Notes

Data monetisation is like teenage sex.

Everyone talks about it. Nobody really knows how to do it. Everyone thinks everyone else is doing it, so everyone claims they are.

The few who actually do it are usually selling something other than the data.

(With apologies to Dan Ariely, who said it first about big data.)

"Data is the new oil." I have heard it in every data boardroom I have sat in, and it is wrong in the one place that sets the price.

Oil is consumed. A barrel burned is a barrel nobody else can burn, and the customer has to come back for the next one. The physics does the seller's work.

Data is not consumed. Your buyer uses it and still has it. So can the buyer's analyst, the buyer's model, and the buyer's next vendor. A file delivered three years ago can still be answering questions today, paid for once.

That is why "they only use it internally" protects nothing. Use is not the risk. Use is the leak. Almost every use leaves something behind that you do not own: a copy, a calculation, a model that needs you a little less.

An oil company never had to write down what the customer may do with the oil once it is burned. A data company has to write down almost nothing else. And the clauses only help at the edges.

So stop pricing data like a commodity. Oil is protected by being used up. Data is protected by what you refuse to hand over, and by what you can still withhold next year.

Data is not the new oil. Oil was easier to sell.

"Flat fee, five years, everyone is happy."

That is a bet. One party learns what the asset is worth, and it is not you.

The buyer uses your data every day. It learns which parts move its numbers and what it would cost to replace you. You know the invoice.

Five years of that is specification debt: everything the buyer has learned about your asset that you have not. It comes due at renewal, when the buyer walks in knowing the price and you walk in guessing it.

The flat fee felt safe because you never had to measure anything. The buyer measured everything.

My bet: most flat-fee data licences signed this year will renew at a discount, or not renew at all.

Stack Overflow licensed its answers to Google and OpenAI. Then the questions stopped.

In 2024 both signed up to its new data API. No price was disclosed. The CEO said what made the data worth paying for: it is "peer-curated, it's voted by the community, and only the accepted answers are ultimately displayed."

The text was never the asset. Anyone could read the text. The votes were: developers telling each other, every day since 2008, which answer actually worked.

In December 2025 Stack Overflow received 3,862 new questions. At its peak it took more than 200,000 a month.

The money looks fine. Revenue went from $98m to $129m in two years, and its owner, Prosus, credits the data deals with most of the first year's growth.

But a licence prices an archive. The votes were never an archive. They were a flow, and the flow came from developers who now ask a chatbot instead.

Stack Overflow is being paid for its past by the models that are emptying its future.

Data nobody maintains any more is compiled data. Compiled data goes first.

"AI makes proprietary data more valuable."

On two conditions.

Your data is good, and nobody can substitute it.

And you have good AI yourself.

Miss the first, and the AI companies buy someone else's data. Miss the second, and they take the value of yours.

One bank paid Delta $8.2 billion last year. That is what a bargain looks like.

Delta reports $8.2 billion of remuneration from American Express in 2025, about an eighth of the airline's revenue, and expects $9 billion this year. At the centre of it, cardholders earn SkyMiles on what they spend, and the bank buys those miles from Delta at a price set in the contract.

Both sides can count the miles. The price per mile is written down and the number of miles is on the statement, so neither side has to guess what the deal is worth.

A bet works the other way. A fixed fee buys access, the buyer learns what the data was worth, and the seller finds out at renewal.

Because the payment is measured, it grows with the spend on the card. The agreement was signed in 2019 and runs to 2029. The $9 billion needs no new negotiation, only more spending.

Measure it, and the payment grows on its own.

Kantar built the panel it sells. That is why nobody can take it back.

Many data businesses run on someone else's data: a retailer's loyalty card, a bank's transactions, a publisher's archive. The partner does the learning, the owner holds a claim, and one day the owner can call it.

Worldpanel's data was never anyone else's. It recruited tens of thousands of British households, gave them scanners, and had them record the shopping they brought home. The households are paid in vouchers, not in a share of what the panel is worth. And Worldpanel never sold the panel raw; it sold what the panel showed, after calibration.

