The Four Attribution Models Founders Confuse

I spent twenty years in analytics and go-to-market before I ran a marketing function of my own. Analytics at Google across Russia and the CIS, then analytics at eBay across 208 countries, then go-to-market at eBay in the US. The job was never only marketing measurement; it was the numbers under whole businesses — product, commercial, operations, the funnel end to end. And in every one of those seats the same argument repeated, and it was almost never an argument about data. It was an argument about a word. Someone would say "attribution," someone else would hear a different thing under the same word, and the two of them would spend an hour disagreeing about a number when the real disagreement was about which question the number was answering.

"Attribution" is one word doing four jobs. Founders use it to mean credit, cause, allocation, and story, and they treat the four as if they were four competing estimates of a single truth, four readings of the same thermometer that ought to agree. They do not agree, because they are not measuring the same thing. Each is a clean answer to a different question. The damage is done not by any one of them but by crossing them: accepting the answer to one question as if it settled another. That error is common, it is expensive, and it has no name, so I am going to give it one. But first the four models, because you cannot see the error until you can see that the four are genuinely different.

The four are four questions, not four methods

The standard way attribution gets taught is as a menu of methods, and it is the wrong frame. You have seen the menu: last-click, first-click, linear, time-decay, position-based, "data-driven." Pick one. Marketing blogs rank them as if you were choosing a more or less accurate instrument, as if moving from last-click to data-driven were like trading a kitchen scale for a lab balance.

It is not, and the menu hides the thing that matters. Every model on that list answers the same question, the credit question, and none of them touches the other three. Choosing among them is choosing a bookkeeping convention, not upgrading your grip on the truth. The real divisions are not inside that menu. They are between four whole questions that all wear the word "attribution," and the four are these.

Model one: credit. "Which touchpoints get the conversion?"

This is the dashboard default, and it is the entire menu above: last-click, first-click, linear, position-based, time-decay, the algorithmic "data-driven" model your ad platform runs. They feel like different models. They are one kind of model with the dial set differently. Each takes the conversions you observed and the touchpoints you observed, and applies a rule for dividing the first among the second. Last-click hands all the credit to the final touch. First-click hands it to the opener. Linear spreads it evenly. "Data-driven" fits the split from your own historical paths. The dial moves; the question does not. Every one of them answers: of the touches we recorded, who gets the line in the ledger.

This is bookkeeping, and as bookkeeping it is exact. If you need to split a sales commission, settle an affiliate payout, or produce a defensible channel report for a board, a credit model is the right tool and it is doing its job perfectly. The trouble starts only when the ledger gets mistaken for a measurement of cause, which is a different model entirely.

Model two: cause. "What revenue would not exist without this spend?"

This is the question founders almost always mean when they say they want attribution, and it is the one the credit models cannot answer at any dial setting. Cause is counterfactual. It asks what would have happened in the world where you did not run the campaign, and that world is not in your dashboard, because it did not occur. No amount of dividing up the conversions you did get tells you which of them you would have gotten anyway.

Measuring cause requires building the missing world on purpose: a holdout. You withhold the spend from a comparable slice (a set of geographies, a randomized audience, a matched market) and compare. Geo holdouts, randomized lift studies, ghost ads, public-service-announcement placebo tests, switch-back experiments. These are the only methods that measure cause, because they are the only ones that manufacture the comparison the question demands. They are slower, costlier, and lower-resolution than a dashboard, and they are the only honest answer to "did this spend produce revenue that would not otherwise exist." A credit model that confidently assigns 34% to paid social has not answered this question. It has not approached it.

Model three: allocation. "Where does the next dollar earn the most?"

This one is forward-looking, and it is also distinct, because cause and allocation are not the same. Knowing that a channel is incremental does not tell you whether the next ten thousand dollars into it returns more than the next ten thousand into a different one. Channels saturate. The tenth dollar into a small, hot channel can beat the millionth into a large, tired one. Allocation is a question about marginal return, about the shape of the curve, not about credit for what already happened.

