Outgrace, home

How marketing attribution models differ, and where each misleads

Each attribution model divides credit for a sale by a different rule. Before any budget moves, know what each rule rewards and what it cannot see.

Ask three people in one marketing team which channel brought a sale and you will probably get three answers. Each one may be right under the rule that person uses. An attribution model is that rule: a fixed way of sharing credit for a conversion among the touches that came before it. Pick the rule first and the answer is decided before anyone opens a report.

How it works

Every model answers the same question: how should credit for a sale be divided among the marketing touches a buyer met on the way? A touch might be a paid search click, an email, a conference badge scan or a call from a salesperson. The models differ only in the division rule, and each rule carries an assumption that is easy to forget once the dashboard is live.

Last-click gives all the credit to the final touch before the sale. It is simple, it matches what the checkout system records, and it is why branded search and retargeting so often look excellent. Its blind spot is the buyer who first heard of the company from a podcast or a trade article, then returned three weeks later by searching for the brand name. Last-click pays the search and forgets the podcast.

First-click does the reverse and credits the first recorded touch. It is useful for seeing which channels bring new people in, and it overvalues discovery. A channel that starts many journeys but never closes them will look like a star, while the nurture work that turns interest into a purchase receives no credit at all.

Linear models split the credit equally across every touch. They feel fair, which is their appeal, but they assume each touch matters as much as every other. A ninth email and the single demonstration that won the account receive the same share. Linear rewards long, busy journeys and tends to flatter channels that appear often and do little.

Time-decay models give more credit to touches closer to the sale and less to earlier ones. The idea is sound: recent contact is usually more influential than contact long ago. The trouble is the decay rate, which someone has to choose. Set it differently and the ranking of channels changes, yet the report presents the output with the same confidence whatever the setting.

Position-based models, sometimes called U-shaped, fix a share for the first and last touches and divide the remainder among the middle ones. The shares are a convention the team agreed, not a finding. They are easy to explain to a board and easy to defend in a meeting, and they can still be wrong by a wide margin for any particular business.

Data-driven or algorithmic models estimate credit from the paths of buyers who converted and of those who did not. They can capture patterns that no hand-written rule would encode. Their weaknesses are practical. They need a large volume of clean, linked journeys, and they are only as good as the tracking behind them. Where cookies are blocked, devices are not linked, or offline sales leave no record, the model is confidently fitting an incomplete picture.

All of these models describe association within recorded paths. None of them, on its own, shows what a channel caused. A buyer who was already going to purchase may have clicked an advert on the way, and the model will give that advert credit. Measuring the effect of spending requires a comparison that attribution does not provide, usually a holdout or a controlled test in which some audiences see less of a channel and the outcomes are compared.

A useful habit is to write the blind spot of each model beside its output, in the same plain words every time. For last-click the note might read: credits the final recorded touch and ignores earlier contact. For a linear model: credits every recorded touch equally, whether or not it mattered. The discipline is less about choosing the perfect model than about refusing to read any model's output without its note.

Who it affects

The problem falls hardest on companies that sell through several channels and have one person who owns measurement. Take a hypothetical manufacturer with a paid search budget, a trade show programme and a small email list. Its marketing lead runs last-click because the CRM makes it easy. The trade show team, which feeds the early stages of the buying process, sees its results credited to nobody. The arguments that follow are about the rule, not the people, and they recur in every budget round.

Leaders who sign off budgets are affected too. A board pack that shows return by channel looks like evidence. In most cases it is a summary of a rule. A chair who asks which model produced the numbers, and what that model cannot see, will reach a better decision than one who asks only for the total.

Agencies and software vendors often supply the model as a default setting. The default is a decision too, and it is rarely the one the business would have chosen on purpose. Ask who set it, when it was set, and whether your own team can change it without a new contract or a support ticket.

What leaders get wrong

  • They treat the model as neutral. Every attribution model encodes a choice about credit, and someone made that choice.
  • They compare channels across models without noticing. A ranking under last-click can differ from a ranking under linear, and the report rarely says which one it used.
  • They move budget on small differences. Two channels separated by a few points in a model's output are not reliably different, and the model's own assumptions can explain the gap.
  • They skip the question of what was never tracked. Offline influence, word of mouth and sales conversations sit outside the data, and the model quietly assigns them zero.
  • They let the model's history decide the future. A data-driven model trained on past spending learns where budget already went, so it may favour the channels that were funded before.

Most of these mistakes are not technical. They come from reading a number as a fact when it is an output of a rule. The fix starts with a plainer question: what was this rule built to answer? If the answer is which channel first introduced a buyer, the report should not be used to decide which advert to keep running this month.

What to do about it

Choose the model before the data arrives, and write its blind spot in one sentence beside it. A sensible test is whether the finance lead could explain, without notes, which buyers the model ignores. If nobody can answer that, the model is not yet understood well enough to move money. The choice should be recorded in the same document as the budget, so that it can be questioned later.

Next, decide which decision the model is for. A model that shows where buyers first meet the brand is a different tool from one that shows which campaign to keep this month. Many teams use one report for both purposes and get poor answers to each. Separate the questions, and let each model answer only the one it was built for.

A useful exercise is to take one period of data and run it through two models. If the top three channels are the same under both, the decision is less sensitive to the rule. If they differ, the choice of rule has become the decision, and it should be made by someone who can be held to it, with the reasoning written down.

Then test the largest budget line with a comparison that does not depend on the model. Hold back spend from a sample of comparable audiences for long enough to see a difference, and compare the outcome with what the attribution report predicted. Where the two disagree, believe the test, and investigate why the model was wrong. The model is then useful for what it does well, such as tracking how paths change over time, rather than for judging causes it cannot see.

Keep the attribution report, but place it beside a short note: which model was used, which touches were not captured, and which decisions it should not drive. Readers who see the note will ask better questions than readers who see a clean table. Record every change to the model as well, because a silent change is the moment most reports stop being comparable with the year before.

If the tracking depends on customer data collected across websites, apps or devices, the rules on consent and retention are a question for counsel, not the marketing team. Put that question to them before the tracking is extended, not after it has been built into the budget process.

Newsletter

The Briefing

One email a week: the leaders we profiled, the ideas worth stealing, and nothing else.

Weekly, free, one click to leave. Read our privacy policy.