Level 8 · Conversion by device and source

Counts and rates from a boolean column

converted is True or False for each session. That makes it two measures in one:

web_sessions["converted"].sum()    # 1193 sessions converted
web_sessions["converted"].mean()   # 0.0298, about 3% of all sessions

Inside a groupby the same trick works per group. One agg with two keys builds everything Priya needs:

combos = web_sessions.groupby(["device", "source"], as_index=False).agg(
    sessions=("session_id", "size"),
    conversions=("converted", "sum"),
    conv_rate=("converted", "mean"),
)

The mean of converted is exactly conversions / sessions, so you don’t need a separate division. Five sources times three devices gives 15 combinations.

Filter the groups, not the rows

The rule “at least 1,000 sessions” is about groups, so it has to come after the aggregation. The summary is an ordinary DataFrame, so a normal mask does it:

combos[combos["sessions"] >= 1000]

Four combinations drop out, all on tablets: email (447 sessions), social (642), direct (599) and paid (876). Tablets are only about a tenth of the traffic, so they rarely reach the threshold. Tablet email is the interesting loss: 41 conversions out of 447 is 9.2%, which would rank second. It may be real, but a few lucky days could move it a lot. That’s exactly why Priya asked for a floor.

Reading the ranking

Email converts best on every device, and desktop email leads at 12.8%. Social is last everywhere, with mobile social at 0.4%. The biggest combination, mobile organic with 7,252 sessions, converts at only 1.5%. Lots of traffic and a strong rate are different things, and a budget discussion needs both columns side by side. Keep sessions and conversions in the table even though the ranking only uses the rate: they let anyone reading it judge how solid each number is.

Common mistakes

• Filtering before grouping, for example on pages_viewed, which changes the rates themselves.
• Using "count" on converted for conversions. It counts sessions, not True values.
• Averaging per-device rates to get a per-source rate. Groups have different sizes, so divide totals instead.
• Writing the threshold as > 1000 when the ticket says “at least 1,000”. On other data that would drop a combination with exactly 1,000 sessions.