Attribution, honestly
Perfect attribution does not exist, and any tool that promises it is selling you a model rather than a measurement. What you can do is understand why the numbers disagree, choose the right number for the decision in front of you, and run a small number of experiments that tell you what actually changed. This guide is about making decisions with imperfect data instead of pretending the data is better than it is.
Why the numbers do not add up
Add up the conversions reported by every platform you advertise on and you will usually arrive at more sales than your business made. This is not fraud. It is the inevitable result of several systems, each with its own view of the world, each asked to report on itself.
Four mechanisms cause most of the double counting.
- Different windows. One platform counts a conversion seven days after a click. Another counts one day after a view. A purchase that followed both a click and a view inside both windows is claimed twice, legitimately, by two systems that do not talk to each other.
- Modelled conversions. Where tracking signals are missing — blocked cookies, restricted app tracking, cross-device journeys — platforms estimate the conversions they believe they caused. Those estimates are useful for optimisation and should not be treated as a measurement of revenue.
- Self-reporting. The platform selling you the inventory also grades the exam. There is no malice required for the grade to be generous.
- No shared identity. Nobody has a single view of the person who saw an ad on a phone, clicked on a laptop and bought in a shop. Each system sees a fragment and reports the fragment as if it were the whole.
The consequence is that platform-reported numbers are excellent for relative comparison within one platform, and poor as a statement of fact about your business.
Last-click flatters the end of the journey
Last-click attribution gives all the credit to the final interaction before the sale. It is the default almost everywhere because it is easy to compute and easy to explain.
It is also structurally biased. Consider a customer who read a guide in March, saw three social ads in April, and converted after searching your brand name in May. Last-click credits the brand search. The social campaign that did the persuading and the content that built the trust get nothing. Brand search looks like a marvel and paid social looks like a waste, so the budget moves to brand search, and the top of the funnel quietly starves.
That failure mode is extremely common, and it is self-reinforcing. Brand search and retargeting will always look good under last-click, because they reach people who were already close to buying. Top-of-funnel activity will always look bad, because its job is to create the demand that something else closes. Judge the whole system by last-click and you will systematically defund the part that creates growth.
The other models, and what each one hides
Every attribution model is a rule for distributing credit. None of them is correct, because credit is not a physical quantity that exists in the world. Each is a lens, and each lens has a blind spot.
| Model | What it credits | Blind spot |
|---|---|---|
| Last click | The final interaction before conversion | Everything that created the demand |
| First click | The first known touch | Everything that closed the sale |
| Linear | Every touch equally | The reality that touches differ in value |
| Position based | First and last touch most, the middle shared | Still an arbitrary split |
| Data driven | Whatever the model estimates from your data | It is still a model, and it is only as good as the data behind it |
There are two other approaches worth knowing about, because they escape the model problem rather than refining it. Marketing mix modelling works at the level of aggregate spend and outcomes over time, and it can see channels that leave no click trail at all, such as offline, brand and sponsorship. Its weakness is that it describes correlation at a high level and says little about the individual campaign. Incrementality testing measures the thing everyone actually wants to know, and we will come to it.
A practical position: use data-driven or position-based attribution for day-to-day channel comparison, use last-click only where the journey genuinely is one step, and never let any single model settle a large budget question on its own.
Platform conversions versus blended CAC
Platform-reported conversions are what each advertising system believes it produced. Blended customer acquisition cost is what you actually paid.
Blended CAC is total acquisition spend across all channels in a period, divided by the number of new customers acquired in that period. It is crude, and its crudeness is the point: it cannot be inflated by a generous attribution window and it includes the agency retainer, the content team and the tools alongside the media.
Its limitations are real. It gives you no channel-level insight, so it tells you whether the machine works rather than which part to change. It lags, because a customer acquired today may have been influenced weeks ago. And it is distorted by seasonality if you compare short windows. Use it as the constraint, not the guide: the blended number is the one that has to hold, and the channel numbers are the ones you steer with.
The essential pairing is blended CAC with customer value. A CAC figure means nothing on its own. What matters is how much a customer is worth, how long it takes to earn back what you paid to acquire them, and whether that payback fits your cash position. A business can be profitable with a high CAC and a long customer life; a business can be dying with a low CAC and no repeat purchase. Put payback period on the dashboard next to CAC, and you will make better decisions than any attribution model delivers.
