There is a habit in data analysis that looks like rigor but is actually its opposite.

You look at a dataset. You form an intuition about what's probably true. You build a query that confirms it. You write it up. The analysis "worked."

The problem is that you never tested anything. You found what you were looking for because you only looked where you expected to find it. The uncomfortable alternative — that your intuition was wrong — never got a real chance to surface.

This is confirmation bias dressed up as analysis. And it produces findings that feel true, read well, and guide decisions in the wrong direction.

The alternative: build to be proven wrong

A few months ago, I ran a retention analysis on an online retail platform. Before writing a single query, I wrote seven specific, falsifiable predictions — each with a defined confirmation condition and a defined rejection condition.

I was genuinely convinced most of them were right.

Five were rejected.

Here's what those five rejections looked like:

And the two I got right:

I'll be honest — when H1 came back rejected, I sat with the result for a few minutes hoping I'd written the query wrong. I had been so certain the business had a retention problem. Finding out it didn't felt like losing the point of the whole exercise. It took a moment to realize: this is better. A business that thinks it has a retention problem and doesn't is about to waste serious money. Finding that out early is the entire value of the analysis.

What the rejections actually taught me

H1 told me the 72.4% repeat rate was driven by wholesale buyers and resellers — not casual consumers. The business didn't have a retention problem. It had a wrong-fit acquisition problem. That's a completely different solution.

H6 was the most counterintuitive. Customers with the most returns retained at 96.1%. Once I stopped expecting the opposite, the reason became clear: these were resellers returning unsold inventory as part of their normal cycle. Returns were a loyalty signal. Treating them as a churn risk would have been a costly mistake.

H5 killed a popular playbook. "Get them to buy twice and they're yours forever" is conventional retention wisdom. This dataset said no — loyalty here built gradually across every purchase, not sharply at purchase two. The recommendation changed entirely because of that rejection.

In each case, the rejection didn't just say "you were wrong." It said here is something about how this business actually works that your intuition missed. That's more valuable than three confirmations.

Why most analysts skip this

Writing hypotheses before you analyze feels slower. It introduces friction. And it creates a record of being wrong.

That last part is the real reason. In professional settings, being wrong feels like weakness. The safer framing is "I explored the data and found these interesting patterns" — no predictions, no accountability, no record of what you expected.

But that safety is the problem. Exploration without prediction is documentation. You're describing what happened, not testing whether your model of the world is accurate.

The format that makes it work

The hypothesis format I use before opening any query editor:

Hypothesis: Specific, falsifiable prediction Confirmation condition: What the data needs to show Rejection condition: What the data needs to show instead Business implication if confirmed: What follows Business implication if rejected: What follows instead

That last line is the one most people skip. If you can't say what you'd recommend when wrong, you haven't thought through the problem deeply enough.

Why this changes the output

The deliverable changes when hypotheses are defined upfront.

Instead of "here are the interesting patterns I found," the output becomes "here are the questions I asked, here is what the data said, and here is what that means for the business."

That structure is more defensible. It separates what the analyst expected from what the data showed — the most important distinction in any analytical document.

In this case, the five rejections shifted the entire recommendation. Not toward retention spending — toward acquisition strategy. The business didn't need to keep customers better. It needed to attract more of the right ones from the start.

Five of my seven hypotheses were wrong. The analysis became stronger because of it. And the resulting recommendations shifted acquisition strategy, not retention spending — which is what happens when you let the data answer honestly instead of confirming what you already believed.

Full methodology, hypothesis framework, and code → GitHub

The retention analysis referenced here was conducted on the Online Retail II dataset (UCI Machine Learning Repository) using Python, pandas, and matplotlib.

I tested seven hypotheses against the Online Retail II dataset (UCI ML Repository) using Python and pandas—cohort retention curves, first-session behavioral splits, and CLV segmentation.

When I began this analysis, I was searching for a retention problem.

That was the expectation.

Two years of transactional retail data. Over a million rows. Roughly 6,000 customers. Tens of thousands of invoices.

Classic e-commerce terrain.

And from experience, one expects the usual story:

Acquire customers. Convert them once. Lose most of them shortly after. Repeat the cycle.

That is the way of the world.

So naturally, I assumed the repeat purchase rate would reveal weakness.

Instead, the data contradicted me almost immediately.

What the data revealed

72.4% of customers made more than one purchase.

Now pause there.

Most e-commerce businesses celebrate when repeat purchase rates cross 30% for consumer retail.

Industry averages often sit somewhere between 20–40%.

This business was sitting at 72.4%.

Not slightly above average. Not “performing well.”

Abnormally high.

And strangely enough, the framing around the business did not reflect how unusual that number truly was.

The analysis I built to discover a retention problem had accidentally proven the opposite:

The business was already retaining people exceptionally well.

