Marketing attribution modelling

Attribution modelling splits the credit for each sale among the channels a customer touched. You see last click, first click and linear side by side on the same conversions, so you know how much of a channel's figure comes from the choice of model.

Last reviewed

One month, three models, example

Paid social: 60 conversions under last click, 510 under first click

Example data
  • Paid search

    Last click 410First click 230Linear 355

    230 to 410

  • Paid social

    Last click 60First click 510Linear 255

    60 to 510

  • Email

    Last click 430First click 40Linear 210

    40 to 430

  • Organic search

    Last click 100First click 220Linear 180

    100 to 220

Linear gives each touch on a path an equal share of the order.
Conversions
1,000 orders in one month, from the order system
Lookback window
30 days, held fixed
Same paths under every model
1,000 orders, the same paths and the same 30-day window under every model.
ChannelLast clickFirst clickLinearRange
Paid search410230355230 to 410
Paid social6051025560 to 510
Email4304021040 to 430
Organic search100220180100 to 220
Total1,0001,0001,000
Example data: an invented month for an online shop, eleven kinds of path. The customers and their paths are the same in every column. The rule for sharing credit changes, and paid social's count moves by 450 orders.
Show the numbers
The path table, example data. S paid search, P paid social, E email, O organic search. Touches in order, first to last.
PathOrders
S140
E40
O100
P60
P → S120
P → E150
S → E90
O → S60
P → S → E90
P → O → S90
O → S → E60
Total1,000

What this fixes

Each team reports its own modelSocial quotes first click, search quotes last click, and together they claim more orders than the shop took.
The window moved and nobody saidSomeone shortened the lookback window in spring, and every channel's trend bends in the same week.
The model no one can explainCredit shifted when the account moved to a data-driven model, and no one can say which paths moved it.
Credit quoted as proofLast click gives email a big number that gets used to defend its budget. In the example above, 390 of email's 430 last-click orders had touched another channel first.

What an attribution modelling project hands over

The path tableIt holds one row per order from the order system, with every recorded touch inside the window from GA4's BigQuery export or your own event data.
The three-model comparisonEach channel's count appears under last click, first click and linear, refreshed monthly.
The model noteOne page records the reporting model, the lookback window, the conversion and the date each was set.
The monthly channel tableEach channel's figure appears with its range across the models and its rule, as figures do in the KPI library.
The test shortlistChannels rank by the width of their range, and the widest is the first candidate for a holdout test.

Paid social goes from 60 orders to 510 when the model changes, and not one customer did anything differently.

How a marketing attribution model is set up

  1. 1

    Scoping

    We agree the conversion, orders net of cancellations from the order system, then the channels and the decision the numbers feed.

  2. 2

    Path table

    Touches join to orders, the lookback window cuts them, and orders with no path stay in view.

  3. 3

    Comparison

    The three models run on the same rows, and any channel whose range is wider than the decision can absorb goes on the shortlist.

  4. 4

    Handover

    The SQL, the path table and the model note are in your warehouse, and the monthly refresh runs without us.

Where multi touch attribution goes wrong

Credit read as causeRule-based models share out orders among the channels that were there, and none can say how many would have come anyway. A holdout test on the channel the decision hangs on answers that, as part of marketing measurement.
A window that driftsCut the lookback from 30 days to 7 and every channel that comes early in a path loses credit overnight. The model note records the model and window, and a change starts a new series.
Paths with holesDeclined cookies, a switched device or a poster leave touches the path table never sees. Every comparison shows the count of orders with no path, and if it grows, the fix is tracking.

Every column in the comparison adds up to the same orders, so the only thing that changes between columns is the rule.

Attribution modelling for ecommerce

EcommerceEach order has a timestamp and a customer, so it matches to its touches and the order system stays the count.

Attribution splits credit between channels. What each channel adds to sales in total, TV and radio included, is a question for marketing mix modelling.

Questions people ask

Which attribution model should we use?

The one that matches the decision: first click for which channels find new customers, last click for what closes the sale. Report on one with its lookback window written down, and show the other two beside it.

Do you set up marketing attribution for ecommerce brands?

Yes. Each order from the order system, net of cancellations, gets matched to its touches inside the agreed window, and orders with no path stay counted and shown.

What about data-driven attribution?

It goes in as a fourth column if your GA4 property or Google Ads account uses it. It still only shares credit among the touches your tracking recorded, so only a holdout test shows cause.

Is multi touch attribution the same as incrementality testing?

No. Attribution shares out credit for conversions that happened, and a holdout test counts the orders that go missing when a channel pauses. The channel with the widest range across models is the first to test.

How long should the lookback window be?

Long enough to cover first touch to purchase in your own data. We set it where the path table's curve of days to order flattens, and it holds until a dated line in the model note changes it.

Do two of your teams credit the same channel differently?

On the call we look at whether your order data and touch data can be joined into one path table.

Book a call

Free 30 min call