Your Brand’s Marketing is Data-Driven. Do Your Sources Agree?
“Five years ago, a customer looking for a service typed a query, and we could see it, bid on it and measure it. Today, a meaningful share of that same intent goes to an AI assistant instead. There is no query to buy and no results page to rank on. There is an answer, and you are either in it or you are not.”
Topics
Michiel Das has spent the last fifteen years watching marketing’s definition of “data‑driven” quietly change under its own feet. Early in his career, data‑driven meant something narrow and mechanical: choose a keyword, buy the click, and count whatever happened inside the platform’s attribution window. It was a world of clean charts and clear responsibilities, where the job was to optimise the part of the journey you could see and stay quiet about the part you could not.
That model worked as long as the purchase happened close to the click. It became far less convincing when he spent eight years in the automotive industry, where the customer often bought a car in a dealership months after the first search, and the most important moments of the journey lived outside any dashboard.
“The brands that did not depend on generic keywords were playing a completely different game,” he says. “If somebody searches for your name, the important part of the battle is already won, and the click is almost an administrative detail. Every euro invested in a brand lowers what you have to pay to be found later, and it keeps paying you back for years, which no performance channel does.”
As CMO of S1ervice Club, that insight has shifted his focus downstream of the click. At the same time, the data itself has stopped being what people think it is. “When somebody tells me they are data-driven, what I want to know is what they do when their sources disagree,” he says.
In his own teams, that means working with several imperfect indicators at once: what the platforms report, what the CRM says about the quality of what actually arrived, what retention looks like weeks later, and what people say in their own words. “When most of them point the same way, we move.”
In this interview, Das unpacks how he has rebuilt marketing around that philosophy: from the signals he now treats as non‑negotiable, to the experiments he refuses to run, to the guardrails he puts around AI agents that can design and launch tests autonomously.
The result is a picture of data‑driven marketing that is less about dashboards and more about judgment, and less about running more tests than about asking bigger questions.
“Access to data has never been so democratic. Any junior can ask an AI to analyse a dataset and get a competent report in thirty seconds, and it will usually be right. The challenge nowadays is understanding what those numbers mean in the context of this specific business.”
How has your definition of “data-driven marketing” evolved from keyword-centric optimisation to multi-signal, multi-touchpoint modelling?
At the beginning of my career it meant something quite narrow: choose a keyword, buy the click, and count whatever happened inside the attribution window. I spent eight years in the automotive industry, where the purchase happens in a dealership months after that click, so we optimised the part of the journey we could see and stayed fairly quiet about the part we could not.
That period taught me the lesson that has shaped my understanding of marketing since: the brands that did not depend on generic keywords were playing a completely different game.
If somebody searches for “family SUV”, you are in an auction against everyone else, paying for the right to be considered. If they search for your name, the important part of the battle is already won, and the click is almost an administrative detail. Every euro invested in a brand lowers what you have to pay to be found later, and it keeps paying you back for years, which no performance channel does.
The real power lies in combining that brand-centred focus with everything that happens after leads fill in a form: what really predicts whether the business works or not is whether the client is still ordering four months later. Once you accept that, cost per lead becomes an operational metric rather than a marketing objective, and the interesting work moves downstream of the click, into activation, retention and the quality of the match.
The numbers themselves have also stopped being what people think they are. Since consent banners arrived, a large share of your visitors are never tracked individually, so the platforms fill the gaps with models and hand you back a figure that looks precise but is partly a prediction. And a growing share of discovery now happens somewhere you cannot see at all, inside an AI assistant or a WhatsApp group.
That’s why when somebody tells me they are data-driven, what I want to know is what they do when their sources disagree. In my specific case, I tend to work with several imperfect indicators at once: what the platforms report, what the CRM says about the quality of what actually arrived, what retention looks like weeks later, and what people tell us in their own words. When most of them point the same way, we move.
This is also the part I find most interesting right now. Access to data has never been so democratic. Any junior can ask an AI to analyse a dataset and get a competent report in thirty seconds, and it will usually be right. The challenge nowadays is understanding what those numbers mean in the context of this specific business, and turning them into a decision that is not the same generic recommendation the same model is giving your competitor.