When I interviewed Guillaume Bacuvier, who ran Kantar Worldpanel, I asked how it could sell the same categories to brands that compete with each other. His first answer was ownership: the data was owed to no one. No revenue share, no profit share.

I call this owner-enriched by construction. One way to own the whole rake is to find data nobody else has. The other is to make it.

The moat can be built, not only found.

AI is making your data cheap in four ways.

Three made headlines. The quietest never will.

It arrives as a polite email about budget.

Your client licenses 400 fields and uses three. An analyst with an agent rebuilds those three in an afternoon. The renewal comes back smaller, and nobody announces anything.

My bet: compiled data goes first.

If your moat only worked while compiling was expensive, you no longer have a moat.

The other three, each with a deal behind it, are in today's newsletter. Link in the first comment.

A retailer's data can work twice with the same brand.

Once in retail media, where the brand pays to reach the retailer's shoppers.

Once at the negotiating table, where both sides plan the category from the same facts.

Same data, two very different uses. Many retailers run neither.

In my whole career, I have never sold raw data.

There's very little money in it. Unless your data is exclusive and can't be substituted, the seller is on the losing side of the deal.

The money is in what you build on top of it.

"We could bring it in-house any time."

Tesco did. The last slice cost nearly five million pounds a point, because the rake was on loan and the loan had grown.

You get the data back. The value stays with the partner who learned your customers.

The Clubcard data was always Tesco's. dunnhumby did the learning on it, and Lord MacLaurin, then Tesco's chairman, told the board: "you know more about my customers after three months than I know after thirty years."

Tesco never needed the data back, because it already had the data. It had to buy the company: 53 percent in 2001 for a reported £30m, 84 percent by 2006, and when the founders sold their last 10 percent at the end of 2010, it fetched a reported £48m. Per percentage point, the last stake cost more than eight times the first.

The gap was not Tesco's customers alone. In those nine years dunnhumby took what it had learned on Tesco's data to other countries and other clients, and the company's price rose with it. Tesco paid the market price for value grown on its own data.

Your partner's rake is on loan from you. It sits on your data, and when the data is yours alone, you hold a claim on it. The claim is real, and it depreciates. Every year the partner learns, calling it costs more, and the part you get back for free is the part that was never worth much on its own.

How much would it cost you to leave?

Price what they have learned since you signed.

Google now pays more than 400 news publishers for AI use of their journalism. That is not a market. It is a panel.

On 15 October Google said its News AI pilot had passed 400 publications, with CNN, the FT, Axios and Forbes among the new names. No price is public.

It does not need to be. The buyer is also the company that decides how much search traffic each of those 400 publishers gets tomorrow. Nobody negotiates hard with the hand that ranks them. Google does not have to dictate the price. It is the only party at the table that could.

Nine days earlier, six scientific publishers went the other way through Redpine, a Swedish start-up: each publisher sets what an AI agent may use and what it pays per transaction.

My bet: when the pilot cheques stop, the publishers still earning from AI will be the ones who set their own price.

Publicis agreed to buy LiveRamp at a 29.8% premium. It can buy the platform. It can't buy the trust.

Part of what made LiveRamp valuable: no agency group controlled it.

More than 25,000 publisher domains and over 500 technology and data partners connect through it.

LiveRamp owns the technology. It doesn't own the relationships that make the network worth joining.

That is a rake on loan.

Publicis already owns Epsilon and Lotame. Its promise: LiveRamp stays "a neutral, interoperable platform." No customer restricted.

I believe them. It doesn't matter.

A pledge is not an architecture. Whoever owns the platform sets its prices, its roadmap and its priorities.

Every rival already owns its own matcher. Omnicom has Acxiom. Dentsu has Merkury. WPP has InfoSum.
LiveRamp was where they could all meet.

My bet: the pledge is kept, and rivals steer their clients elsewhere anyway. Not because access is denied. Because a competitor now owns the room.

