Data without a story is just noise. Drama school taught him that first.

I moved to New York and realised it was expensive. While I was directing theatre I answered a one-line newspaper ad and ended up at D.E. Shaw, the hedge fund. They owned Juno Online, an AOL competitor, and were hiring people straight out of university based on grades and the fact that you had a pulse. I had a pulse. I ran the customer service department, got interested in customer service reporting and the reports we were producing for marketing, moved into product management for internal tools and infrastructure, did a series of data-tagging implementations and analysis, and worked into data analytics, which turned into data science and Pickaxe.

The link from drama is real. Data is useless without a story to explain what it means. The discipline I took from theatre is keeping the audience and giving them a so what and a now what. The VP of marketing is busy. You've told them what about the data. So what? What's important about what you're saying? And then what should they do next?

Craig Teitelbaum was a modeller at Juno. He had done work for a roofing company predicting, from the volume of flyers they sent out, how many people would call asking for an estimate. For Juno, we were sending out millions of CDs (the coasters, for anyone who remembers). Craig could say: for this volume of coasters sent on these dates, we need to buy this much call-centre time in advance. If there's a Super Bowl ad or a Monday Night Football ad, we need this many hours. He was incredibly accurate. The ability to take the maths and tell a business what it needs to prepare for, with a range of uncertainty around the prediction, is what I fell in love with. Same problem, in many ways, still applies to streaming today.

Marketers want to know how much they need to spend to hit subscriber, viewer, or ad-revenue goals, and how to allocate that spend. The recurring streaming question is what to watch next. Every streamer we work with treats finishing a season as a point of retention risk. The user has just hit the end. They're at risk of leaving. I came in, I saw my thing, oh, here's this other thing, maybe I won't cancel yet, maybe next week. So the question is what to surface, and how.

It pops up a question: are you interested in this show? The user swipes or taps yes, and the show goes onto a list to be shown back to them. If they swipe no, the algorithm learns what they don't want. The user is actively building their own recommended-for-you algorithm. Product is focused on usability and the recommendation engine; marketing is using an old-fashioned technique of we don't care how we do it, let's just get them to tell us what they want. The response has been excellent.

Across streaming services we work with, the falloff points are: do you click on episode one of season one? Can you finish it? Can you make it to episode two? If you finish episode two, you're probably going to watch the rest of the season. That's how most of this content works, with some variation by genre; mystery requires the reward of the resolution, comedy doesn't quite the same way.

Phase zero is going live. Tools up, pixels in, everything tracking. Nobody really cares about performance yet. Day two or three the question becomes is it working? Then it becomes which parts are working?, where you start slicing the data and pulling the higher-level levers (which network, which message, which landing-page variant) while the networks pull the lower-level levers with their own AI. The next phase, six to nine months in, is reflection. Marketing-mix modelling and econometrics come in. Walk, run, fly. The fly part is where you put it all together: how do these levers interact, where are we cannibalising, where are we spending too much on the wrong thing, what happens if you give me X million dollars for the next year. That's where the sophisticated picture lives.

You went from I waste half my budget but I don't know which half to the brief promise of the cookie, where for a brief period you could know everything about every user. The walled garden then started closing. After iOS 14, the death of cookies in Chrome (then the rebirth, then the death, then not yet), roughly half your users are now unidentifiable. So the methods marketers use have to assume unreliable attribution. The difference between this state and the time before is that you do have a small known pool, like a Nielsen or BARB panel, that lets you extrapolate to the anonymous majority through modelling. That known pool is a major building block going forward.

Mix is the core product. It addresses where to spend money to optimise the business, lets marketers divide their budget, pick the target they're trying to improve, query against their data, and get a calculator that will tell them: if you spend a million dollars next week, this is how many customers, viewers, or installs you'll get, with the confidence level across each of your channels. Marketers can define the channels at whatever level matters: awareness inside Meta, acquisition inside Meta, out of home, cinema, and so on.

The other key feature is business variables. Marketers can ask what if I change my price? What if I run a 50% offer? What if I drop the paywall to zero free articles this month for a regional disaster? (We do work for newspaper groups during natural disasters where the local paper will go free, sponsored, and we can model what that does to paying-subscriber acquisition.) Modelling those business features alongside the spend model is the genuinely useful part.

