Level 5 · What last-touch credit really says

How conversions get credited

In web_sessions, a purchase shows up as converted = true on one session, and that session has one source. Grouping conversions by source credits each purchase to the last touch: the channel of the visit where the purchase happened.

A real customer journey usually has several visits. Someone sees a paid ad, browses anonymously, signs up, gets the newsletter a week later and buys from the email link. Last-touch credit gives the whole purchase to email and nothing to paid.

Why email looks cheap

The 2025 numbers, spend aggregated per channel before dividing:

38_322.22 / 187    # email:  $204.93 per conversion
340_778.33 / 55    # paid:   $6,195.97 per conversion

Now look at who each channel reaches. 89% of email sessions belong to a known customer, against 29% of paid sessions. In this data only known customers can convert. Email talks to people who already buy from Harbor & Pine; paid mostly brings strangers. The cost gap is partly real efficiency and partly a difference in audience that the attribution rule turns into credit.

What the data cannot tell you

• Orders. Sessions carry no order id. 2025 has 1,805 booked web orders and 500 converted sessions, so the session data does not even cover every web order.
• Earlier touches. A paid visit that led to a later email purchase is invisible in a last-touch count.
• What happens if you move budget. Cost per conversion is an average over what you already bought. Email cannot reach more people than are on the list; cutting paid may shrink the list over time.

How to say it to Finance

“Under last-touch credit, email cost $205 per converted session in 2025 and paid search $6,196. Email mostly reaches existing customers, so the gap overstates its efficiency for growth. To learn what paid really adds, pause it in a few regions for a month and compare.”

Interview note

“What are the limits of last-click attribution?” is a common marketing-analytics question. Name the bias (credit to the final touch, so retention channels look best), give a number like the one above, and suggest a holdout test as the way to measure what a channel actually adds.