Attribution windows for messaging

Attribution is the question of which message gets credit for which outcome, and the answer depends almost entirely on how long your customers take to decide. Most businesses inherit a default window from an analytics tool and never revisit it.

Match the window to the decision

First touch, last touch, or both

For a small business the honest answer is usually to record both and report last touch, because last touch is the one you can act on. But keep first touch in the data, because it is what tells you whether messaging is *starting* conversations or merely closing ones that email started.

A programme where messaging is always last touch and never first is doing less than it appears — it is harvesting demand rather than creating it, and that is worth knowing before you scale it.

The practical implementation

1. 1Tag every link with a consistent scheme, so the outcome carries the message identity. The scheme.
2. 2Record the send with its play, contact and timestamp. You cannot attribute what you did not log.
3. 3Record the outcome with its own timestamp, on the same contact.
4. 4Join with an explicit window rather than "most recent", so the window is a decision you can change and re-run.
5. 5Keep an unattributed bucket. Outcomes with no message in the window are real information, not a bug.

-- Last-touch attribution with an explicit window.
with attributed as (
  select
    o.id            as outcome_id,
    o.kind          as outcome_kind,
    m.play,
    m.sent_at,
    row_number() over (
      partition by o.id order by m.sent_at desc
    ) as recency
  from outcomes o
  join messages m
    on m.contact_id = o.contact_id
   and m.sent_at <= o.occurred_at
   and m.sent_at > o.occurred_at - interval '30 days'  -- the window
)
select play, outcome_kind, count(*) as outcomes
from attributed
where recency = 1
group by play, outcome_kind
order by outcomes desc;

Test how much the window is doing

Run the same query at 7, 30 and 90 days. If your conclusion flips between them, your result is a function of the window rather than of the channel, and you need a held-back control group instead of better attribution. How to run that comparison.