That interpretation layer is where people and companies will separate themselves over the next few years, with those that manage to go beyond the reports and generic recommendations being able to stand out.
Industry benchmarks this year put AI referrals at around 1% of website visits, which is why plenty of teams still dismiss it. That number misleads in both directions. Those visits convert several times better than organic ones, and far more importantly, most of the influence never becomes a click at all.
Can you describe the most important data signal you track today that wasn’t on your radar five years ago?
There is an old line in marketing that your brand is what people say about you when you are not in the room. For decades, that was a lovely metaphor and almost impossible to measure. It has now become literal, because a growing share of the conversations about your category happen inside an AI assistant, with nobody from your company present, and the assistant is the one making the recommendation.
Five years ago, a customer looking for a service typed a query, and we could see it, bid on it and measure it. Today, a meaningful share of that same intent goes to an AI assistant instead. There is no query to buy and no results page to rank on. There is an answer, and you are either in it or you are not.
At the same time, industry benchmarks this year put AI referrals at around 1% of website visits, which is why plenty of teams still dismiss it. That number misleads in both directions. Those visits convert several times better than organic ones, and far more importantly, most of the influence never becomes a click at all.
Somebody asks Gemini, ChatGPT or Claude for a recommendation, gets three names, then searches for one of them directly. In your analytics, that appears as direct traffic and a branded search, and the conversation that actually made the decision leaves no trace.
The solution I’ve been implementing is checking for these searches with a set of prompts written the way a real potential client would ask, in the languages of the market, looking at how we are described and whether what is said about us is even accurate. When the answer is wrong, the fix is content, PR, third party sources and SEO content.
The second signal I would not have valued five years ago is what people type into the conversational agents after visiting your website. In the case of delivery drivers looking for jobs, for instance, some don’t fill in the job form on our website, because they want to consult their AI assistant to know more about your company. They ask the agent whether they need their own van, whether the deposit comes back, whether they can work afternoons because they pick their kids up at four, whether the contract is a real contract, etc.
These are thousands of objections a month that could be partially picked up by hosting yearly focus groups with twelve people in a room who knew they were being watched.
Thanks to the concept of synthetic users (AI-generated digital personas designed to simulate the behaviours, thoughts, and responses of real human target audiences), we’re now able to interview hundreds of drivers every month, generating a continuous stream of what actually worries people. All of that information goes straight into the landing pages, the ads, the onboarding and the product roadmap.
How do you decide which hypotheses are “worth” experimentation time, and can you share a case where saying no to experiments was as important as saying yes?
The first question is what it is worth if we are right. Not whether the idea is interesting, because most of them are, but what actually moves in the business: two points of conversion on a page that thirty people visit a month is not worth three weeks of anybody’s time.
The second question is whether the result would change a decision. If we are going to carry on doing exactly the same thing whether it wins or loses, we are actually not experimenting. Those questions alone kill half the ideas that get proposed, including plenty of mine.
The tests that are worth the time are almost never about execution. Button colours, subject lines and headline variants are cheap and they give you a clean number, but they rarely drive fundamental changes. The valuable ones challenge an assumption about the customer.
The clearest no I have given was in a services marketplace, where we spent months running tests to lower the cost per lead and got very good at it. It was the wrong objective. We were not short of customers asking for the service. We were short of the right professionals to deliver it: qualified, close enough to that customer, and free on exactly the days and hours that customer needed.
So every extra lead went into a queue. Somebody who calls on Tuesday needing the service by Friday will not wait two weeks while we look for the right person, and they should not have to. They go somewhere else, and we have paid twice: once for the lead, and once in reputation, because theirs is the version of the story their friends will hear. Cheaper leads were making that worse rather than better, and the dashboard kept telling us we were improving.
We stopped the programme and put the effort into two things instead. The first was recruiting and keeping the professionals on the other side of the marketplace. Most companies treat that as an HR problem when it is really a marketing one: you have to reach them where they actually are, convince them that working with you beats the alternative, and give them enough reason to stay that they are worth what you paid to find them.
We applied the same discipline to that side of the business as we did to the customer side, and it moved considerably more revenue.
