alternative. It usually sets a ceiling rather than a price, since contractors can't fully recreate years of real history.
Market: compare with deals that resemble yours
The closest public comparisons for an operating company are archive sales by closing startups, reported at $10,000 to $100,000 per company (Gizmodo, reporting on Forbes, April 17, 2026). Platform deals sit far above that: Reddit disclosed $203.0 million in AI data licensing contracts over two to three years (Reddit, Form S-1 filed with the SEC, February 22, 2024). Use the first kind as a reference point and the second only as a ceiling. Our guide to what company data is worth lists the reported figures.
Income: only if the money keeps coming
A one-time sale has no income stream to value. A license with scheduled refreshes does: new records delivered every quarter or year for a fee. If that's on the table, estimate the fees over the term, then discount them for the chance the buyer ends the contract early. It's also a reminder that a refresh deal can be worth more than a bigger one-time payment.
How the buyer sees it
Buyers value data by what it does for their models: whether it fills a gap, how much it improves results on the tasks they care about, and whether they can get something similar elsewhere. Researchers have methods for measuring how much individual examples contribute to a model, but you won't see those numbers. What you can control is how clearly you describe what's in the dataset, because a buyer that understands it is more likely to price it well.
What raises the value
- Complete records with outcomes. Tickets with resolutions, code changes with tests, deals with results.
- Years of history in one system. Consistency and volume matter more than breadth.
- Specialized work. Records few other companies could supply.
- Clean rights. Contracts and privacy notices that allow the use, and a clear redaction plan.
- Non-exclusive terms. The ability to license the same records to more than one buyer.
- Refreshes. New records on a schedule.
Mistakes to avoid
- Pricing from headlines. Platform deals tell you the market exists, not what your records will fetch.
- Treating cost as value. Ten years of storage costs don't make ten-year-old records valuable.
- Forgetting the cost of the deal. Legal and cleanup time come out of the price.
- Giving exclusivity away. If a buyer wants it, price it separately and put an end date on it.
Once you have a floor and a range, the licensing agreement is where the price gets protected: acceptance terms, milestones and what happens if the buyer walks away.
Common questions
What is the best data valuation method?
There isn't a single accepted one. Researchers group the methods into cost-based, market-based and income-based approaches, and each answers a different question. For a dataset you plan to license, use costs to set your floor, market comparisons for a range, and income only if there will be recurring fees.
Is our data an asset on the balance sheet?
Usually not, if your company created it. Under IAS 38, internally generated items like customer lists are not recognised as intangible assets, so most company datasets have no book value until someone pays for them. Purchased data can be different. Ask your accountant how licensing income should be recorded.
Can a buyer tell us what our data is worth?
A buyer can tell you what it will pay, which is not the same thing. Its offer reflects what it is building and what it already has. That is why it helps to know your own floor, and to compare more than one offer if you can.
Sources
- Diane Coyle and Annabel Manley, Bennett Institute, University of Cambridge, July 2022: What is the value of data? A review of empirical methods
- IFRS Foundation: IAS 38 Intangible Assets
- Reddit, Form S-1 filed with the SEC, February 22, 2024: Reddit, Inc. registration statement
- Gizmodo, reporting on Forbes, April 17, 2026: Failed companies are selling old Slack chats and email archives to train AI