Neutrality is not a feature you can acquire. It's a fact about who owns you.

I have data, I will make money.

That was the title of my lecture yesterday at my alma mater, Università Cattolica del Sacro Cuore, for the students of the Master in AI.

The sentence hides the mistake most data owners make. Nobody pays you for having data. Value goes where the work is: to whoever turns the data into a decision someone will pay for. Own the data, leave the work to someone else, and you capture a fraction of what your own asset is worth.

I gave a similar session a year ago, and the material has changed more than I expected. The reason is the book I'm writing, where I'm pinning down the rules behind data monetisation from 20+ years in the seats where data was bought and sold.

This year's lecture was built on it, and testing the rules on a room of sharp students from all over the world was a privilege, and a useful stress test.

Thank you Massimo Chiriatti for the invitation, and thank you to the students for your time and curiosity.

"Data product" is the most expensive phrase in the enterprise.

It sounds like a business. Too often it is a table with a roadmap.

The data-mesh playbook says start with the consumer. In practice, too many programmes start with the data they already have and go looking for someone to use it. That's how a catalogue fills up with products nobody uses.

Internal or external, direct or three hops downstream, the test is the same: which decision, made by whom, does this change?

If your team can't answer that in one sentence, you haven't got a product. You've got expensive inventory.

Publishers signing with OpenAI are being paid for their content. They're also handing over the front door.

In September it was BCCL, owner of the Times of India, and the Indian Express Group: reportedly around $5 million and $3 million a year. News Corp's 2024 deal was reported at more than $250 million over five years, including technology credits.

What ChatGPT shows: their reporting and archives, summarised, with a link back.

What they hand over: the moment the reader asks the question. When it's asked in ChatGPT, it happens in someone else's product. A link back is not a visit.

Saying no doesn't bring the reader back either. That's the trap.

Renewal season has already started elsewhere. Reddit has reportedly been negotiating a successor to its $60 million a year Google deal, and discussed ending Google's AI access.

My bet: none of these deals renews at a higher annual fee for the same rights.

UMG just signed the Spotify deal again, this time one layer up.

On 10 September it announced a multi-year licensing and product deal with ElevenLabs, the AI audio company. Fans will remix, mash up and reinterpret music from participating artists and songwriters on a new platform.

ElevenLabs will launch it and operate it.

The labels got paid by Spotify. Spotify got the listener.

Now the fan who cares enough to remix a song will do it on someone else's product too. Every remix will show which songs, which artists and which sounds fans actually want to play with. That data will be created on ElevenLabs' platform, not UMG's.

A rake on someone else's platform is a rake on loan.

Infillion is buying Foursquare. The trade press read it as ad-tech consolidation.

I read it differently.

Foursquare spent a decade doing what every sensible data business does. It did the enrichment itself. It sold direct. It kept its share.

It still ended up one link short of the buyer's decision.

This week's newsletter is about that one link: where it sits, who is about to hold it, and the question every data business should ask of its own.

When everyone with a car park has a media network, nobody has one.

McDonald's wants a $1 billion ad business. In August it started piloting ads on the drive-thru order boards of 450 of its company-owned US restaurants.

Simon, America's largest mall operator, launched a media network in late August. Circle K relaunched its own the same month.

Amazon, Walmart and a few others will be fine. My bet for the rest: within 5 years, most will be running on someone else's platform, or gone.

Best Buy has already chosen. From February it will sell ads on its own-brand TVs through Amazon's ad tech, on Amazon's TV software.

Retail media was pitched to boards as a high-margin business of their own. For most, it will be a rake on loan with their logo on it.

We charged a few hundred euros for years. The customer was willing to pay two thousand.

The directory business I worked in sold a listing for a few hundred euros a year. Then we built a website product on the same data and sold it for two thousand.

Nothing underneath had changed. Same advertisers, same categories, same records the company had kept for decades.

Here is the part that took me a while to understand. Two thousand euros was roughly what the average Belgian advertiser spent with us in a year, on everything. You would assume the money moved across, that they bought this instead of something else. A lot of them did not. Much of it was new spend, sitting on top of what they were already paying.