The tool built for this is marketing-mix modeling: a top-down regression of total revenue against total spend per channel over time, with controls for season, price, promotion, and baseline demand. It is aggregate, not person-level. It never names a customer or a touch. It will not tell you who converted, and it does not try. What it estimates is the response curve, so you can find the point on each channel where the next dollar stops pulling its weight. It is the right model for "what should the budget be," and it is the wrong model for "who gets credit for this specific sale," which it cannot answer and was never built to.

Model four: story. "How does the customer narrate their own path?"

The fourth model is the one quantitative people dismiss and then regret dismissing. It is self-reported attribution: the "How did you hear about us?" field, the post-purchase survey, the line in the sales call where the prospect says they had been reading your stuff for a year. It is qualitative, it is biased by memory and recency, and it is not numerically additive; you cannot sum it into a clean pie.

It also catches the one thing the other three structurally miss: demand that leaves no click. The podcast someone heard. The Slack group where your name came up. The conference hallway. The screenshot forwarded in a DM. None of that appears in a touch-based model, because there was no tracked touch; it is underweighted in a mix model, because it is hard to isolate as a spend line; and it is invisible to an incrementality test you did not think to run on a channel you did not know was working. Self-reported attribution is how you find out the channel exists at all. Treated as a story, a source of hypotheses, it is irreplaceable. Treated as a ledger, totaled and trusted to two decimal places, it lies.

The Attribution Mismatch

Here is the error, and it is worth naming precisely because it is invisible while you are committing it. An Attribution Mismatch is using a model built to answer one question to make a decision that requires a different one. The model runs, it returns a number, the number is quantitative and therefore feels like evidence, and nobody notices that it is evidence about a question nobody asked.

The defining case, the one I have watched cost real money in more than one company, is this. A founder wants to know where to put the budget, which is an allocation question, model three, or sometimes a keep-or-kill question, which is causation, model two. They open the dashboard. The dashboard shows a last-touch credit report, model one, because that is the model that is free, instant, and already running. They read the credit report and they reallocate the budget. They have answered an allocation question with a bookkeeping convention. That is the mismatch.

And it is not a harmless approximation, because the mismatch has a direction. Last-touch credit systematically over-rewards the channels that harvest demand and under-rewards the channels that create it. Branded search, retargeting, the final email: these sit at the bottom of the funnel and catch people who were already coming, so they collect the last click and look spectacular. Content, organic social, PR, the long top-of-funnel work that made the person want you in the first place: these rarely get the last click, so on a last-touch report they look like waste.

Trust that report to allocate, and the move is obvious and wrong. Cut the "underperforming" demand-creation channels, pour the money into the "high-ROI" harvest channels. For a quarter the harvest channels keep converting the demand still in the pipe, and the dashboard looks vindicated. Then the demand that those channels were harvesting stops arriving, because you defunded the thing that created it, and the harvest channels decline with a lag long enough that nobody connects the decline to the decision. The mismatch covered its own tracks. The credit model was not lying about credit; it was answering its own question correctly. It was being asked the wrong one.

This is the real content of the pillar I keep returning to: attribution is the only honest metric. The honesty does not live in the number. It lives in the match between the number and the decision. A precise answer to the wrong question is not partial honesty. It is a confident error wearing the costume of rigor, and it is more dangerous than no number at all, because no number invites caution and a clean dashboard invites action.

The one question

The fix is not a better model. There is no single model that answers all four questions; the search for one is itself a symptom of the confusion. The fix is a habit, and it is one question asked before you trust any attribution figure: which of the four is this, and is it the question my decision actually needs?

Run the decision, not the dashboard, and the mapping is short.

• Splitting a commission, settling an affiliate payout, reporting channel credit to a board? You want credit. A touch-based model is correct, and which dial you set it to is a policy choice, not a truth claim. Pick one, write it down, stop relitigating it.
• Deciding whether to keep or kill a channel? You want cause. Nothing but an incrementality test answers it. If you cannot run one, the honest move is to say "we will run this for ninety days and watch total revenue, and if there is no lift we turn it off," and to know that you are estimating, not measuring.
• Deciding where the next block of budget goes across many channels? You want allocation. That is a mix model, or at minimum marginal-return reasoning, and it is emphatically not a last-touch report.
• Trying to find the demand you cannot see, the dark-social and word-of-mouth the trackers miss? You want story. Ask people. Read the surveys for hypotheses, then go test the promising ones with model two.

