---

From Marzipan Bars to Modern Product Strategy:
What Conjoint Analysis Still Teaches Us

---

Back in my university days, I spent weeks staring at survey data about chocolate bars. Not exactly what you’d expect from someone who would later live and breathe Microsoft Cloud and AI, but that project shaped how I think about product development to this day.

Our mission: use conjoint analysis to understand why marzipan bars live in the shadow of other flavors and how to design a bar that people would actually pick. Today, in a world of digital products, AI features and app marketplaces, the lessons from that study are still surprisingly relevant.

---

From marzipan bars to modern product strategy

---

In our seminar, we worked on the very glamorous category of “Riegelware” – those snack bars you see at the checkout, also known as “quengelware” in Germany because kids spot them, beg for them, and parents eventually give in.

The brief sounded simple:
Why do marzipan bars underperform compared to other flavors, and how could we change that?

To answer that, we didn’t just ask, “Do you like marzipan, yes or no?” We treated it like a full new product development process:

• Identify customer needs and segments
• Understand competitors and the overcrowded snack shelf
• Build different product concepts (flavor, coating, size, price, add-ons, packaging)
• Test those concepts using conjoint analysis before ever launching a new bar

If you replace “chocolate bar” with “SaaS feature” or “AI add-on”, you get the same basic pattern that modern product teams and startups still follow today: generate ideas, refine concepts, test before launch, and reduce the chance of a flop.

Back then, the numbers were brutal: in some FMCG categories, over 70% of new products failed. Concept tests and conjoint analysis were one way to reduce the odds of burning budget on things nobody really wanted. Today, with app stores and cloud services overflowing, the failure rate hasn’t magically disappeared—it just moved into the digital space.

---

What conjoint analysis really does (in human language)

---

Conjoint analysis sounds like something you’d only do with three cups of coffee and a statistics textbook next to you. In reality, it answers a very human question:

“When people choose between products, what really matters—and how much?”

Instead of asking “Do you like marzipan?”, we showed respondents different combinations of product attributes:

• Price (for example: €0.49, €0.59, €0.69, €0.79)
• Chocolate coating (milk, dark, no coating)
• Flavor (plain chocolate, cream filling, caramel, nougat, marzipan, coconut, coffee, …)
• Portioning (one bar, two pieces, multiple pieces)
• Weight (small snack vs. bigger bar)
• Add-ons (no add-on, biscuit, wafer, nuts, cereals, fruit)

Participants chose which bar they would buy from these combinations. From those repeated choices, we used hierarchical Bayes estimation to compute part-worth utilities—essentially, how much each feature and value increases or decreases the likelihood of a choice.

Today, product teams do something very similar, just with prettier dashboards:

• A/B tests in apps instead of paper questionnaires
• Feature flags instead of hypothetical flavors
• Data science pipelines instead of SPSS + seminar room

But the logic is the same: people make trade-offs, and instead of guessing, you measure those trade-offs.

---

What people really wanted and why marzipan struggles

---

When we crunched the numbers, some patterns were wonderfully intuitive—and some were a bit painful for marzipan fans.

First, the obvious one: price matters.
As expected, lower prices generated higher utilities. Between €0.59 and €0.69, many respondents were nearly indifferent; below that, the price became a real positive driver. That’s still true today: even in premium niches, price elasticity is very real, especially in crowded categories like snacks or app subscriptions.

Flavor-wise, the data was crystal clear:

• Classic chocolate flavor was the top favorite
• Cream fillings performed very well, especially among women
• Caramel and nougat were also strong
• Marzipan, coconut, and coffee flavors scored significantly lower on average

Marzipan wasn’t a total disaster—but it clearly played in the second league. That already hinted why marzipan bars sit in the corner while chocolate, caramel and nougat dominate the center shelf.

For other attributes we saw similar patterns:

• Milk chocolate coating beat dark chocolate and “no coating” comfortably
• A bar split into two pieces felt just right—easy to share or save, but not over-fragmented
• Bigger weights increased perceived value up to a point; beyond about 65–80g, the utility started to level off or drop
• Add-ons like plain or biscuit were preferred over cereals and especially over fruit pieces, which scored poorly—if people reach for a chocolate bar, they apparently don’t want disguised health food

Interestingly, when we split the data by gender, we didn’t get a completely different world—but we did see nuanced differences:

• Women were more price-sensitive overall
• Men cared more about size (very small bars scored worse with them)
• Women rated cream fillings and cereal components higher
• Men leaned more towards nuts and kept a stronger preference for milk chocolate

These are exactly the kind of insights that still drive segmentation and positioning today: the product that works best for one segment might not be the winner for another.

