Field notes on attribution, from the point of view of someone who has watched budget meetings turn into arguments over a single column in a report.

Attribution is a rule for splitting the credit for a sale across the marketing touches that came before it. That is all. It is accounting, not discovery. Keep that in mind and most of the confusion goes away.

One customer, one sale, four answers

A shopper sees an Instagram ad for a ₹4,000 backpack. Two days later she searches the brand name on Google and clicks a paid result. Three days after that she opens a reminder email and buys. Three touches, one order.

  • Last click: email gets all ₹4,000. Search and Instagram get nothing.
  • First click: Instagram gets all ₹4,000.
  • Linear: each channel gets one third, about ₹1,333.
  • Position-based, often set at 40% first, 40% last, 20% spread over the middle: Instagram ₹1,600, search ₹800, email ₹1,600.

Same customer, same money, four reports. A time-decay model, which gives more weight to touches nearer the purchase, is a fifth. Every one of them is internally consistent and none is “correct”.

What each rule quietly rewards

Last click is the default in many tools because it is simple and needs only the final step. Its bias is plain: it favours channels that sit at the bottom of the funnel, such as brand search, email and retargeting, which mostly harvest demand that something else created. Cut the Instagram budget because it shows poor last-click sales and, a few months later, brand search volume falls too. The credit went to the wrong place and the money followed it.

First click does the reverse. It flatters awareness channels and ignores whatever turned interest into a purchase. Linear is democratic but treats a fleeting banner impression like a deliberate product comparison. Position-based and time-decay add judgement through weights, and the weights are your assumptions, not measurements.

Opinion: if you must choose a single model for reporting, last click is defensible only if you also look at something that sits outside it. Otherwise you are measuring who was standing nearest the till.

The problems underneath all of them

The data is incomplete. Journeys are stitched together using cookies, logged-in IDs or device identifiers. People switch phones and laptops, browsers restrict third-party tracking, consent choices remove people from the data (and India’s Digital Personal Data Protection Act, 2023 pushes organisations to be clearer about consent), and offline touches such as a hoarding or a shop visit are missing altogether. Any model works only on the journeys it can see.

Credit is not causation. Attribution says who was present, not who changed the outcome. A returning customer who was going to buy anyway still hands credit to whichever email she clicked last. The only way to know what a channel adds is to compare against a situation without it.

Platforms mark their own homework. Each ad platform counts conversions using its own window and rules, so adding up what Google, Meta and the email tool each report often exceeds your actual orders. If the sum of reported sales is more than ₹5 crore and your finance team shows ₹4 crore, the extra ₹1 crore is double counting across tools, not magic.

Ways to get closer to the truth

Incrementality tests are the cleanest answer. Hold out a random group from seeing a channel’s ads, or switch a channel off in a few matched cities, and compare sales with the groups that still see it. The difference is what the channel truly adds. It costs a little revenue during the test but gives you something attribution cannot.

Marketing mix modelling uses aggregated spend and sales over time, with seasonality and price controlled, to estimate each channel’s contribution. It does not depend on tracking individuals, which is useful given the privacy shifts, but it needs a couple of years of reasonably varied spending data and careful interpretation.

A practical stance for a small team: report last click for daily operations, review a multi-touch view monthly to see which channels assist, and run a holdout test once or twice a year on the biggest budget line. When the three disagree, that disagreement is useful information, not an error to be fixed. Where the budget is large, trust the experiment over the model.

Figures and names in the examples are illustrative.

Leave a Reply

Trending