The most important thing we've learned is to allow markets to be themselves and not assume the same model applies to all. Always break out by country. Even Rest of World or Rest of Europe can be dangerously misleading. Benelux is not a country. Spain and Portugal are different countries. When you tag everything correctly and look at the model, product A often has a totally different reception in Portugal versus Spain. Brand loyalties are regional. Beverage companies in particular have local brands that are strong in one country and don't even exist in another.

US versus Mexico is huge in North America, less so in Europe. UK matches against Germany are intense in the UK. There was a day during 2022 when three matches went into overtime; in the US those were happening during the work day, so all the viewing data was on mobile phones with almost nothing on CTV. In the UK they were after the workday, so a great deal was on CTV at home and out-of-home in bars and pubs. Same matches, totally different behaviour patterns. The discipline is reading those patterns into the model rather than averaging them out.

The biggest impact has been people finally understanding the importance of their data and data quality. With ChatGPT and the hallucinations, suddenly everyone understands the principle: there's a pipeline, it needs to be trained, the source material has to be correct and up to date. It feels obvious now. It wasn't to most marketing teams six years ago. The second change is a sense of horizon: people understand that a great deal is now possible. Generative AI isn't necessarily great at data analysis, modelling, or maths, but it can do those things if you give it a calculator and a model to look at. The expectation has been raised, which raises the budget for the work that genuinely uses it.

The year before we launched Pickaxe Plus, we sent out an April Fool's press release announcing an S-VOD about S-VODs: an over-the-top streaming service whose content was about content, where the recommended-for-you list would just recommend other S-VODs you might want to consider, and our data would just be data about data. I get cold emails to this day asking if we want developers for Pickaxe Plus or marketing services for Pickaxe Plus. Every time I have to say you do know that was a joke? It's clearly a joke.

We don't require a college degree. We have PhDs and we have college dropouts. There is no single path to competency or excellence. Developers in particular have a deep culture of learning and documenting work for the next person; that ethos has been at the core of what we've built. Develop for the person next to you. Outside development the same is true: analysis, marketing, every field is full of learning, and only some of that learning is institutional. There's nothing that replaces on-the-job training.

Love of learning and the ability to teach yourself when there isn't a formal route. Teamwork. Even developers, who can look like lone actors, are always interacting with someone else's API, someone else's connector, someone else's data downstream. Marketers are never working alone; they're always working with stakeholders, setting expectations with management about budget and revenue.

Openly, no jerks. We're virtual. The test isn't would you have a beer with this person? It's would you want to work with them? Do you enjoy being in a meeting with them, on Slack with them, can you have fun with them? There's always a tomorrow. There's a next day. You're stuck together. So can you learn, can you work with others, are you fun to be around. Those are the three things.

Every company is a data company. Whether you sell greeting cards in a store at Boots or CVS, or you're a streamer trying to figure out exactly how many seconds of Call the Midwife is getting watched, you're trying to understand what brought a customer in, what they did, how much they paid. The difference between industries is the frame and the scale. A streamer plans out budget for weeks or months and tracks subscription starts and unique viewers. A greeting-card business doing Mother's Day cards is the same modelling problem, just compartmentalised into a different time window: 20 to 30 days of marketing lead-in, a 5-day purchase window where many sons buy on the way to brunch. The model is the same. The human layer of what is happening in this business in this industry is what separates a generic model from a useful one.

For sporting events, the lead-up matters: how much is it being talked about on social, what day of the week is the match. A World Cup match on a Thursday plays differently from one on a Monday night. Saturday night events are massive. The same applies to Valentine's Day on a weekend versus a Monday. Year-on-year comparison requires human editing of the calendar: you compare Super Bowl weekend with Super Bowl weekend, not the same calendar date. Sporting events are a little like a liturgical calendar; they adjust every year. In twenty years we'll be working with people who weren't born yet, modelling Rosh Hashanah sales for them with the same calendar discipline.

Data quality. Some consolidation around ad networks. Identity, still. A bunch of sci-fi things, like how accurate are the subdermal chips, is the sun affecting them, are the algorithms suggesting things we would want or just showing us a limited view. Are we spending too much money; we'll still be worried about that in twenty years. The pure cold business view: are we getting the right customers, do we have enough of them, are they sticking around long enough. That part seems pretty eternal.