---

Why concept tests still matter in a world of AI and cloud

---

You might ask: “Nice snack bar story, Uwe, but what does this have to do with my cloud transformation or AI roadmap?”

A lot.

The mechanics haven’t changed:

• In consumer goods, we mix price, flavor, packaging, size
• In software, we mix features, UX flows, pricing models, support levels
• In cloud and AI, we mix service tiers, data residency, AI capabilities, compliance guarantees

The risks haven’t changed either. Whether it’s a marzipan bar that nobody buys or a cloud product nobody activates, launching the wrong thing at scale is expensive. That’s why concept tests, conjoint analysis, and structured experiments are still gold—especially when you move fast with cloud-native services and AI features.

The difference today: we don’t have to wait weeks for survey data and manual calculations. With Microsoft’s ecosystem and modern analytics, we can:

• Run near real-time experiments across regions and segments
• Feed telemetry into product decision loops
• Use AI to detect patterns in usage and preference
• Combine classic survey-based research with behavioral data from real users

In other words: the snack-bar methodology grew up, moved to Azure, and learned to work with streaming data and AI. But the question it answers is still the same:

“What combination of attributes gives this product the best chance of success—for whom?”

---

The limits of conjoint analysis (and why humility is part of good product work)

---

Even in our student project, we ran into the limitations that every serious study faces—and those are still very relevant for today’s product teams.

First, respondent fatigue. Some participants told us quite openly that the survey felt too long and that they stopped paying attention to details like price toward the end. That’s the reality in many research setups: people get tired, take shortcuts, and the data becomes noisier.

Second, the “number of levels” effect. Attributes with more levels often look more “important” in the analysis, simply because there are more utility steps between the best and worst option. That can distort perceived importance and tempt decision-makers to tweak the wrong lever first.

Third, the assumption of compensatory decision-making. Conjoint models often assume that people mentally add and subtract utilities (“this flavor is worse, but the price is better, so overall I still pick it”). Real humans often don’t behave like that. They use heuristics:

• “Never pay more than €x for a bar”
• “Always pick milk chocolate”
• “No fruit pieces in my chocolate, ever”

Later research showed that only a portion of respondents truly follow additive, compensatory rules. Others use threshold-based or simplified decision strategies. That doesn’t make conjoint useless—but it means you should treat it as one strong lens, not the single source of truth.

Translate that to today’s world and you get a clear message:

No matter how shiny your models, telemetry dashboards, or AI assistants are—reality is always messier than the model.

That’s why the best product teams combine:

• Quantitative modeling (conjoint, telemetry, funnel data)
• Qualitative insights (interviews, usability tests, field research)
• Continuous validation after launch (usage metrics, churn, feedback)

---

From marzipan bars to cloud products: what I still use from this study

---

Looking back at this seminar from today’s perspective, when I work with enterprises on Microsoft Cloud, Azure, AI and modern application architectures, a few core principles have stayed with me:

• Don’t fall in love with ideas, fall in love with evidence. Marzipan might be your personal favorite, but if the preference structure says “chocolate + cream + fair price”, that’s the direction your mainstream product should explore—unless you consciously target a niche.
• Design for segments, not for averages. The “average” respondent in our study didn’t really exist. Men and women had different priorities; in real markets, age, income, context, and use cases add even more layers. In cloud and SaaS, that’s your enterprise vs. SMB, regulated vs. non-regulated, core vs. edge workloads.
• Prototype on paper before you prototype in code (or in factories).
  A well-designed concept test can kill bad ideas before they cost real money. Today, that might be a combination of user story mapping, Figma prototypes, simulated pricing pages, or low-fidelity feature toggles. Same mindset, different tools.
• Accept that no method is perfect—but disciplined imperfection beats guessing.
  Yes, conjoint has theoretical weaknesses. Yes, survey fatigue is real. But a structured, data-informed view of preferences is still dramatically better than the HIPPO (Highest Paid Person’s Opinion) method.

In a weird way, that marzipan project was my first serious lesson in data-informed product leadership—long before cloud economics, FinOps, or AI-driven analytics came into my daily work.

Stay clever. Stay customer-obsessed. Stay insight-driven.
Your Mr. Microsoft,
Uwe Zabel

---

🚀 Curious how market research, data-driven product design, and modern cloud strategy come together? Follow my journey on Mr. Microsoft’s thoughts—where cloud, AI, and business strategy converge.
Or ping me directly—because building the future works better as a team.