In this episode of TECS Lens – Emerging Founders, Antonis sits down with George Filippou, founder of MarSci, to explore how marketing analytics and AI-driven measurement are reshaping modern decision-making.
Link to the Youtube channel: https://www.youtube.com/@TECSLens
The number that started everything
A few years ago I was responsible for budget allocation at a large advertiser. Google Ads on one side, Facebook on the other, a healthy budget, and a team that knew what it was doing.
Then I looked at where the revenue was coming from. Almost 80 percent of it was attributed to direct traffic.
If you have an analytics account open right now, go and check your own. There is a good chance you will see something similar.
Direct traffic is supposed to mean someone typed your address into a browser and arrived with no help from anything else. Eighty percent of revenue arriving that way is not a finding. It is a gap. We were running sophisticated campaigns across multiple channels and the measurement system was quietly telling us that none of it mattered.
That was the moment I stopped trusting the dashboard. It is also, more or less, the reason I started a PhD. My first publication was an attempt to decode that direct traffic and give the credit back to the channels that actually earned it.
Consumers stopped buying in straight lines
The industry still carries an assumption from a simpler era: someone sees an ad, clicks the ad, lands on the site, adds to cart, pays. One session, one journey, one channel to thank.
Think about how you bought anything recently. A friend sends you a link on WhatsApp. You look at it on your phone. You search the brand later on your laptop. You mention it to your partner over dinner. Your partner buys it for you.
That is a normal journey for a pair of shoes. It is not even a considered purchase like a car or a house. And almost none of it survives contact with a last-click report.
I would not say last-click attribution is always wrong. There are businesses where it is a reasonable approximation. The myth is not the method. The myth is the belief that people buy in a linear way, and that belief is what keeps budgets pointed in the wrong direction.
Why we built causal-driven attribution
The idea behind MarSci had been in my head for years, from working with advertisers across Europe at Facebook and TikTok and seeing the same problem in every meeting. Good teams, clear strategies, and no reliable way to know what was working.
The concrete starting point came about two years ago with causal-driven attribution, or CDA, which is now core technology for us. CDA uses aggregated data to produce the kind of insight you would normally need user-level data to get. No clean rooms. No months of pipeline work before you see anything. It is cheaper to compute, easier to feed, and easier to read at the end.
When the model started working, it was clear this was not a feature. It was the foundation of a company.
The hard part was proving it
Digital marketing has no ground truth. Ten channels will each claim credit for the same sale, and there is no way to open the box and check who was right. That makes validating any new methodology genuinely difficult, and it makes it doubly difficult when your method deliberately avoids user-level data, which is exactly what makes it attractive to advertisers in the first place.
We had a technology we believed in and no clean way to prove it. There is no clever answer to that problem. You just have to do the work.
So we stopped production and built the missing piece: a full methodology for generating synthetic datasets with known ground truth, which we then open sourced so other people can test our claims, and their own.
That was the hardest stretch we have had. It was also the most useful thing we did.
From dashboards to decisions
Legacy analytics are deterministic. They show you what happened, and then they stop. Everything after that point, interpretation, prioritisation, the internal argument about what to do next, is left to humans.
The shift now underway is toward probabilistic systems that go further. Not just insight, but a recommendation. Not just a recommendation, but a way to validate the change before anyone commits budget to it.
This is also where we part company with a long-running industry habit. For years the plan was to educate marketers into becoming analysts. I think that dream is behind us. A marketing manager cannot be a data scientist, and honestly should not have to be. The goal is a future where someone runs a marketing mix model without ever needing to know that is what they just did.
Accuracy is not the point. Direction is.
Much of the industry competes on model robustness, and robustness matters. But probabilistic modelling is fundamentally about direction. You are working with data that has already been pre-processed by media platforms, without clean first-party signal, in a world with no ground truth.
So we validate on the action rather than on the insight. Did the recommended change produce the outcome we expected? That question is answerable, and it is the one clients actually care about.
Three tests decide what we build:
- Does it move top line or bottom line for an advertiser?
- Is it genuinely different from what already exists?
- Do our clients want it? We ask them directly, usually on WhatsApp.
If the answer to all three is yes, we build it. If not, it waits.
The moment it clicked
The best feedback we have ever received arrived as a WhatsApp message from a client. A screenshot, some celebration emojis, and four words:
What on earth is that?
They could not have explained the method behind the result and they did not need to. They could see what it was worth to their business. We stopped for a moment and celebrated that one, because it meant the work made sense to someone who has never heard of causal inference and never needs to.
What we are actually trying to do
We are not building another SaaS tool for a narrow list of accounts. We are trying to democratise marketing measurement, and that only works at scale and only works in the open.
More than 160 advertisers are on the platform today, running close to 200 models a week between them. The methodology is peer reviewed, the tooling is open source, and the ambition is considerably larger than the current numbers.
It ain’t going to be easy. That is one of our values, and we picked it deliberately.
If you are running four or five active channels and you have someone accountable for the budget, we would like to hear from you. Come and see what your media mix actually looks like.