Level 6 · Boss: who converts on the website?

A funnel dataset

web_sessions has one row per visit to the website: when it started, the device, the traffic source, how many pages the visitor viewed and whether the visit ended in a purchase. It’s 40,000 rows, and every question in this level is a filter followed by a count, a share or a sort.

Step 1: count with a mask

You don’t need to filter rows to count them. A mask is a column of True and False, and True counts as 1:

mask = (
    (web_sessions["device"] == "mobile")
    & (web_sessions["source"] == "paid")
    & web_sessions["converted"]
)
mask.sum()

Notice the last condition has no == True. converted is already boolean, so it’s a mask on its own. Writing web_sessions["converted"] == True works too, but it’s redundant. .sum() on a mask returns a NumPy integer; wrapping it in int() gives a plain Python number.

Step 2: a rate is a mean

The share of sessions that converted is the number of True values divided by the number of rows, which is exactly the mean of a boolean column:

engaged = web_sessions[web_sessions["pages_viewed"] >= 10]
engaged["converted"].mean()

This is one of the most useful tricks in pandas. Conversion rate, return rate, share of orders with a coupon: each is .mean() of the right mask, on the right set of rows. The “right set of rows” is the important part. Conversion “among engaged sessions” has a different denominator from conversion over all sessions, and mixing them up is one of the most common errors in funnel reporting.

Compare it with the overall rate, web_sessions["converted"].mean(). The engaged rate is much higher, which is expected: someone who browses ten pages is far more likely to buy. That doesn’t mean making people view more pages would make them buy; engagement and buying intent move together. Keep correlation and cause separate when you present this.

Step 3: a ranked list

email = web_sessions[web_sessions["source"] == "email"]
(
    email.sort_values(["pages_viewed", "session_id"], ascending=[False, True])
    .head(15)[["session_id", "device", "pages_viewed", "converted"]]
    .reset_index(drop=True)
)

Many sessions share the same page count, so session_id as the tie-breaker makes the list reproducible. Filter first, then sort: sorting 40,000 rows and filtering later gives the same answer but does more work, and it’s easier to make mistakes in the order of steps.

Reading the results like a marketer

• A conversion count without a denominator is hard to judge. “51 paid mobile conversions” means little until you know how many paid mobile sessions there were.
• Sessions without a customer_id are anonymous visitors. In this dataset, only known customers convert, which mirrors how many stores track purchases.

Common mistakes

• Using and between masks instead of &.
• Computing the step 2 share over all sessions instead of the engaged ones.
• Sorting by pages_viewed only, so the 15th row changes between runs.