At least... a certain kind of people.

The first assumption failed again

My immediate conclusion was simple:

“Good. Retention works. Scale acquisition.”

Bring in more users. Let the retention engine handle the rest.

Reasonable logic.

But before accepting that conclusion, I needed to understand who exactly these returning customers were.

Because metrics speak.

And a number as high as 72.4% is rarely random.

So I kept digging.

53,628 invoices across 5,942 customers.

That averages roughly 9 invoices per customer over two years.

For a retail platform selling gifts, décor, and novelty items?

That is unusually high.

Then I examined returns and cancellations.

And this is where the analysis stopped behaving normally.

Customers with 4 or more returns retained at 96.1%.

Non-returners retained at 58.3%.

Read that again.

The people returning the most products were also the people staying the longest.

Completely backwards.

Or so it seemed.

And then the picture clicked.

Not mathematically.

Humanly.

What was actually happening

This was never truly a casual consumer customer base.

Not primarily.

The behavioral patterns gave them away.

Frequent purchases. Frequent returns. High invoice volume. Sustained engagement over time.

These were not ordinary shoppers buying birthday gifts and scented candles on impulse.

These were resellers.

Wholesale-oriented buyers.

Small operators running miniature businesses through the platform itself.

And suddenly everything made sense.

The returns were not signs of dissatisfaction.

They were inventory behavior.

Unsold items returned. Fast-moving products reordered. Demand tested in cycles.

The “customer” was behaving less like a shopper and more like a merchant.

That realization changed the entire meaning of the dataset.

The 72.4% repeat purchase rate was no longer merely a retention metric.

It was evidence of a business model hiding inside another business model.

The question changed

Initially, the question was:

“How do we improve retention?”

But the data forced a different question entirely:

“How do we acquire more customers who already behave like the ones that stay?”

That is a radically different problem.

Improving retention assumes the current customer base is fundamentally correct and merely needs encouragement to remain longer.

But the data suggested something else:

The loyal customers were already loyal.

The issue was not retention.

The issue was customer fit.

The 27.6% who disappeared after one purchase were not failed loyalists.

They were different people entirely.

Casual buyers. Campaign-driven traffic. One-time purchasers with no structural reason to return.

And no amount of discount emails or loyalty points can manufacture long-term intent where none existed in the first place.

The signal appeared early

This was perhaps the most interesting part of the entire analysis.

I separated customers into two groups:

One-time buyers

Repeat buyers

Then I examined only their first-session behavior.

Before retention mechanics. Before remarketing. Before loyalty campaigns.

The differences were already visible immediately.

Repeat buyers:

Spent 32% more in their first session

Purchased 26% more items

Explored 19% more product categories

The people who stayed behaved differently from day one.

Not subtly.

Structurally.

Because they were not shopping recreationally.

They were sourcing inventory.

And this means something important for businesses:

Loyalty often announces itself early.

The activation threshold exists in the data.

A customer who explores broadly, spends aggressively, and purchases with volume in session one is not merely “engaged.”

They are statistically more likely to become long-term revenue.

Which means businesses can act immediately.

Wholesale onboarding. Dedicated account management. Personalized retention sequences. Priority servicing.

The signs appear long before the second purchase occurs.

What this means beyond this dataset

This lesson extends far beyond one UK retail platform examined by a man sitting in Lagos with a notebook, SQL queries, and too many tabs open at midnight.

The deeper lesson is this:

Assumptions distort analysis.

We often decide what the problem is before the data speaks.

And once the mind accepts a narrative, it begins searching merely for confirmation.

The original assumption here was textbook e-commerce reasoning:

“Retention is weak.”

But reality proved stranger.

The business possessed an unusually loyal wholesale core hidden beneath aggregate averages.

Two customer worlds existed simultaneously:

A loyal reseller ecosystem

A casual consumer periphery with natural churn

Combined together, they produced one healthy-looking average.

But averages are diplomatic.

Segmentation tells the truth.

The uncomfortable recommendation

And now we arrive at the part most growth teams dislike hearing.

Stop trying to retain people who were never going to stay.

Not every lost customer is a failure.

Some customers are simply wrong-fit acquisitions.

The one-time buyer who arrived through a campaign, purchased impulsively, and vanished afterward is not necessarily a retention problem.

Sometimes he is merely evidence of poor targeting.

And pouring loyalty programs, re-engagement emails, discounts, and incentives into such users often becomes an expensive attempt to negotiate against reality.

The wiser path is harsher.

Let them go.

Redirect the effort toward finding more people who resemble the loyal core.

More resellers. More wholesale-minded buyers. More structurally aligned customers.

Because growth is not merely about increasing numbers.

It is about increasing the right numbers.

And the data will usually tell you who those people are...

...if you ask it honestly enough.