No one of these is the truth alone. The honest stack uses several and triangulates, lets the credit ledger, the lift test, the mix model, and the customer's own account disagree, and treats the disagreement as information about which question each is answering rather than as noise to be averaged away. The dishonest move is the opposite: reach for the single model whose answer flatters the decision you have already made, and present it as data. The clearest tell that you are looking at an Attribution Mismatch is that the number agrees with what someone already wanted to do.

Where this reaches its limit

Naming four models cleanly invites a false comfort, that if you just pick the right one per decision you have solved measurement. You have not, and the boundary deserves to be stated as plainly as the framework.

Cause is genuinely hard, and often you will not get to measure it. Many channels are too small, too entangled, or too slow to run a clean holdout against, and for those you are reasoning under real uncertainty, not measuring. Mix models need years of varied spend to separate the channels, and most companies have neither the history nor the variation. Self-reported data degrades the moment you treat it as more than a lead. And every one of these models can be specified to flatter; the math does not protect you from a motivated analyst, it only changes which step the motivation hides in.

So the four models do not deliver certainty. What they deliver is the ability to know which kind of uncertainty you are standing in, which is the difference between an honest "we believe this channel is incremental but cannot yet prove it" and a dishonest "the data says paid social drives 34% of revenue." The first names its question and its limit. The second has crossed two models and hidden the seam. The whole discipline is keeping that seam visible.

The word "attribution" will go on doing four jobs; it is too entrenched to retire. The repair is not a new word for the word. It is the reflex, before you trust any figure that wears it, to ask which of the four you are holding, and whether it is the one your decision was asking for.

---

The fastest way to see which attribution questions your funnel is currently answering by accident is to have someone trace them. The 72-Hour Growth Diagnostic maps where your reported numbers and your real causal picture diverge, names the mismatches that are steering budget, and returns three prioritized fixes, produced in 72 hours. The output is the proof.

— Linara Bozieva, Founder, Ravenopus

---

In one paragraph, and a few common questions

The four attribution models are four different questions sharing one word. Credit (touch-based: last-click, first-click, linear, position-based, "data-driven") asks which observed touchpoints split the conversion; it is exact bookkeeping and says nothing about cause. Cause (incrementality: holdouts, lift studies) asks what revenue would not exist without the spend; it is the only model that measures causation. Allocation (marketing-mix modeling) asks where the next dollar earns most at the margin. Story (self-reported surveys) asks how the customer narrates their own path and catches the untracked demand the others miss. The costly error, the Attribution Mismatch, is using one model to make a decision that needs another, most often reading a last-touch credit report to allocate a budget. The repair is one question asked before trusting any attribution number: which of the four is this, and is it the question my decision actually needs.

Which attribution model is the most accurate? Malformed question. They answer different questions, so accuracy is only defined relative to one of them. Touch-based is exact as bookkeeping and near-useless as causation; incrementality measures cause but cannot split a multi-channel budget. Match the model to the decision instead of crowning one.

Why is last-click misleading? It is not misleading as bookkeeping. It becomes misleading the instant it is used to allocate budget, because it over-credits demand-harvesting channels and under-credits demand-creating ones, so it steers money toward the bottom of the funnel and quietly starves the top.

What is an Attribution Mismatch? Using a model built for one question to decide another. The defining case is reading a credit report and using it to make an allocation call. The number looks quantitative and is therefore trusted, but it was never an estimate of the thing being decided.

How should a founder actually measure marketing? Start from the decision, name which of the four questions it is, then pick the model built for that question: credit for bookkeeping, incrementality for keep-or-kill, mix modeling for allocation, self-reported for finding untracked demand. Triangulate across several, and trust least the single number that agrees with what you already wanted to do.