The second was brand. A customer who arrives after comparing prices on three websites picks whoever is cheapest that morning and leaves the month somebody undercuts you. A customer who arrives because they already know your name, or because somebody they trust mentioned you, accepts waiting a couple of days for the right person, argues less about price and stays for years.
Two years later, we were the most recognised brand in the category, which turned out to be the cheapest acquisition asset we ever built. None of that produces a satisfying chart at the end of the month, which is exactly why it is so easy to keep postponing.
If I had to compress it into one line: fewer tests, bigger questions. Ten well-chosen experiments a year on things that could change the strategy are worth more than two hundred on things that cannot.
Anything that only exists in somebody’s head is a bottleneck. This shift also implies that everything has to be written down as prompts, playbooks and instructions the agents can use, or the whole system stops the week that person is on holiday.
What operational changes were necessary to move from campaign-centric testing to always-on experimentation?
The campaign calendar is a leftover from a world where you bought media in blocks and people were only reachable at certain moments. That world is gone. People search, compare, ask an AI assistant, read reviews and decide whenever it suits them, at eleven at night if that is when the problem appears, and they have no idea that your campaign starts in September.
In our industry, it is even more obvious. Churn in delivery and logistics runs above 65% a year, so there is never a week when a client does not need drivers.
Three things had to change. The first is that content stopped being a project. Our website now runs on a framework with a Git-based CMS, managed from a code editor with Claude, so a page is versioned, reviewed and live in minutes, in any of our markets. That sounds like a technical footnote, but in most companies, the real bottleneck in marketing is the queue in front of whoever controls the website, and removing it changes what a team is willing to try.
The second is that we connected the tools to each other. We have automation tools that are all connected to our AI agents, so the data sources and the reporting talk to each other without anyone clicking anything. The agents pull from analytics, Search Console and the CRM, and what lands on Monday morning is a written summary of what changed and what deserves attention, rather than another dashboard nobody opens.
The third change is the one people underestimate, which is the roles. Any marketing team now needs fewer people who own a single channel, and more who can move between data, content and automation.
Somebody has to look after the automations the way a developer looks after code, because when they break, they break quietly. And somebody has to own the written version of the brand: the tone, the vocabulary, the claims we are allowed to make. That document is now an input into every agent we run, so it stopped being a PDF that just sat on a shelf without anyone actually using it.
The last piece is cultural. In a team that works this way, anything that only exists in somebody’s head is a bottleneck. This shift also implies that everything has to be written down as prompts, playbooks and instructions the agents can use, or the whole system stops the week that person is on holiday.
Most of the horror stories are badly written objectives. Tell a system to minimise cost per lead and it will find you the cheapest human beings on the internet, which most of the times means unqualified applicants.
With AI agents increasingly able to design, launch and iterate experiments autonomously, what guardrails do you put in place?
My worry is not that an agent goes rogue, but that the output is mediocre and nobody notices, because we’re not sure what the tool is doing. Building software has become extraordinarily cheap: what used to take an engineering team a year can now be assembled by three people in a month, and the agents optimising your funnel are optimising your competitors’ with the same models and much the same playbooks.
That combination is what damages a brand. An AI agent writing forty landing pages a week will produce something acceptable on every one of them, and something excellent on none. You’ll risk ending up with a website that says nothing, in a voice that belongs to no one.
So the first guardrail is visibility. Every automation needs to have an owner, a log, and a sample that a human reads every week. If that person cannot explain in one sentence what a workflow does and where it could go wrong, it shouldn’t be running.
The second is the objective you hand it. Most of the horror stories are badly written objectives. Tell a system to minimise cost per lead, and it will find you the cheapest human beings on the internet, which most of the times means unqualified applicants.
The third is how much autonomy each type of decision earns, depending on how expensive it is to be wrong. Meta descriptions, bid adjustments, internal reporting and first drafts can be run on their own. But anything that speaks in the brand’s voice, or that a customer will read as a promise, goes past a human before it goes out.
So my honest advice, after fifteen years of watching tools come and go, is to automate everything that repeats itself, and put the time you win back into the one thing that does not scale and does not copy: being a brand people actually want to belong to. That is where this whole industry is heading, and the sooner you start, the further ahead you will be when everyone else catches on.
ALSO READ: The Customer Journey Doesn’t Care Who Owns the Platform
