Marketing Attribution Models, Compared

Every growth team I have worked with hits the same argument eventually. Paid search says it drove the sale. Paid social says it started the whole thing. Email says the customer only converted because of the nurture sequence. Content marketing quietly points out that nobody would have heard of the product without the article that ranked first. They are all looking at the same conversion, and they are all claiming it. Attribution is the referee in that argument, and the uncomfortable truth is that the referee is guessing.

I run multi-channel growth where the output of this argument is not academic. It decides where next quarter’s budget goes. Get attribution wrong and you starve the channel that actually creates demand while overfeeding the one that happens to be standing closest to the checkout button. I have watched teams do exactly that, cut brand and content because “it does not convert,” and then wonder why paid search got more expensive and less effective six months later.

So this is a plain comparison of the standard marketing attribution models, what each one is good and bad at, and, more importantly, the mental model I use so that no single number runs the budget. None of these models is truth. Some are just less misleading than others for the question you are actually asking.

What attribution is really trying to answer

Strip away the tooling and attribution answers one question: which touchpoints deserve credit for a conversion, so you can decide where to spend more. That is it. A customer sees an ad, reads a review, clicks a search result, opens an email, and eventually buys. Attribution assigns credit across those touches. The credit becomes the numerator in your channel math, and that math sets budget.

The reason this matters so much is that attribution feeds directly into cost per acquisition. If you over-credit a channel, its apparent CAC looks great and you pour money in. If you under-credit a channel, it looks expensive and you cut it. Since CAC only means something next to lifetime value, the whole exercise is upstream of the LTV and CAC decisions that determine whether the business is actually growing profitably. Attribution is not a reporting nicety. It is the input that decides what you buy more of.

Why attribution is genuinely hard

If customers saw one ad and bought immediately, none of this would be difficult. They do not. Real paths have several touches across several channels over days or weeks, sometimes months for a considered purchase. That is the first problem: multi-touch paths mean the credit has to be split, and there is no physical law telling you how.

The second problem is cross-device. Someone discovers you on a phone during a commute, researches on a work laptop, and buys on a tablet at home. Three devices, one human, and your analytics often sees three strangers. The third is the consideration window. When the gap between first touch and purchase is long, the early touches fall outside most tracking windows entirely, so they simply never get counted.

Then there is the privacy shift, which has made all of this materially worse. Third-party cookies are disappearing, iOS restricts tracking, and consent banners mean a large share of users are never measured at all. The deterministic, follow-the-user-around tracking that attribution quietly depended on is eroding. So attribution is hard for structural reasons, and the ground under it keeps moving.

Last-click: simple, popular, and usually wrong

Last-click gives all the credit to the final touch before conversion. It is the default in most tools and the most widely used model in the world, and it is easy to see why. It is simple to explain, it needs no assumptions, and everyone understands it. The last thing the customer clicked gets the sale.

The problem is that it ignores everything that came before the final touch. It systematically over-credits bottom-funnel channels, branded search being the classic example. Someone already decided to buy, types your brand name, clicks the ad, converts, and paid search claims a customer that discovery, content, and word of mouth actually produced. Optimise hard on last-click and you will keep pouring money into the channel that closes demand while defunding the channels that create it. That is the single most common and most expensive mistake I see.

First-click: the mirror-image bias

First-click does the opposite. All credit goes to the first touch that introduced the customer. It answers a real and useful question, which is where do new customers first hear about us, and for that specific purpose it is fine.

As a budget model, though, it over-credits discovery and ignores everything that moves someone from aware to buying. A channel can be brilliant at generating first touches that never convert without a lot of expensive nurturing downstream. First-click will make that channel look like a hero. It is the exact inverse of last-click’s error, and pairing the two is a decent quick sanity check, but neither should run the budget alone.

Linear: fair, simple, and a little naive

Linear splits credit equally across every touch in the path. Five touches, twenty percent each. Its appeal is that it stops pretending any single touch did all the work, and it is easy to explain to a room.

