Failure case: −20% for the quarter. What did I do wrong in the segmentation hypothesis?

Not all of my cases are successes. This one is a failure 20% to the quarter plan. And that is why it is still useful to me. I’ll tell you what I did wrong in the segmentation hypothesis and what lesson I learned. Without emotions, by numbers. For context - I have and separate 6-step checklist on how to diagnose a failed campaign, but in this case the error was in the strategic hypothesis itself, not in the execution.

It infuriates me when marketers only show success in their portfolios. “I grew from +25%, +18%, +30%.” From such graphs, I understand one of two things: either the person did not try anything significant and received a guaranteed result, or he is holding back half of the experience. Both are weak positions. That’s why I’m dealing with my failure openly.

Segmented by demographics - should have segmented by behavior. This cost 20% of the quarter's revenue and taught me that demographics in 2025 is the laziest ICP categorization.

Context: project and task

E-com project in the category “premium accessories for smart-home”. The average bill is 18,000 ₽, the margin is 35%, the standard buyer at the beginning of 2025 is a man 28–45 years old, above average income, urban resident.

Performance budget for the quarter - 3.6M ₽. KPI - revenue 24M ₽ (i.e. ROAS 6.7×). At this point, the project had a one-year history: ROAS fluctuated stably around 5.5–7×, KPI was historically met.

In January 2025, I took on the quarterly plan. I decided that current campaigns were working “okay, but not optimally”, and proposed a hypothesis: deepen segmentation audience.

A hypothesis that seemed to work

The current segmentation at the time of my arrival was split into 3 broad groups:

• Men 28–35 (young families, new buildings)
• Men 35–45 (ready repair, upgrade)
• Women 30–45 (impulse purchase for gift)

I suggested breaking them down into 6 groups along finer demographic axes:

• M 28–32 “new apartment” - starting a smart-home from scratch
• M 33–38 “family, children” - priority on safety (cameras, sensors)
• M 39–45 “premium upgrade” - replacement of existing devices with a top-level one
• F 28–35 “gift for husband/partner” - emotional driver
• F 36–45 “gift to son/brother” - for a teenager
• Uni 25–35 “technical enthusiasts” - no gender, focus on technical characteristics

Logic: deeper segmentation = more accurate creative = higher CTR = lower CPL. For each segment there is a separate set of creatives with a unique offer and a separate landing page.

Budget plan: 600K ₽ for each segment, 100K ₽ for creative production per segment, the rest for traffic.

What the data showed after 4 weeks

A month after launch, revenue is 22% below plan. Specifically:

This is not a "standard swing". This is a sure failure in all three metrics at once. And the failure correlated with the transition to new segmentation - the old system had worked stably for 6+ months before that.

What did I do to understand the reason

The first reaction is typical: “The creatives are burning out, we’ll update them.” I rolled back part of it to the old version for control. A week later, the old version in the control pulled up to historical indicators, the new one continued to sag. So, the problem is not in creativity.

Further - deeper. I picked up CRM, unloaded all the project clients for the last 12 months and started looking at shopping patterns, and not on demographics.

What was discovered:

• Real large buyers (check 30K+) are not young men 28–32 in new buildings, as I assumed.
• Large average checks are concentrated at "serial smart-home builders" — people who buy 2–4 times a year, expanding the system.
• This group is not localized by gender/age - there are M 30, M 50, and F 35.
• What unites them behavior: have already purchased in the past, open email newsletters, access the site from a direct source, not from advertising.
• My new segmentation missed this group - it was “blurred” between 4 different demographic baskets.

Root error

Main mistake: I applied demographic segmentation to a product that requires behavioral.

Demographic segmentation (age/gender/geolocation) works for:

• Low average receipt (where there is an impulse purchase)
• Narrow ICP with explicit age restrictions (children's products, pension products)
• Brand campaigns for awareness, not sales

For a premium smart-home with repeat purchases and an average bill of 18K ₽ - you need behavioral:

• Newbies (never purchased) vs repeat customers
• Active on the site (3+ visits per quarter) vs one-time
• Cart abandoned vs cart not collected
• Email-engagement (open mailing lists) vs advertising only

My hypothesis “6 demo segments = higher CTR” broke down in reality: “serial builders” don’t buy based on demo targeting—they buy based on behavioral triggers.

What did I do next quarter

After 4 weeks, I closed the new segmentation. Admitted the mistake to the client and presented a fix plan.

Q2-2025 was redone into 4 behavioral segments:

• “Serial builders” - repeat buyers, look-alike in their behavior (40% of the budget)
• “Active researchers” - visited 3+ times, did not buy (30% of the budget)
• “Beginners with potential” - a cold audience, divided by interests, not by demo (25% of the budget)
• “Email-active without purchase” - reactivation through remarketing (5% of budget)

Q2 Results:

Q2 not only closed the gap in Q1, but also surpassed KPI by 12%. The client stayed - the lesson was accepted, trust was restored through the result.

What lesson remains - for all projects after

The main lesson is prosaic: Before changing your strategy, figure out which axis the buyer is actually segmented on. It's not always demographics. It could be:

• Behavior (repeated/new/exploratory)
• Purchase context (jobs-to-be-done - “why did the person buy it”)
• Awareness level (cold/warm/hot)
• Financial profile (check size, frequency)
• Source of income (advertising/organic/referral)

Demographics is the laziest axis. It’s convenient to do because it’s available in any advertising account. But that doesn't mean she works for your product.

What do I do now before any new segmentation?

1. Custdev on CRM data. I take the top 50 clients by revenue and look for patterns in their journey. It takes 1 business day and gives 80% of insights.
2. The hypothesis “on which axis to segment.” Not “how to break down demographically,” but “what differentiates buyers from each other in reality.”
3. Test for 30% of the budget, not 100%. Before translating the entire performance, I test the hypothesis on a quarter of the budget for 4 weeks. If it works, I scale it up. If not, I lose 4 weeks on one quarter, and not a quarter on everything.
4. Control group. I never remove the old mesh 100%. I leave 20–30% on the old segmentation as baseline. Without control, it is impossible to understand whether the new is really better.

What to pick up on Monday

If you launch a new segmentation tomorrow:

1. Don't start with demographics. First, behavioral data from CRM.
2. Custdev top 50 clients = 1 day of work = 80% correct hypotheses.
3. The hypothesis is tested on 30% of the budget, not 100%. Never.
4. A control group on the old segmentation is mandatory.
5. 4 weeks trial period. Less - it’s too early to judge, more - you’re wasting time.
6. If it doesn’t work, admit the mistake to the client in 4 weeks, not in a quarter. It's difficult, but reputation demands it.

If you now have the task of “reconsidering segmentation” and don’t want to step on the same rake, write @dipustovalov. Custdev and a hypothesis usually take a week, then a test plan and control.

Related materials: ROAS calculator, unit economics calculator, ROAS benchmarks for e-com.