Payments are experiencing issues due to temporary restrictions in Russia. If your payment does not go through, please submit a support request.Our support team is available 24/7 — we are always here to help with hosting and server issues.We are now accepting requests for dedicated server rental and colocation services in our data center.Reminder: we recommend enabling backups for additional data protection.A new VPS/VDS lineup with NVMe storage and improved performance is now available.Maintenance work on some servers has been completed. All services are operating normally.
Article5 min readViews0

Conversion grew in every segment but fell overall: checking traffic composition

Tutorial example with computers and phones: 4.4% converts to 3.6% despite growth in both groups. We calculate weights and separate the report explanation from proof of causes.

Comments 0

Tutorial diagram: session shares for PCs and phones shift from 80/20 to 20/80, while overall conversion drops from 4.4% to 3.6%
In this article

The store's overall conversion fell from 4.4% to 3.6%. Yet, the share of sessions with at least one order increased on both computers and phones. Such a report does not necessarily contain an arithmetic error: the proportion of the groups themselves may have changed. Before tackling the task of restoring the previous conversion rate, it is necessary to restore the traffic composition.

Let us examine a fully hypothetical example. This is not General iT data, nor is it the result of an experiment. Its purpose is to show why the sum of indicators can sometimes move in the opposite direction to each individual segment, and what calculation to request from an analyst.

First, let us agree on what we are counting

In the example, conversion equals the number of sessions with at least one recorded order divided by all considered sessions. If two orders occur in a single session, the numerator still counts as one session. This is not the number of orders per visitor, nor the share of paid purchases. These indicators must not be substituted when comparing.

Google Analytics documentation describes session share of a key event similarly: sessions with the key event are divided by all sessions. However, the specific event name and the conditions for its registration determine the meaning of the result. A checkout event does not prove payment or receipt of goods.

For simplicity, we have only two non-overlapping groups covering all 1,000 sessions per period: computers and phones. Tablets and unknown devices are absent in this model. In a working report, they must not be silently skipped: the sum of the groups must match the selected total.

Within groups, the rate is generally lower

In the first period, computers account for 800 sessions. A checkout was recorded in 40 of them: 40 out of 800, or 5%. Phones account for 200 sessions, with 4 checkouts: 2%. Together, we get 44 sessions with a checkout out of 1,000, or 4.4%.

In the second period, the composition changes: 200 sessions from computers and 800 from phones. On computers, a checkout occurred in 12 sessions, so conversion is 6%. On phones, there were 24 such sessions, so conversion is 3%. Both groups increased by one percentage point. But the overall total is 36 sessions with a checkout out of 1,000, or 3.6%.

The cause is visible in the weights. In the second period, far more sessions fall into the group with the lower conversion rate. When calculating the overall result, 3% is assigned a weight of 80%, while 6% is assigned a weight of 20%. A simple average of the two percentages, 4.5%, does not describe this population: the groups differ in size.

The decline in the overall metric here is 0.8 percentage points. This is not a 0.8% decrease relative to the previous value. For relative change, one would divide the difference by the original 4.4%; it is important to choose the comparison method in advance and not mix units in a single report.

Recalculation with previous weights

To see the influence of group metrics separately, we retain their shares from the first period: 80% computers and 20% phones. Substituting the second-period conversion rates: 0.8 × 6% + 0.2 × 3% = 5.4%. This is a standardized calculation assuming a constant composition, not the actual conversion rate of the second period.

If we imagine the same 1,000 sessions with the previous distribution, the arithmetic yields 48 sessions with an order on computers and 6 on phones, totaling 54. These 54 are not additional orders received or a forecast. We are only changing the weights in the formula to explain the difference between the total and the segments.

It is useful to place three figures side by side: the actual 4.4% for the first period, the actual 3.6% for the second, and the hypothetical 5.4% assuming the same composition. The last figure helps frame a question about changes in traffic sources and audience, but it does not negate the fact that the number of recorded sessions with an order has decreased.

What this calculation does not prove

One cannot conclude from this example that the new version of the site improved sales. Within each group, traffic sources, product assortment, seasonality, and the share of returning customers could have changed. We did not conduct random assignment between versions, nor did we assess statistical uncertainty. A positive difference within a segment does not, by itself, establish causality.

One also cannot declare phones a problem solely because their share increased. A new mobile audience may be beneficial to the business, even if the path to purchase differs. To address this, comparable data on acquisition costs, payments, and order outcomes are required. The sample table provides no information about the profitability of these groups.

Before discussing design, verify that data collection conditions are identical: the event, filters, period boundaries, time zone, and device attribution rules. Cross-check the registration of multiple permitted test cases against the expected result. Changing a counter setting can render metrics incomparable, even if the formula is correct.

For the analyst task, I would ask to show the numerator and denominator for each segment, the group shares, and the recalculation with fixed weights. Based on these data, the next check can be selected: changes in audience composition, a specific checkout step, or the quality of event registration. A single overall percentage is suitable for monitoring the final result; the budget for fixes should be assigned only after explaining what it consists of.

Discussion 0

Share your experience and ask questions. Comments without links appear after editorial review.

No comments yet. Start the discussion.