Marketing needs a spine

A marketing measurement question sounds simple: did the spring campaign sell more breakfast cereal? Answering it means walking a chain. This money bought this placement. This placement was seen by this person. This person bought a box at a supermarket on a Tuesday. That box was worth six dollars fifty at the checkout and less to the company that made it. There are four joins, and the answer is only as good as the weakest of them.

That chain is the measurement spine. Nearly all of the argument in our industry happens above it, about which method should allocate the credit. The spine itself gets little attention, and it is usually where the problem is.

The chain holds only if the business has agreed what a customer, a campaign and a sale are. Where it does not hold, the result is a number nobody can act on.

The same chain breaks in other industries, at different joins. A mortgage ad is clicked in March and the loan settles through a broker in June, so the lender sees the settlement and not the click. A superannuation fund sees a rollover arrive weeks after a campaign, with no record of what prompted it. A gym sees a membership sign-up and waits a year to learn whether the member was worth acquiring.

Three methods dominate marketing measurement, and each copes with the broken joins in a different way. We use all three. The useful question is which links in the chain each one observed and which it inferred, and most measurement stacks cannot answer it.

Probabilistic models are a black box

As identifiers disappear, more of the chain is estimated and less of it is observed. The platforms fill the gaps with their own models. When a visitor declines cookies, Google estimates the conversions it can no longer see and adds them to the advertiser’s reports. Google says the modelling recovers more than 70 percent of the click-to-conversion journeys lost to consent choices.⁴ The advertiser sees the result, and none of the assumptions, the training data or the error around it.

A modelled number can be accurate. Without access to the model it cannot be checked, and it comes from the company that sold the media. Open models show another way. Google’s marketing mix model, Meridian, is open source, so an analyst can read its assumptions and run it on their own data.⁵

The platform marks its own homework

The platforms that sell the media also decide what counts as a result. In March 2026 Meta changed its rule for crediting a sale to a video ad. A purchase within a day of a view used to need ten seconds of watching to earn credit, and it now needs five.³ Meta said the shorter threshold better reflects how people watch short video, and cited its own figure that 46 percent of online purchases from Reels ads happen within the first two seconds of attention.³

Meta has not published how many more conversions the new rule brings into its reports. Viewing falls away quickly in the first seconds of a video, so more views now qualify, possibly many more. In the same update Meta stopped counting likes, shares and saves as clicks and moved the conversions that follow them into a separate category. The net effect for any one advertiser can only be measured from outside the platform.

Meta sets the rule that decides what its own media is credited with, and can change it when it chooses. Every platform that reports on its own sales is in the same position, and each has an interest in a larger number.

A number you cannot act on is reporting

A mix model returns a coefficient for paid search. That is a real finding. Acting on it means deciding which campaigns, which locations and which audiences change on Monday morning, and the coefficient is expressed in a unit nobody can address.

To act on it you have to travel back down the spine to the objects you can change: a budget on a campaign, an audience, a piece of creative, one product in one retailer. If the measurement was never built on those objects, the trip back down is guesswork.

The system that measures and the system that acts have to describe the same things. Otherwise every action is a translation of the measurement, made by a person, in their head, at speed, on a Monday.

Ontology is a heavy word for something plain. It is a written, agreed statement of the things a business measures and acts on: what each one is, what distinguishes one from another, which links between them are allowed, and what evidence has to be present before a link can be asserted.

A schema records that a column holds a string, and a semantic layer enforces a definition when someone runs a query. The ontology is the definition being enforced. Three parts of it hold the spine together.

Identity is the weakest join

The usual approach is a customer key computed inside whichever source system had the best data: a hash of phone number, falling back to email, falling back to name. It works, and it cannot be audited, corrected or extended, because the logic lives inside a vendor’s product.

The alternative costs one more table. A party is the resolved person. Identity rows are the evidence for that resolution: an identifier type, a hashed value, the source it came from and a strength. Email, phone and loyalty number are strong, and two parties sharing one are merged. A device ID, a session or a click ID is weak, and attaches to a party without ever causing a merge.

Weak identifiers are the ones that decay. Cookies expire, click IDs are stripped and a tablet gets handed to somebody else. Held as evidence with a strength, they let the graph thin out as they go, instead of merging two people who once shared an iPad.

Consent sits on the person, because the action does

Consent should be an attribute of the party, scoped by purpose such as marketing or measurement, and carrying the source it came from and the time it was captured.

Modelled per source instead, consent becomes as many regimes as there are connectors, with no single answer to whether a record could be used and no way to honour a withdrawal across systems. That is a legal problem and an operational one, because the far end of the spine is an action aimed at a person. A stack that can measure someone it is not allowed to contact has a fault built into it.

Every link records how it was made

The join between an exposure and a transaction needs a field recording how the join was made: identity match, order reference, geography or model.

That one field does two jobs. It lets you report which parts of the spine were observed and which were estimated, which is the honest form of an attribution number. It also lets two methods run side by side over the same data, so when you change method you can see how much of the movement came from the world and how much from the arithmetic.

Why now

Every year less of the spine can rest on a technical identifier, and more of it has to rest on a person the business already knows. That turns identity resolution from an optimisation into a requirement. The legal question is getting sharper too, and being able to state why a record could be used is a question about entities before it is a question of policy.

The interface has changed as well. When someone asks a system a question in plain English, the answer is only correct if the words map to defined things. “How many new customers did the spring campaign bring in?” cannot be answered until customer, new and brought in each have one written meaning. Language models made the interface easy and made the ontology load-bearing.

Consumers should be protected from being followed around the internet, and governments are right to act. How they do it decides who is left able to prove that advertising works.

The evidence so far points one way. When the GDPR took effect in Europe in 2018, websites cut the number of third-party technology vendors they used for European visitors by 15 percent. They kept the biggest vendors, concentration in that market rose by 17 percent, and the shares held by Google and Facebook rose with it.¹

When Apple required apps to ask before tracking people across other companies’ apps, Meta estimated the change would cost it about $10 billion in 2022.² Meta could absorb that, because it still sees what billions of signed-in people do inside its own apps. A newspaper, a broadcaster or an independent ad network has no equivalent, and needs data to move between companies to show an advertiser what its audience bought.

A form of regulatory capture

A platform with billions of signed-in users holds its identity graph inside its own walls. Rules that stop data moving between companies leave that graph intact and cut off everyone who has to join data across companies to prove a result: traditional publishers, emerging media businesses and the independent technology that serves them. Advertising money follows proof, and the proof would sit with the platforms.

We think the effect is a form of regulatory capture, whether or not anyone intends it. Rules written to protect consumers would protect the position of the largest platforms, and leave their competitors unable to prove what their ad products deliver. A press that cannot show what its advertising is worth struggles to fund itself, and a weaker press leaves a democracy with fewer independent sources of information.

Australia is in the middle of this decision. In 2023 the government agreed in principle to define and regulate targeted advertising, and as of early 2026 it had given no timeline for the second tranche of reform that would do it.⁶ The ACCC has already found that Google holds a dominant position across the ad tech supply chain, and recommended measures to make that market more transparent.⁷

Openness is the sustainable path

Openness gives consumers protection without handing the proof to the platforms. Consent can be held on the person and honoured across every system that touches them. Each join can record how it was made. Models can publish their assumptions, and measurement can be built so that an advertiser, a publisher and a regulator can all inspect it.

In an age of AI and data sovereignty, that path leads to more control and more choice for consumers. Marketing regains its spine when a business can see the return on its advertising directly, without taking a platform’s word for it, and when every media owner, large or small, can prove what its audience is worth.