So a few hundred euros was never what the listing was worth. It was the most that shape could reach. What the customer was willing to spend sat somewhere else entirely, and the company had read one as the other for years before I got there.

That is what under-capture looks like from the inside. It does not arrive as a loss. The revenue line is fine, nobody complains, and you never find out what the customer would have paid because you never put anything in front of them that could carry it.

The company eventually filed for Chapter 11 anyway. An underpriced asset and a broken balance sheet were never the same problem, and fixing one does nothing for the other.

Spirit sold its customer data for ten million. Google kept the rest.

Spirit Airlines put its customer data up at auction and Google took it for ten million dollars, paid once, at closing, with no claim on anything built from it afterwards. Google handles the enrichment. Whatever that corpus turns out to be worth inside Google's models, it is worth it to Google.

Ten million looks like the price of a corpus. It is the price of a corpus somebody else still has to finish.

In the same week, CuriosityStream told its investors that licensing its library to the AI labs was running at around seventy-three percent gross margin, and that it intended to keep going. It clears the rights, builds the datasets and deals with the labs itself, so there is nobody in the middle taking a cut of the work.

Almost the same product, and the same kind of buyer. What separates ten million paid once from seventy-three percent paid over and over is which of the two had already done the work that makes a corpus sellable.

By August, Spirit had no move left. That was settled years earlier, in whatever meeting decided rights clearing and curation were somebody else's job, and a bankruptcy auction is not where you revisit it.

Value goes where the work is. The work gets done long before the buyer turns up.

The headline prices held for seventeen years. What could be charged for changed.

NYSE charges the same three numbers for OpenBook today as it did in 2009. Five thousand a month for access, sixty for a professional user, fifteen for a qualifying nonprofessional. They are in the March 2009 SEC order and they are on the pricing guide dated 14 May 2026.

The title of that 2009 order tells you what was actually going on. It approves a nonprofessional fee, and it revises the unit of count that determines the device fees payable by data recipients. Even then, the pricing decision included deciding what counts.

The order defined subscribers to include both individual people and devices. On today's schedule there are three non-display categories at six thousand dollars each, and a redistribution fee of three thousand, sitting on the same page as the sixty dollar user.

None of it is buried. It is enumerated, with product codes, on a public document. This is a deliberate pricing decision and the seller is the one who makes it.

In the negotiations I have sat through, the rate gets the argument. What counts as a chargeable use gets almost none.

Before you negotiate the price, find out what you can be charged for.

Three prices for the same company in six years: 6.9 billion, 12.5, 7.7. Nobody did anything wrong. A sentence was deleted.

For twenty-four years, every company that wanted to be paid by the US government needed a number only Dun & Bradstreet could issue. The government paid about nineteen million a year for it. That was never the money.

Then the government started writing its own number. No revenue line disappeared that day. The price did, one renewal at a time.

Most data businesses run on a key they did not make. This edition is about how to find yours, and what it is worth to whoever holds it.

RBS owned the data. Its partner owned the enrichment and the customer. The partner just sold for $230 million, cash.

The data stayed exactly where it was. The price went to the layer built on top of it.

Reward spent more than ten years building card-linked rewards on RBS transaction data, and ran the same model across dozens of banks. The offers, the merchant deals, the programme the cardholder actually sees: all Reward's. In February Rezolve, a Nasdaq-listed AI commerce company, paid $230 million for it. Not for the data. Reward doesn't own the data, it owns everything that makes the data worth acting on.

I know this structure from the inside. I was COO at Reward for many years, when we launched the RBS programme, and later CPO at dunnhumby. The data owner watches a partner build on top of their substrate and thinks the arrangement is secure because they control the supply. It is secure until someone prices the enrichment.

The rake goes to the party that owns the enrichment and the customer relationship, not the party that owns the data underneath it. If a buyer priced your data partnership tomorrow, which side would they pay for?