Attribution assigns credit for a conversion to the marketing touchpoints that preceded it. Every model does this differently, every model is wrong, and the useful skill is knowing which direction each one is wrong in.
The models
First touch. All credit to the first interaction. Over-credits awareness channels and ignores everything that happened afterwards. Useful for understanding what creates demand.
Last touch. All credit to the final interaction. The default in most analytics tools and the most misleading in B2B, because the last touch before a purchase is frequently a branded search — which the earlier channels created. Last-touch reporting reliably makes brand search look like your best channel and content look worthless.
Linear. Equal credit to every touchpoint. Simple and fair-seeming, and it treats an incidental newsletter open the same as the demo that closed the deal.
Time decay. More credit to touchpoints closer to conversion. Reasonable for short sales cycles, and it systematically undervalues early-stage content in long ones.
Position-based. Typically 40% to first, 40% to last, 20% spread across the middle. A pragmatic compromise that credits both demand creation and closing.
Data-driven. Uses the platform’s own modelling to assign credit based on observed patterns. Better in principle, opaque in practice, and it requires substantial conversion volume to produce stable results.
Why all of them are wrong
Three structural problems that no model solves.
Not everything is trackable. A recommendation in a private community, a conversation at a conference, a podcast listened to in a car. These influence purchases and appear in no attribution report. In B2B this is a large share of real influence.
Correlation is not contribution. A touchpoint appearing before a conversion was not necessarily a cause. Branded search is the clearest example: the person was going to buy, and searching your name was how they navigated there.
Privacy changes have degraded the data. Cookie restrictions, tracking prevention, mobile app privacy controls and consent requirements mean a meaningful share of journeys are now incomplete or invisible. Models built on complete journey data are modelling something that no longer exists.
Using attribution sensibly
Three practices make it useful despite the limitations.
Compare models rather than choosing one. Run first touch and last touch side by side. A channel that looks strong in first touch and weak in last touch is creating demand; the reverse is capturing it. Both are necessary and the gap is the insight.
Use it directionally, not for precise budget allocation. Attribution can tell you a channel is roughly three times more productive than another. It cannot tell you it is 2.7 times, and treating the decimal as real leads to overconfident decisions.
Validate with experiments. The only way to know a channel’s genuine contribution is to turn it off and see what happens. Geographic holdouts, time-based pauses and audience splits all work, and all beat any model.
Incrementality, which is the real question
Attribution asks which touchpoints preceded the conversion. The question that matters is which conversions would not have happened without the spend.
A branded search campaign may be credited with a great deal of revenue and contribute almost none, because those people would have found you anyway. A retargeting campaign may show excellent attributed returns while mostly reaching people who had already decided.
Testing it is straightforward in principle: pause the channel in half your markets for a defined period and compare total conversions, not attributed ones. The results frequently surprise people, and they are more actionable than any model output.
Self-reported attribution
An underused and effective supplement: ask new customers how they heard about you, as a free-text or lightly structured field at signup or on a sales call.
The data is messy and biased by recall, and it captures the channels attribution cannot see at all — podcasts, word of mouth, communities, conferences. Compared alongside platform attribution, it fills exactly the gaps the platforms have.
Ask at the point of purchase rather than later, keep it optional, and read the free text rather than only the categories.
B2B specifically
Long cycles and multiple stakeholders break most attribution tooling.
Four adjustments help. Attribute at the account level rather than the contact level, since several people from one company will each have their own journey. Extend the lookback window to cover your actual sales cycle, which is frequently longer than the default 30 or 90 days. Track influence on pipeline rather than only on closed revenue, so you get signal before the cycle completes. And accept that the picture will be incomplete.
A workable reporting setup
For most organisations: a first-touch and last-touch comparison as the standing report, self-reported attribution collected at signup, one incrementality test per quarter on the largest spending channel, and a simple blended efficiency measure — total marketing spend divided by total new customers — as the sanity check that no model can distort.
That last measure is unsophisticated and hard to argue with, which is exactly its value in a conversation where every channel owner has a model that favours their channel.
Marketing mix modelling, briefly
As individual-level tracking degrades, aggregate approaches have returned to favour. Marketing mix modelling uses statistical analysis of spend and outcomes over time, across channels, to estimate each channel’s contribution without tracking individuals at all.
The advantages are real: it is privacy-robust, it captures offline and untrackable channels, and it measures incremental contribution rather than correlation with a conversion.
The limitations are equally real. It needs a long history — typically two years or more of weekly data — and enough variation in spend to separate the channels. It produces channel-level answers rather than campaign-level ones, and it cannot tell you which advertisement to run.
For most mid-sized organisations it is not yet practical, and the holdout-test approach gets you most of the way for a fraction of the cost. For larger advertisers with substantial spend across several channels, it is increasingly the primary method.
Presenting attribution to a sceptical audience
Attribution reports often cause arguments because each stakeholder arrives with a model that favours their channel.
Three habits defuse this. State the model and its known bias on the slide, every time. Show two models side by side rather than one, which makes the uncertainty visible and prevents the number being treated as fact. And include the blended efficiency measure — total spend over total new customers — as an anchor that no model can inflate.
The goal of the conversation is a budget decision, not a correct number. Framing it as “this is roughly the picture and here is how we will test it” produces better decisions than defending a figure nobody can verify.
Attribution within a single channel
The models above allocate credit between channels. A related question is how to judge performance inside one, and the answer is usually the channel’s own diagnostics rather than the attribution model.
In paid search, for example, the relevant signals are search terms, win rate against budget and the relevance measures the platform exposes — see Quality Score. Using a cross-channel attribution model to tune a single campaign imports all the model’s bias into a decision that did not need it.
Frequently asked questions
Which attribution model should I use?
Compare at least two rather than choosing one. If forced to pick a single default, position-based is a reasonable compromise for B2B.
Is data-driven attribution better?
In principle, where there is enough conversion volume. In practice it is opaque, varies between platforms, and each platform’s version tends to favour that platform’s channels.
How long should the lookback window be?
Match it to your actual sales cycle. If deals take four months, a 30-day window attributes almost nothing correctly.
How do I measure channels with no tracking at all?
Self-reported attribution and holdout tests. For a podcast or event, compare traffic, branded search volume and signups in the period against a baseline.