The weakness is in the word equal. Touches are not equal. A throwaway impression and a deep comparison-page visit both get the same slice, which flatters low-value touches and dilutes the ones that genuinely moved the decision. Linear is naive in the specific sense that it refuses to make a judgement about which moments mattered. It is a reasonable step up from single-touch models and an honest default when you have no basis for weighting, but do not mistake fairness for accuracy.

Time-decay: recency as a proxy for influence

Time-decay gives more credit to touches closer to the conversion and less to earlier ones. The logic is that recent interactions were more influential in the final decision, which is often true, especially for shorter sales cycles.

It is a sensible middle ground and better than last-click because early touches still get something. The catch is that recency is only a proxy for influence, not the thing itself. For products with long consideration windows, the touch that actually mattered might have been an article read weeks before purchase, and time-decay will quietly discount it toward zero. You are still assuming a shape for the curve rather than measuring real contribution.

Position-based: crediting the bookends

Position-based, or U-shaped, weights the first and last touches most heavily, typically forty percent each, and splits the remaining twenty across the middle. The reasoning is intuitive. The touch that introduced the customer and the touch that closed the sale both did something special, and the middle touches were supporting cast.

For a lot of businesses this matches reality better than any of the single-touch or flat models, because it respects both discovery and conversion at once. It is my usual recommendation when a team wants one rule-based model to standardise on. The limitation is that the weights are still a guess dressed up as a decision. If your middle touches are where the real persuasion happens, U-shaped will under-credit exactly the work that matters.

Data-driven attribution: best in theory, black box in practice

Data-driven or algorithmic attribution tries to model actual contribution instead of applying a fixed rule. It looks at converting and non-converting paths and estimates how much each touch changed the probability of conversion. Done well, this is the most accurate touch-based approach available, because the weights come from your data rather than someone’s intuition.

The catches are real. It needs volume, a lot of it, to produce stable estimates, so smaller programs cannot use it reliably. It is a black box, which means when it tells you a channel deserves more credit you often cannot fully explain why, and that makes it hard to defend in a budget meeting. And it still only models what it can observe, so everything the privacy shift hides from tracking is still invisible to it. It is the best touch-attribution model and still a model, not the truth.

Why last-click persists despite all this

Given everything above, why is last-click still everywhere? Because it is legible. A CFO understands “this channel produced these sales” instantly. It requires no assumptions to argue about, it is consistent over time, and it is what the tools default to. There is also a defensible narrow case: for pure bottom-funnel, high-intent channels where you genuinely only want to know what closed already-warm demand, last-click answers that question honestly.

The failure is treating it as the whole picture. Last-click worship, optimising the entire budget to it, guarantees you underinvest in everything that creates demand upstream. Use it as one lens with a clear known bias, not as the scoreboard.

The harder truth all click models miss

Here is the thing that no touch-attribution model, not even the algorithmic one, handles well. All click attribution undercounts channels that create demand but do not earn the last click. Brand building, content that gets read but not clicked through, a recommendation from a friend, a podcast mention, seeing your name enough times that a brand search feels natural later. These create real demand and leave little or no clickable trail, so the models credit the channel that happened to catch the click at the end.

This is why teams that live entirely inside attribution dashboards drift toward cutting brand and content. Those channels look weak in every click model because their contribution is structurally invisible to click tracking. The demand they create shows up as someone else’s conversion. If your only measurement lens is touch attribution, you will keep making this mistake, and it compounds.

What actually measures causation

To get past correlation you need methods that measure incrementality, the lift a channel actually causes rather than the conversions that flow through it. This is where the real answers live.

Incrementality testing and holdouts are the gold standard. You withhold a channel or campaign from a randomised group and compare against those who got it. The difference is causal, full stop. Geo experiments do the same thing across markets, turning spend up or down in matched regions and reading the difference in outcomes. These are the only methods that tell you what would have happened without the spend, which is the actual question behind every budget decision. They are harder to run and you cannot test everything constantly, but for big bets they are worth the effort many times over.