This analysis was conducted using the Online Retail II dataset (UCI Machine Learning Repository) with Python, pandas, and matplotlib. Full methodology, hypothesis testing process, and project code are available on GitHub.

When discussions on e-commerce conversion arise, men often rush toward checkout.

Cart abandonment. Payment friction. Long forms. Forced account creation before purchase. The usual suspects.

And yes, these things matter.

But upon examining a year's worth of Google Analytics data from the Google Merchandise Store — over 903,000 sessions across one full year — I found something rather unsettling:

The real problem begins long before checkout is ever reached.

In fact, long before desire is even formed.

What the funnel revealed

Upon building the acquisition funnel, the figures appeared thus:

Now pause at the second line.

Only 14% of all sessions resulted in a product view.

Which means that roughly 86% of visitors arrived at the store, wandered briefly, and departed without meaningfully engaging with a single product.

That is the wound.

Not checkout.

Not payment.

Not shipping.

The leakage occurs before interest is properly born.

And strangely enough, everything after product view appears relatively functional. Not extraordinary, but functional.

40% from product view to cart. 45% from cart to checkout. 52% from checkout to purchase.

The lower funnel survives.

The upper funnel hemorrhages.

The first assumption failed

Naturally, my first instinct was to blame traffic quality.

Wrong audience. Weak targeting. Low-intent visitors.

Simple enough.

So I segmented the data.

275 unique traffic sources. Grouped by product-view rate. Sources under 100 sessions removed to reduce noise.

And then came the strange part:

Not one high-volume traffic source qualified as “high quality.”

None crossed a 60% product-view rate.

So I moved further.

Desktop. Mobile. Tablet.

Same pattern.

Then geography.

121 countries examined. 30 passed the minimum threshold.

Again:

No standout quality segment.

At this point, the pattern itself becomes information.

When every segmentation cut reveals the same weakness, the issue no longer belongs to the segments.

It belongs to the thing itself.

What is actually happening here?

The Google Merchandise Store sells officially branded Google merchandise:

Hoodies. Water bottles. Journals. Caps. Stickers.

Recognizable products. Functional products. Even exclusive products.

And yet exclusivity alone does not create desire.

That is the crucial distinction.

A thing may be unique without being aspirational.

People arrive at the store through many routes:

Some are Google enthusiasts. Some arrive from YouTube or blog content. Some are developers merely curious. Some likely stumbled upon the store accidentally through campaigns or referrals.

But most arrive without a strong internal desire to own Google-branded apparel.

And therein lies the fracture.

The brand possesses technological authority, yes.

But authority in technology is not automatically authority in lifestyle identity.

Luxury brands understand this deeply.

Men do not purchase certain products merely because they function.

They purchase symbols. Associations. Statements. Identity projections.

The product becomes social language.

Google merchandise, for many visitors, has not yet become such language.

And because of this, visitors arrive... but do not linger.

What this means for businesses

Conventionally, low conversion rates are answered with more traffic.

Run more ads. Increase impressions. Scale acquisition.

But what happens when the majority of visitors disengage before desire even begins?

More traffic merely multiplies disengagement.

One does not solve leakage by increasing water pressure into a broken pipe.

One first examines the fracture.

Diagnosing the real issue

Now, pre-engagement dropout can emerge from several places.

If it is a landing-page problem:

Visitors are not efficiently routed toward products.

The issue then becomes navigational.

Poor product discovery. Weak calls-to-action. Unclear hierarchy. Insufficient visual pull.

This is a UX problem.

If it is a traffic-quality problem:

Certain channels are attracting low-intent users.

This becomes an acquisition problem.

Audit the channels. Reduce spend where engagement remains weak. Retarget differently.

In this dataset, platforms like m.facebook.com, t.co, and reddit.com showed high traffic volume with relatively weak engagement.

Reasonable candidates for scrutiny.

But if it is a desirability problem:

Then funnel optimization becomes cosmetic.

For no amount of checkout refinement can save a product people never desired to explore in the first place.

And this, I believe, is the deeper issue revealed here.

Because the low engagement persisted across:

• Devices

• Countries

• Traffic sources

• User segments

The consistency of the weakness eliminated many alternative explanations.

The pattern pointed repeatedly toward the same conclusion:

The problem was not merely visibility.

It was pull.

The practical lesson

Before optimizing checkout flows...

Before purchasing funnel software...

Before endlessly A/B testing button colors...

Build the full funnel first.

Observe where desire dies.

For sometimes the issue is indeed checkout.

Sometimes it is payment friction.

But other times?

86% of visitors leave before seeing a single product.

And no checkout optimization on earth will rescue users who never intended to stay.

This analysis was conducted using Google Analytics session data queried in Google BigQuery. Full SQL methodology, funnel construction, and findings are available on GitHub.