CommerceLens
Business

Decision Lab

Observational, not causal

This page translates five statistical findings from elsewhere on the site into business decisions, without claiming causation.

Every finding below is a real, observational pattern in this dataset. None of it was produced by a randomized experiment, so none of it should be read as proof that changing one variable causes a change in another. The governing rule for this entire page:

Forbidden phrasing
“Discounts cause purchases.”
Correct phrasing
“Purchase probability is higher among discount-exposed sessions in this observational dataset.”

Every finding below follows this same pattern: an observed, quantified difference, paired with an explicit statement of what would have to be true for a causal read to hold — and usually isn’t.

Finding 1
Discount exposure and purchase
What did we observe?
Sessions exposed to a discount purchased at a rate of 33.5%, versus 22.3% for sessions with no discount exposure — purchase probability is higher among discount-exposed sessions in this observational dataset.
What statistical concept supports this?
Conditional probability / two-group comparison: P(Purchase | Discount) vs. P(Purchase | No Discount).
What does it mean?
Discount exposure and purchase move together in the data we have: sessions where a discount was shown are more likely to end in a purchase than sessions where it wasn't.
What business decision could it inform?
Worth testing (e.g. via a randomized holdout or A/B test) whether proactively surfacing discounts to under-converting segments lifts purchase rate, before committing to a broad discount-allocation strategy.
What assumption or limitation should a decision-maker know?
Discount exposure was not randomly assigned in this dataset. Customers who were already more interested in buying may have been more likely to seek out or qualify for a discount (self-selection). We cannot conclude “discounts cause purchases” from this comparison alone — only that the two are associated.
Finding 2
Add-to-cart and purchase
What did we observe?
P(Purchase | Added to Cart) is 55.0%, far above the baseline P(Purchase) of 27.5% across all sessions.
What statistical concept supports this?
Bayes' theorem / conditional probability, comparing a conditional rate against its unconditional baseline.
What does it mean?
Once a session includes an add-to-cart action, the odds that it ends in a purchase rise sharply compared to a session picked at random.
What business decision could it inform?
Prioritize cart-abandonment interventions (reminder emails, retargeting, saved-cart nudges) for sessions that reach the cart stage but stall, since this subgroup already carries much higher intent than the general session population.
What assumption or limitation should a decision-maker know?
A higher conditional purchase rate reflects stronger intent among cart-adding sessions, not a guarantee that any individual cart will convert, or that an intervention will change the outcome for that individual.
Finding 3
Customer segment differences in conversion
What did we observe?
Conversion rate rises across behavioral segments: New 20.5%, Occasional 26.4%, Regular 30.1%, Loyal 39.0%.
What statistical concept supports this?
Categorical association / group comparison across a multi-level segment variable.
What does it mean?
Sessions from customers with more purchase history convert at a higher rate than sessions from newer customers, in this dataset.
What business decision could it inform?
Allocate marketing spend and messaging by lifecycle stage — e.g. heavier trust-building and incentive content for New/Occasional segments, retention-focused content for Regular/Loyal — rather than a single blanket campaign.
What assumption or limitation should a decision-maker know?
The segment itself is defined by past order count (previous_orders), so part of this gap is definitional: Loyal customers have already demonstrated repeat buying behavior. This is not evidence that a New customer would convert at Loyal-segment rates if simply given a Loyal-segment treatment.
Finding 4
Return rate differences across categories
What did we observe?
Footwear has the highest observed return rate at 16.3%, while Grocery has the lowest at 7.7%.
What statistical concept supports this?
Categorical association / comparison of proportions across product category.
What does it mean?
Return likelihood is not uniform across categories: Footwear orders in this dataset are returned meaningfully more often than Grocery orders.
What business decision could it inform?
Prioritize return-reduction investment (sizing guides, clearer product imagery/descriptions, tighter quality checks) toward Footwear, where the return-rate gap suggests the largest potential recovery.
What assumption or limitation should a decision-maker know?
Return rate here is descriptive of this sample's mix of price bands and products within each category, not a controlled, product-quality-adjusted comparison. Differences could partly reflect price point, sizing complexity, or product mix rather than category-level quality alone.
Finding 5
Acquisition channel differences in conversion
What did we observe?
Email sessions convert at 31.3%, compared with 24.8% for Affiliate — the highest and lowest converting channels in this dataset.
What statistical concept supports this?
Categorical association across acquisition channel, a multi-level categorical variable.
What does it mean?
Sessions acquired through different channels convert at meaningfully different rates in this dataset.
What business decision could it inform?
Use channel-level conversion as one input (alongside cost-per-session and margin) when reallocating acquisition budget across channels, rather than optimizing purely on traffic volume.
What assumption or limitation should a decision-maker know?
Channel conversion differences may reflect audience intent differences at acquisition rather than channel effectiveness in isolation — e.g. Email traffic often already carries higher purchase intent than colder discovery channels, so this comparison should not be read as a pure measure of channel quality.
Across all five findings on this page, the underlying data is observational: variables were not randomly assigned by an experiment, so every relationship described above is an association, not a demonstrated causal effect. Decisions informed by these findings should ideally be validated with a controlled test before being scaled.