Attribution Modelling: What the Models Won’t Tell You

Attribution modelling is the practice of assigning credit to the marketing touchpoints that contribute to a conversion. At its simplest, it answers the question: which channels and interactions drove this sale? In practice, it is one of the most contested, imperfect, and commercially consequential decisions a marketing team makes.

Every attribution model is a set of assumptions wearing the clothes of data. The model you choose does not reveal the truth about your customer experience. It reflects a particular theory about how customers make decisions, and that theory has consequences for where budget goes next.

Why Attribution Modelling Matters More Than Most Teams Admit

Attribution is not a reporting exercise. It is a budget allocation mechanism. Whatever model you apply to your conversion data will directly shape which channels receive more investment and which get cut. That makes it one of the highest-stakes analytical choices in performance marketing, yet most teams treat it as a default setting they never revisit.

I spent years running agencies where paid search consistently appeared as the top-performing channel in every client report. The reason was almost always the same: last-click attribution. Paid search sits at the bottom of the funnel. It catches people who have already decided to buy. Of course it looks efficient. The model was designed to make it look that way.

The problem is not that paid search is unimportant. It often is important. The problem is that last-click attribution makes it look like the only thing that matters, which means display, social, content, and email get starved of budget because they cannot claim the final click. Over time, you hollow out your upper funnel and wonder why your paid search costs keep rising. You have been competing against yourself.

If you want to go deeper on how analytics decisions connect to broader marketing measurement, the Marketing Analytics and GA4 hub covers the full landscape, from tracking setup to budget decisions.

What Are the Main Attribution Models and What Do They Actually Assume?

There are several attribution models in common use. Each one encodes a different theory about how customers behave.

Last-click attribution gives 100% of the credit to the final touchpoint before conversion. It is simple, easy to implement, and deeply misleading for most businesses with any meaningful customer experience. It tells you what channel closed the sale, not what built the intent.

First-click attribution gives 100% of the credit to the first touchpoint. This is useful if you are trying to understand awareness and acquisition, but it ignores everything that happened between introduction and conversion. It is the mirror image of last-click, with the same structural flaw.

Linear attribution distributes credit equally across every touchpoint in the experience. It sounds fair, but it treats a brand awareness display impression the same as a high-intent product page visit driven by paid search. Equal weighting is not the same as accurate weighting.

Time-decay attribution gives more credit to touchpoints closer to the conversion event. This makes intuitive sense for short sales cycles where recency genuinely signals intent. For longer, more considered purchase journeys, it still undervalues the early interactions that created the opportunity in the first place.

Position-based attribution, sometimes called the U-shaped model, splits credit between the first and last touchpoints (typically 40% each) and distributes the remaining 20% across the middle. It acknowledges that both acquisition and conversion matter, which is a more honest starting point than single-touch models.

Data-driven attribution uses machine learning to assign credit based on observed patterns in your own conversion data. Google’s version, now the default in GA4, trains on your account’s historical data to weight touchpoints according to their statistical contribution to conversion. It sounds like the obvious answer. It is not without problems, which I will come to shortly.

The Problem With Data-Driven Attribution Specifically

Data-driven attribution has become the default recommendation from most platforms and a lot of agencies. The pitch is compelling: let the algorithm figure it out. Remove human bias. Let the data speak.

The issue is that data-driven attribution is only as good as the data it trains on. If your tracking has gaps, and almost every business has tracking gaps, the model learns from an incomplete picture. It does not know about the podcast ad that introduced someone to your brand, the word-of-mouth recommendation from a colleague, or the trade press article that moved a prospect from consideration to intent. It only knows what it can see, and it builds a confident-looking model on that partial view.

Forrester has written about the persistent gap between what marketing measurement promises and what it delivers, and their analysis of measurement snake oil is worth reading for anyone who feels uncomfortable about how much confidence vendors project onto inherently noisy data. The discomfort is warranted.

There is also a structural conflict of interest to be aware of. When Google’s data-driven attribution model runs inside Google Ads, it is assigning credit to touchpoints using data that Google can see. It cannot see what happened on Meta, in email, or offline. The model is not wrong, exactly. It is incomplete. And incomplete models, presented with algorithmic authority, can be more dangerous than simple ones that everyone knows are imperfect.

How Sales Cycle Length Should Shape Your Model Choice

One of the clearest frameworks I have used when advising clients on attribution is to start with the sales cycle, not the channel mix.

For a business with a short, transactional sales cycle, last-click or time-decay models are more defensible. The experience is simple. Someone sees an ad, clicks, buys. There are not many touchpoints to apportion. The gap between the model and reality is small.

Early in my career, when I was working on a paid search campaign for a music festival at lastminute.com, the attribution question was almost irrelevant. Someone searched for festival tickets, clicked the ad, bought. The experience was a straight line. The revenue showed up the same day. In that context, last-click told you everything you needed to know, and the simplicity was a feature, not a limitation.

For businesses with longer sales cycles, that simplicity becomes a liability. A B2B software company with a 90-day evaluation process, multiple stakeholders, and a mix of content, events, email, and sales outreach cannot run its marketing on last-click attribution without systematically defunding the channels that create pipeline. The model and the reality are too far apart.

The honest answer for longer cycles is usually a combination of position-based or data-driven attribution, supplemented by incrementality testing and qualitative research. No single model will capture it cleanly. That is not a failure of the tools. It is an accurate reflection of how complex purchase decisions actually work.

What Incrementality Testing Adds That Attribution Cannot

Attribution models measure correlation. They tell you which touchpoints appeared in the journeys of people who converted. They do not tell you whether those touchpoints caused the conversion.

