Measurement Is a Mess. AI Gives It Shape and Purpose

For most of my 25 years working in media, data, measurement and technology, I have believed that measurement is the most important thing in advertising. I still do. I have a phrase I use rather too often: the direct-response budget for unmeasurable media is zero. It is deliberately provocative, but the underlying point is serious. Money follows confidence. Google and Meta did not become advertising giants simply because they accumulated enormous audiences. They also persuaded advertisers that money put into their machines could be connected to results. As cookies and mobile identifiers became less reliable, direct first-party conversion signals strengthened that position further.

I came to this view from several different directions. Early in my career I worked across search and display, at a time when even doing both was relatively unusual. Search was intoxicating because it appeared so measurable. Someone searched, someone clicked, someone bought something. Years later at MediaCom I was able to look across the entire media mix. I saw the real power of television, worked with media mix modelling and helped social become part of the agency proposition. Since then I have worked through my own businesses with some of the largest advertisers in Britain and globally. The broader my view of marketing became, the more important measurement looked — but also the less clean it became.

That is the uncomfortable bit. People who want to be data-driven don't like messy data. Unfortunately, marketing measurement is messy. A CEO or finance director quite reasonably wants to know whether the money worked and where the next million pounds should go. They do not particularly want six methodologies, four caveats and a spirited argument between the search team and an econometrician. So organisations gravitate towards answers that look definitive. Last-click attribution is a wonderful example. Almost everyone sophisticated enough to use it knows it is wrong as a description of causality, yet enormous decisions are still made from it because it produces a number. A clean wrong answer can be psychologically much more attractive than a complicated, uncertain one.

The reality is that good measurement systems often disagree. Your media mix model may tell one story about Meta. Meta's own incrementality testing may tell another. A geo holdout might produce something different again. Attribution may insist that all three are idiots. None of this necessarily means measurement has failed. Different methods observe different things and carry different weaknesses. A model built from relatively flat digital spend may struggle to identify an effect that becomes clearer when spend varies. A platform experiment may be methodologically excellent while still leaving you conscious that the platform is, to some extent, marking its own homework. Attribution can provide wonderful tactical signals while being a dreadful answer to the question of what would have happened without the advertising. All models are wrong; some are useful.

The important thing is that the mess has a shape. One of my first mentors had learned his trade building systems to predict the weather. Weather is spectacularly messy: thousands of signals, incomplete information, interacting systems and no possibility of knowing precisely what is going to happen everywhere. Yet we model it because getting better at understanding that mess is enormously useful. At the trivial end, a better forecast means you are less likely to get caught in the rain without a brolly. At the serious end, forecasting can be literally life or death — for ships at sea, or increasingly in predicting the conditions in which forest fires can start and spread.

Nobody expects the forecast to be perfect before acting on it. We expect it to give the mess enough shape to make a better decision. Marketing measurement should work the same way. If one methodology says YouTube is highly incremental and another suggests a much smaller effect, the interesting question is not simply which number should go into the board deck. It is why might both results look the way they do? Perhaps one method is missing variation. Perhaps another is affected by audience selection. Perhaps the answer differs by customer group, geography or campaign type. From that mess you can form a narrative — or, more properly, a hypothesis — about what is really happening.

Crucially, that hypothesis has to be useful. It should help you predict what happens next and tell you what to do. If we believe YouTube is genuinely incremental for a particular audience, but not to the degree platform attribution suggests, we can increase spend in carefully selected areas while running an independent holdout elsewhere. Reality gets another vote. We learn something, improve our understanding and make the next decision from a slightly better position. Great weather forecasting does not control the weather; it helps us make better decisions in a world where the weather remains uncertain.

Doing this properly requires much more than measurement expertise. You cannot really understand paid search performance unless you understand the difference between brand and generic search, auction dynamics and how campaigns are actually being managed. You cannot interpret television properly without understanding how television is bought, reach, scheduling and the realities of broadcaster inventory. The same applies to social, display, retail media and every other discipline. I have spent much of my career deliberately working across channels and I have never met the person who possesses practitioner-level knowledge of all of them. The poor person responsible for measurement is therefore being asked to understand a system whose detailed operation no single human can completely hold in their head.

This is where AI changes the equation. For years marketing has become progressively better at producing measurement. We have attribution systems, incrementality studies, econometrics, dashboards, brand studies, experiments and more data than any sensible human being could want. Yet a remarkable amount of useful analysis withers on the vine. A study gets commissioned, turned into a deck, presented in a meeting and gradually becomes an archaeological artefact in SharePoint. The missing step has been turning all that measurement into continuous understanding. AI gives us a way to combine different forms of evidence, identify where they agree, explain where they conflict, incorporate context and maintain several possible explanations rather than forcing everything into one reassuring number.

So the first transformation is measurement into insight. But insight is not enough either. I am not particularly excited by an AI system that studies everything and announces that we should increase YouTube spend by 14 per cent. That is a prettier spreadsheet. A useful recommendation needs to understand the channel in detail. It might say that the evidence supports increasing YouTube investment against particular audiences, using particular tactics and formats, while explaining that the incrementality evidence conflicts with the media mix model. It should then recommend an independent geo holdout alongside the increase, select appropriate regions and establish what result would increase or decrease our confidence in the hypothesis. That is insight becoming action.

Crucially, this does not mean replacing the people who understand marketing. It means finally being able to use more of what they know. The best search specialist has knowledge that the modeller does not. The best TV planner knows things the search specialist does not. A strategist, social specialist and measurement scientist will all see the same evidence through different lenses. Historically we tried to combine this intelligence through meetings, decks and organisational memory, with results that depended rather heavily on who happened to be in the room. AI allows those perspectives to be captured, applied repeatedly and even allowed to disagree. Instead of pretending there is one omniscient marketing brain, we can maintain competing hypotheses and ask what evidence would help us choose between them.

Speed matters too. Marketing decisions do not politely wait six weeks for an analysis project to conclude. Money is being spent tomorrow morning. Creative is being produced. Audiences are changing. Competitors are moving. A brilliant answer that arrives after the decision has often lost most of its value. AI can compress the journey from data to analysis, from analysis to insight, from insight to recommendation and from recommendation to executable action. That does not mean letting a language model do arithmetic it should not be doing. Deterministic code, statistics and appropriate machine learning should still do the maths. Generative AI is valuable because it can orchestrate those tools, reason across their outputs, combine them with expertise and help decide what should happen next.

And this is where the obsession with perfect measurement becomes actively unhelpful. We do not need perfection. We need to level up. Better understanding. Better use of specialists. Better consistency. Better speed. Better tests. Better decisions. Moving from a mediocre decision made with false certainty to a stronger decision made with an honest understanding of uncertainty is progress. Turning a one-off study into knowledge that influences every subsequent campaign is progress. The mess remains. We simply understand its shape better.

Measurement is not the entirety of marketing. Great creative and talented people remain fundamental, and in many cases they create the demand that measurement is trying to explain. But better intelligence should improve creative too. It can tell us which messages appear to resonate, where different audiences respond, which hypotheses deserve new executions and where a surprising result is worth investigating. Creative gives us new things to put into the world; measurement tells us something about what happened; intelligence helps us decide what to make and do next.

That, to me, is the real opportunity for AI in marketing. Not perfect attribution. Not a machine that magically knows the correct media plan. Not another dashboard with a chatbot stuck on the front.

It is the possibility of creating a continuously learning marketing system:

Measure → understand → hypothesise → act → test → measure again.

Every action creates more evidence. Every test gives the mess a little more shape. Every cycle improves the next decision.

Alongside great creativity and talent, that is pretty much the whole game.