Media mix modeling, MMM, is the other approach worth knowing, and it is having a genuine comeback. MMM uses aggregate spend and outcome data with statistical modeling to estimate each channel’s contribution, and crucially it needs no user-level tracking at all. That property is exactly why it is returning to fashion as privacy erodes deterministic measurement. It is top-down where attribution is bottom-up, better at capturing hard-to-track channels like brand, and weaker at fine-grained tactical decisions. The strongest measurement setups I have seen triangulate: attribution for day-to-day direction, incrementality to validate the big claims, MMM for the whole-portfolio view.

How this connects to budget and CAC

Everything here converges on money. The credit a model assigns becomes each channel’s apparent CAC, and apparent CAC decides where the next euro goes. This is why the choice of model is a budget decision disguised as an analytics one. Optimise blindly to last-click and you will systematically overspend on channels that close demand and underspend on the ones that create it, and your blended acquisition cost will drift up while every individual channel dashboard still looks fine.

When I ran an affiliate program turnaround, attribution sat right at the centre of it, because affiliates are one of the channels most exposed to last-click over-crediting. Coupon and loyalty affiliates often catch the final click on demand that other channels created, and a naive model rewards them for conversions they did not cause. Sorting out what was truly incremental versus what was intercepting existing demand was most of the work. The same trap hides inside marketing automation, where the last email before purchase looks decisive; if you are building that infrastructure, it is worth designing measurement in from the start, which I get into in marketing automation architecture.

A practical stance I would actually recommend

After all the caveats, you still have to run a business, so here is the working approach. Pick one primary attribution model and standardise on it, and know its bias cold. I usually reach for position-based or data-driven if the volume supports it, because both respect more than the final touch. Whatever you pick, write down what it over-credits and what it under-credits, and keep that note next to the dashboard.

Then validate the big bets with incrementality. You do not need to test everything, but before a major budget shift, run a holdout or a geo test and check whether the causal lift agrees with what attribution is telling you. When they disagree, trust the experiment. Treat MMM as the portfolio sanity check that keeps hard-to-track channels from disappearing off the map. And build reporting that makes the model’s assumptions visible rather than hiding them behind a single confident number, which is a theme I keep returning to in dashboards that drive action.

Above all, do not treat any model as truth. Attribution is a lens with a known distortion. Use it for direction, cross-check it with a second lens, and reserve causal methods for the decisions that are expensive to get wrong. Getting your measurement plumbing right in the first place, especially in a tool like GA4, makes all of this far less painful, which is worth setting up carefully as I cover in GA4 for growth.

Common mistakes to avoid

A short list of the errors I see most often, gathered in one place. First, last-click worship, running the entire budget to the model that most over-credits bottom-funnel channels. Second, trusting a single model, treating one lens as the answer instead of triangulating. Third, ignoring incrementality, never actually testing whether the spend causes the outcome or just correlates with it. Fourth, over-crediting trackable channels simply because they are easy to measure, which structurally penalises brand, content, and word of mouth. And fifth, forgetting that the privacy shift is degrading your tracking every quarter, so a model that looked accurate two years ago is quietly getting blinder.

The short version

• Attribution answers one question: which touchpoints deserve credit for a conversion, so you can allocate budget. It feeds straight into CAC.
• It is hard because paths are multi-touch, users cross devices, consideration windows are long, and privacy changes are eroding deterministic tracking.
• Last-click over-credits the final touch, first-click over-credits discovery, linear treats all touches as equal, time-decay favours recency, position-based weights the bookends, and data-driven models real contribution but needs volume and is a black box.
• Last-click persists because it is legible and assumption-free, and it is defensible only for narrow bottom-funnel questions.
• All click attribution undercounts channels that create demand without earning the last click, which is why teams wrongly cut brand and content.
• Incrementality testing, holdouts, and geo experiments are the gold standard for causal truth; MMM is returning because it needs no user-level tracking.
• Pick a primary model, know its bias, validate big bets with incrementality, and never treat any model as truth.

---

I am Deepanshu Grover, a Growth Product Manager in Paris. If your channels are all claiming credit for the same conversions, connect on LinkedIn or get in touch.