That distinction matters enormously when you are making budget decisions. A brand campaign that appears in the journeys of high-value converters might look valuable in an attribution report. But if those people would have converted anyway through other channels, the brand campaign is not generating incremental revenue. It is just present in journeys that were going to end in conversion regardless.

Incrementality testing, typically through geo-based holdout experiments or platform-level lift studies, measures what actually changes when you run or pause a channel. It is harder to set up, takes longer to produce results, and requires statistical rigour that most marketing teams are not set up for. But it is the closest thing to a causal answer that most businesses can access without a controlled experiment.

I have seen paid social campaigns that looked mediocre in attribution reports but showed strong incremental lift when tested properly. I have also seen the reverse: channels that claimed significant credit in the model but showed negligible lift when you held them out. Attribution and incrementality do not always agree. When they disagree, incrementality is usually telling you something important.

The Channel Coverage Problem Nobody Talks About Enough

Attribution modelling in most businesses only covers the channels that can be tracked digitally. That is a significant limitation that rarely appears in the attribution report itself.

Offline channels, word of mouth, PR, events, and organic social all influence purchase decisions. None of them appear in a standard attribution model. The model does not flag their absence. It simply distributes credit among the channels it can see, which creates a systematic bias toward trackable digital channels and away from everything else.

When I was growing an agency from around 20 people to over 100, a significant portion of new business came through referrals, industry reputation, and relationships built over years. None of that showed up in any attribution model. If someone had applied a digital attribution framework to our new business pipeline, they would have concluded that our website and a handful of email touches were responsible for most of our growth. That would have been wrong in a way that would have led to very poor investment decisions.

This is not a reason to dismiss attribution modelling. It is a reason to treat it as one input among several, and to be explicit about what the model cannot see. The Forrester perspective on marketing reporting makes a similar point: the value of measurement frameworks depends heavily on what they exclude as much as what they include.

How to Choose the Right Attribution Model for Your Business

There is no universally correct attribution model. There is a model that is most defensible given your business context, your data quality, and the decisions you are trying to make. Here is how I approach the choice.

Start with your sales cycle. Short and transactional: last-click or time-decay is probably adequate. Long and considered: position-based or data-driven, supplemented by incrementality testing.

Audit your tracking coverage before you trust the model. If significant portions of your customer experience are invisible to your analytics, a sophisticated attribution model will produce sophisticated-looking nonsense. Moz has a useful overview of preparing for GA4 and what tracking gaps to anticipate, which is a reasonable starting point for an honest audit.

Be explicit about what the model cannot see. Document the offline and untracked channels that influence your customers. When you present attribution data, make the coverage gaps visible rather than letting the model imply completeness it does not have.

Use the model to inform budget allocation, not to determine it. Attribution data should be one input into channel investment decisions, alongside incrementality testing, competitive context, channel saturation signals, and commercial judgement. A model that tells you paid search deserves 80% of your budget should prompt a question, not a budget reallocation.

Revisit the model choice periodically. Your channel mix changes. Your customer experience changes. A model that was appropriate two years ago may be encoding assumptions that no longer reflect how your customers behave. Attribution model selection is not a one-time decision.

GA4 and Attribution: What Has Changed and What It Means

GA4 shifted the default attribution model from last-click to data-driven attribution. For many teams, this happened without much fanfare, and the implications were not widely communicated. Conversion numbers changed. Channel comparisons shifted. Some teams noticed. Many did not.

The practical effect is that channels earlier in the funnel, particularly paid social, display, and organic search, tend to receive more credit under data-driven attribution than they did under last-click. Paid search, which dominated last-click reports, typically receives less. This is generally a more honest picture. But it also means that historical comparisons between pre-GA4 and post-GA4 data are not straightforward, because you are comparing different attribution methodologies as well as different time periods.

If your team has not explicitly revisited channel performance benchmarks since the GA4 migration, that is worth doing. You may find that channels you were underinvesting in look more valuable under the new model, and that channels you were scaling aggressively look less dominant. Neither conclusion should be acted on without additional validation, but both are worth investigating.

For a broader view of how GA4 fits into a coherent analytics approach, the Marketing Analytics and GA4 hub covers the practical and strategic dimensions in more detail.

The Honest Conversation Attribution Data Should Prompt

The most useful thing attribution modelling can do is start a conversation, not end one. When the data shows that a particular channel is underperforming relative to its budget allocation, the right response is not to immediately cut it. The right response is to ask why, to check whether the model is capturing the full contribution, and to consider what would happen to the rest of the funnel if that channel were reduced.

I have sat in enough budget review meetings to know that attribution data gets used as a weapon as often as it gets used as a lens. A channel owner who understands attribution can always find a model that flatters their numbers. A finance director who wants to cut marketing spend can always find a model that makes a channel look inefficient. The model is not neutral. It is chosen, and the choice has politics attached to it.

The antidote is to be explicit about the assumptions embedded in the model you are using, to present attribution data alongside other evidence, and to treat the output as a prompt for investigation rather than a verdict. That is a harder conversation to have than presenting a clean chart and drawing a conclusion. It is also a more honest one.

Semrush’s breakdown of content marketing metrics is a useful companion read for anyone trying to build a fuller picture of channel contribution beyond last-click conversion data. Content rarely wins on attribution. That does not mean it is not working.

About the Author

Keith Lacy is a marketing strategist and former agency CEO with 20+ years of experience across agency leadership, performance marketing, and commercial strategy. He writes The Marketing Juice to cut through the noise and share what works.

Frequently Asked Questions