A customer sees a connected TV ad on Sunday. Scrolls past a paid social post on Tuesday without stopping. Gets an email Thursday. Searches the brand name on Saturday, clicks a paid search ad, and buys.
Which ad made the sale?
Every reporting system in your stack will answer that question confidently, and they will not agree with each other.
Why does every model give a different answer?
Because an attribution model is a rule for splitting credit, not a discovery of what happened. Apply different rules to that same journey and the story changes completely.
| Model | Who gets credit | What it distorts |
|---|---|---|
| Last click | Paid search | Overvalues the closer, erases discovery |
| First click | Connected TV | Overvalues discovery, erases the close |
| Linear | All four equally | Treats a scrolled-past impression as equal to a click |
| Time decay | Weighted to search and email | Assumes recency equals influence |
| Algorithmic | Distributed by model | Only sees what it can track |
Nothing about the purchase changed. The reported winner changed four times.
This is why we resist the question “which model should we use” and replace it with “what decision are you making.” A model that answers a budget question well can answer a creative question badly.
What do all of them miss?
The touchpoints outside the tracking. Every model above is working from the events it can observe, and a real buying journey is full of events it cannot.
A recommendation from a colleague. A podcast mention. A billboard on a commute. A connected TV impression in a household where somebody else did the buying. A conversation that happened offline and left no record anywhere.
None of these appear. The models divide credit among the channels they can see, which means observable channels absorb the credit for work done by unobservable ones. Digital channels look better than they are, in part because they are the ones being measured.
Why does brand search break every model?
It is the clearest example of the problem. Somebody has already decided to buy. They type the brand name. A brand search ad sits above the organic listing. They click it and convert.
Last click gives that sale entirely to paid search. The truthful answer is that the decision was made earlier by something else, and paid search collected it at the door.
This is not a small distortion. In our own search data, 37 percent of clicks came from people typing our name, converting at a rate nothing else on the site approaches. Any account that blends brand and non-brand together will look healthier than the underlying business is.
Splitting them is the single highest-value change most reporting packs can make.
Do the platforms agree with each other?
No, and they cannot, because each one only sees its own surface and each one is motivated to claim what it touched. Sum the revenue reported by every platform and it usually exceeds what the business actually took. That gap is not fraud. It is the same order counted more than once.
The check is straightforward. Compare summed platform revenue against finance system revenue. Whatever the difference is, that is how much of your reporting is describing the same customers twice.
So what do we actually trust?
Three things, in this order.
Total revenue against total marketing spend. Blunt, unglamorous, and impossible to double count. When platform efficiency improves while this line stays flat, the channels are competing over credit.
Holdout tests. Turn a channel off for a comparable geography or audience, hold everything else steady, and measure the total result. This is incrementality work and it is the only method that answers what would have happened anyway. Run it on your best-performing line items first, because that is where the overclaiming is largest.
Model comparison rather than model selection. Read the same period through two models and look at where they disagree most. The disagreement is more informative than either answer, because it shows you which channels are dependent on a crediting rule to look good.
What we tell clients
Attribution is a set of useful approximations and we treat it that way. We use models to direct attention and tests to make decisions.
Anybody presenting a single attribution number as the truth is either selling something or has not looked closely at how the number was produced. The honest version is less satisfying and considerably more useful, which is roughly the trade we would make on any measurement question.
Frequently Asked Questions
Which attribution model is the most accurate?
None of them are accurate in the sense of being true. Each one is a rule for dividing credit, and the rule you pick decides which channel looks good. The useful question is which distortion you can live with, not which model is correct.
Why does last-click attribution favour search?
Because search is usually the final step before purchase. Somebody discovers a brand elsewhere, then searches its name to buy. Last-click hands the entire result to the channel that happened to be standing closest to the checkout.
Should I use data-driven attribution?
It handles multi-touch journeys better than fixed rules, but it still only sees the touchpoints inside its own tracking. Offline conversations, connected TV, word of mouth and organic discovery stay invisible to it.
What actually proves a channel worked?
Holdout testing. Turn the channel off for a comparable audience or geography, keep everything else constant, and measure total business outcomes rather than platform-reported ones. It is the only method that answers what would have happened otherwise.