Conversion rate: what is being counted?

Eight orders do not necessarily mean eight buying visits. One session can contain several orders, and one recorded ID can appear in several sessions. Before reading a conversion percentage, establish whether it counts events, orders, sessions or recorded identifiers.

One week, several counts

Consider a fictional shop selling downloadable patterns during April 1–7, 2027, using UTC. Include every session in this shop during those seven complete days: 100 sessions associated with 60 distinct recorded IDs. The supplied record fixes session boundaries and assigns one ID to each session; none crosses the week's boundary. These are invented records, not website results or a Google Analytics export.

Here, a completed, paid order means the download was delivered and receipt of payment was confirmed within that week. Eight distinct orders meet both conditions, with no refunds by the closing time. Each order belongs to exactly one of the sessions below. No other eligible session contains a completed, paid order.

The six purchasing sessions in the invented record

Scroll across to see every column. Use the arrow keys when the table is focused.

The other 94 sessions belong to the remaining 55 IDs. The record contains repeat sessions and repeat purchases. It does not establish how many actual people those 60 IDs represent.

Name the fraction before comparing it

Three calculations describe this same week:

• Sessions containing a completed, paid order: 6 ÷ 100 = 6 percent. Each purchasing session counts once, however many orders it contains.
• Recorded IDs with at least one such order: 5 ÷ 60, or about 8.3 percent. ID A counts once despite ordering in two sessions.
• Completed, paid orders per session: 8 ÷ 100 = 0.08 orders per session, or eight orders per hundred sessions.

The last calculation is useful, but it does not establish that eight percent of sessions contained a purchase. Only six did.

An event is not an extra order

The fictional raw event log contains ten purchase-tagged rows. Eight carry distinct order references; two repeat references already present. Matching them to the supplied completion and payment records leaves eight completed, paid orders. The repeated rows do not create two more orders.

The same week's log also contains twenty Buy-button click records, including repeats. Those record button use. They do not, by themselves, establish delivery or payment. Counting them would answer a different question from the one the order records answer.

Read the product's own definition

Google Analytics' Data API documentation separately defines event count and session key event rate. The latter measures the proportion of sessions in which a key event occurred; the API also supports a rate for a specific key event. A rate covering any key event is not automatically purchase-only.

Google's key-event guidance allows collected events to be designated as important business actions. Calling a button click important does not make it a paid order. The event being counted still needs to be named.

Its reporting-identity documentation also describes User-ID, device identifiers and modeling. A report's user count needs that identity context before it is described as a count of people. The fictional ID calculation above is a defined measure for this record, not a substitute for every product's user metric.

When each group’s rate rises but the total falls

An unchanged conversion definition does not make the session mix irrelevant. A rate can rise within every supplied group while falling across the whole record.

Consider a separate invented shop record for May 1–7 and May 8–14, 2027, using UTC. Each complete seven-day period contains 200 sessions, with none excluded. Session boundaries stay unchanged; no session crosses a period boundary. Every session is assigned at entry to exactly one source group, A or B, using the same supplied classification in both periods. Together the groups contain every eligible session.

Count a session once if it contains at least one order whose delivery and payment are both confirmed within that period. Each such order belongs to exactly one session in its period. There are no refunds by either closing time. The table counts those sessions, not order-event rows or people.

Sessions containing a completed, paid order, divided by all sessions in each row

Scroll across to see every column. Use the arrow keys when the table is focused.

Both group rates rose. A went from 20 to 30 percent; B went from 10 to 12 percent. Yet the overall rate fell by one percentage point, with two fewer purchasing sessions among the same total of 200.

A has the higher rate in both periods, but its share of sessions fell from 75 percent to 25 percent. The combined rate weights each group by its share of sessions: 0.75 × 20% + 0.25 × 10% = 17.5 percent in the first period, and 0.25 × 30% + 0.75 × 12% = 16.5 percent in the second. Giving both group percentages equal weight would describe a different mix.

The Stanford Encyclopedia of Philosophy's discussion of Simpson's paradox explains how an association can reverse between grouped and combined data. The overall rate still correctly describes all sessions in each period.

For this record, a useful summary keeps the two findings together: the overall share of sessions with a completed, paid order fell, while the rate within each source group rose and the session mix changed.

These counts do not show that either source caused purchases or that the same people changed their behavior. There is no matched personal history or experiment in the supplied record. The source labels describe how sessions were grouped, not why an order happened.

Write the claim the record supports

Try repairing this summary: “Eight percent of visitors bought.” A supported version is: “Six of 100 recorded sessions contained at least one completed, paid order. Eight such orders came from five of the 60 recorded IDs during the week.”

Before comparing another period, put its event definition, eligible group, counting unit and time window beside this one. A higher rate cannot yet tell you that a campaign caused more purchases, that the orders were profitable or that the next week will repeat them.

For a memoir that keeps business figures beside the work and commitments they describe, find a starting point in Infinite Upside.