Incrementality: the only question that really matters
Every attribution discussion eventually collapses into one question: would this sale have happened anyway? That is incrementality, and it is the only thing that tells you whether a channel is creating value or taking credit for demand that already existed.
You cannot answer it by looking harder at dashboards, because every dashboard is a record of correlation. You answer it by changing something and comparing.
- Geo holdouts. Run a channel in some regions and not others, matched as closely as you can on size and baseline demand, then compare. This is the most practical test for most businesses.
- Spend step tests. Increase or decrease spend deliberately in one channel for a defined period and watch whether total revenue moves by more or less than the spend change.
- Switch-off tests. Pause a channel entirely for a few weeks, in one market, and see what happens to total sales rather than to that channel's report.
- Holdout audiences. Where a platform supports it, withhold a random share of the audience from seeing the ads and compare conversion rates between the exposed and unexposed groups.
These tests cost money and time, they need enough volume to be readable, and they can be contaminated by seasonality, competitor activity and your own other campaigns. Run them with a hypothesis written down in advance, decide what result would change your mind before you start, and accept that a test which returns "no measurable difference" has still bought you something valuable: permission to stop spending there.
Most businesses run too few of these tests and too many dashboard reviews. One honest geo holdout per quarter will teach you more than a year of attribution tooling.
Making decisions when the data is imperfect
You will never have clean data. You still have to decide where the money goes. A few rules make that workable.
- Rank by marginal return, not average. The average return on a channel tells you what the money you already spent did. The question is what the next euro would do. A channel with a great average and a saturated audience is a bad place for more budget.
- Set a floor, not a target. Define the blended CAC above which you will not spend, and treat it as a constraint. Inside that constraint, let channel-level numbers guide allocation rather than dictate it.
- Use platform data for the small decisions. Which creative, which audience, which keyword, which landing page. Relative comparisons within one platform are exactly what those numbers are good at.
- Use blended numbers and tests for the big decisions. Whether to keep a channel, how much to spend overall, whether to hire, whether to scale. These are questions platform dashboards are structurally unable to answer.
- Reallocate on a cadence, not continuously. Moving budget every day means you never learn what any configuration does. Review monthly, change deliberately, and give the change enough time to be judged.
- Fix the plumbing before you buy a tool. Consistent campaign naming, UTMs that survive the journey, server-side event tracking, and revenue recorded in a CRM are worth more than any attribution product. Tools model whatever data you feed them, and most businesses feed them inconsistency.
The honest summary is that attribution is a decision-support system, not an accounting system. Its purpose is to help you make a better bet than you would have made by guessing, not to produce a number that is true to the euro.
Where most people get this wrong
The most damaging mistake is turning off the top of the funnel because of a last-click report. It is an easy decision to make, it improves the reported numbers immediately, and it makes next quarter worse in a way that is hard to trace back. The people who do it are not careless; they are reading a number that is genuinely misleading and treating it as fact.
The second mistake is letting each platform's dashboard be the report. If the social report, the search report and the email report each arrive separately and each claims its own success, nobody is looking at the business. Someone has to own the one number that joins spend to revenue, and it has to be a person, not a tool.
The third is optimising towards modelled conversions as though they were observed ones. Modelling is a reasonable way to steer an algorithm in the absence of signal. It is not a reasonable basis for telling a board what happened. Know which of your numbers are measured, which are estimated, and which are claimed, and never mix them in the same sentence.
The short version
- Platform conversions will sum to more than your sales. That is normal, and it means they cannot be your source of truth.
- Last-click systematically defunds the top of the funnel. Never use it to decide whether awareness activity stays.
- Track blended CAC and payback period alongside platform numbers. The blended figure is the one that has to hold.
- Run one incrementality test per quarter — a geo holdout is usually the most practical — and write the hypothesis down first.
- Steer small decisions with platform data and big decisions with blended numbers and tests. Fix naming, UTMs and CRM revenue before buying any attribution tool.
Want reporting you can actually decide from?
We will join your spend to your revenue, tell you which numbers are measured and which are estimates, and design the smallest experiment that would settle the question you are arguing about.