Marketing attribution models: what each one is for and how to choose

Pull up the same quarter in two tools and you get two winners: GA4 says email drove the conversions, Google Ads says branded search did. Neither report is broken. Each runs a different attribution model, and an attribution model is a rule for splitting credit, not a measurement of what caused the sale. There is no single right model. The teams that get value out of attribution stop hunting for one and match the model to the decision in front of them.

What follows earns that claim. First the short version, for anyone who just needs a defensible default. Then the proof: one real deal, run through all seven models, no two of them telling the same story. From there the piece builds in one line, from what a model is, to which one your motion needs, to when to stop trusting all of them.

#The short answer

If you want a defensible setup without reading further, do this, in this order:

1. Take last-touch as your baseline (last non-direct if you run a store). It's free, it's everywhere, and its bias is the one everyone already understands: it flatters whatever sits closest to the sale.
2. Fix who gets credited before touching any weights. Sales-led B2B: credit the account, anchored at pipeline created. A store: the customer and the order. Self-serve: the user and the signup. Set the window to your real cycle, not the tool's default.
3. Add one positional view for balance. U-shaped for stores and simple funnels; W-shaped for sales-led teams with real lead and opportunity milestones in the CRM.
4. Add data-driven attribution only where volume is real. A store or self-serve funnel earns it; a pipeline that closes forty deals a year does not.
5. Run one setup per motion, and test before big moves. Reconcile motions at revenue, and calibrate the model you trust most with one holdout before it moves serious budget.

The rest of this page is why that order works, what each model actually answers, and where each one lies to you. If you only take one thing, take the list.

#An attribution model is a credit rule, not a measurement of truth

A marketing attribution model is the rule that decides how credit for a conversion is divided across the marketing touches that preceded it. First-touch and last-touch give one touch everything. Linear, time-decay, U-shaped, and W-shaped split credit by position and timing. Data-driven models compute the split from historical paths. Under every model, each conversion distributes exactly 100% of its credit; the model only decides where it lands.

That "only" is the point. The model doesn't discover which channel caused the sale. It allocates the conversion according to a rule you picked, which means every attribution report is an answer to a question you chose, whether you knew you were choosing or not. First-touch answers "what starts journeys." Last-touch answers "what's nearest the close." Neither answers "what worked," because a credit rule can't see cause. That's not a flaw to engineer away; it's what a model is. The practical consequence: arguing about which report is "right" is wasted breath, and picking the rule to match the decision is the whole game.

#One journey, seven models, seven different answers

Here's a deal the way B2B deals actually happen. A marketing manager clicks a LinkedIn ad in week 0 and attends a webinar in week 2; the CRM creates a lead. In week 6 the head of sales, sent by a colleague, types the pricing URL straight into a browser. In week 9 a retargeting ad brings the marketing manager back and a demo gets booked; an opportunity opens. In week 12 the CFO visits the site directly to sign off, and the deal closes in week 13.

Five touches, three people, one account. Run the standard splits over it and every model files a different report:

Splits computed with the standard weightings and rounded; the spread is the point, not the decimals.

Same five touches, and the "best channel" is the LinkedIn ad, or a browser bookmark, or the retargeting campaign, depending on nothing but the rule.

Two rows deserve a second look. Last-touch crowned a typed-in visit, and it usually will in B2B, because the last thing a buyer does before signing is go straight to your site; a report that keeps crowning "Direct" is telling you about your model, not your marketing. And time-decay with its common 7-day half-life gave the first ten weeks of the journey about 1% of the credit, which is what happens when a rule tuned for ecommerce cycles meets a 90-day pipeline. Both are working exactly as designed. The design just wasn't for your question.

One more thing this table quietly assumed: it stitched three people into one journey. Hold that thought; it matters more than the columns.

#The seven rule-based models and the decision each one serves

Every model in the table is a legitimate tool for a specific decision, and a liability outside it:

Definitions of the W's anchors vary by tool. The version that earns its keep in B2B pins all three to your pipeline: the first touch, the touch nearest lead creation, and the touch nearest opportunity creation, with the remaining 10% spread across the rest of the path. That detail is why W-shaped is the day-to-day default for sales-led teams: it's the only rule in the table whose emphasis follows your sales process instead of the calendar.

There's an eighth name you'll meet: data-driven attribution. It isn't a fixed rule but an algorithm that learns the weights from your own converting and non-converting paths. It now dominates the one place most budgets flow through, so that is where we look at it next.

#Google cut its lineup to two models, and that changed the question

Since September 2023, Google Ads and GA4 support exactly two attribution models: last click and data-driven. First click, linear, time decay, and position-based were removed across both products. Google's stated reason: fewer than 3% of conversion actions still used them. Data-driven is the default for most conversion actions.

So what is data-driven attribution actually doing? Instead of following a fixed rule, it weighs your converting paths against your non-converting ones and credits each touch for how much it moved the odds of a sale. Two engines are common. Shapley values borrow from game theory, scoring each touch by its average contribution across every combination of touches it appears in. Markov chains model the journey as a set of transitions and measure how far conversion probability falls when a channel is removed. How each is computed, and how GA4's version differs, is a guide of its own; what matters for choosing a model is what that approach buys you and what it costs.

So inside the biggest ad platform, the model question has already been answered for you unless you actively answer it yourself. Two things are worth knowing about that default.

First, the honest case for it. On a self-serve funnel with hundreds of conversions a month, short cycles, and one person per purchase, data-driven attribution is a reasonable default and probably beats whichever rule someone picked years ago. That's the shape of data the algorithm was built for.

Second, the costs, and they're operational, not philosophical:

• You can't explain it. When data-driven credit moves a campaign's conversions from 31 to 24, there is no rule you can state in the Monday meeting for why. The weights are the model's, not yours.
• It restates history. Algorithmic credit gets recomputed as new path data arrives, so last week's numbers don't necessarily mean this week what they meant then. Week-over-week comparison, the one thing a weekly report exists for, gets soft.
• It steers the money. The attribution model isn't just reporting; it's Smart Bidding's training data. Bids follow credited conversions, so a model that under-credits discovery keywords teaches the bidder to stop buying them, and the resulting drop in credited conversions reads like confirmation. The same feedback loop runs inside every platform's delivery system. The model choice is a spend decision, not just a reporting choice.

None of that makes data-driven wrong. It makes it a specific tool: strong where volume is high and paths are short, expensive where a human has to narrate the numbers. Which is exactly the split that matters for the rest of this page.

Everything Google did here, every platform does in its own way. Meta grades itself with its own attribution settings and windows, and each ad platform's model sees only the touches that platform can claim. Add up the dashboards of three channels and you will count more conversions than you had customers, because each one ran a different rule over a different slice of the journey. So the model conversation worth having lives one level up, in the view that sees all your channels side by side: your analytics, your CRM or store data, or a reporting layer built on them. A credit rule can only arbitrate between channels it can actually see.

#In B2B, who gets credited matters more than how credit is split

Go back to the journey. Three people touched it: the marketing manager who clicked and attended, the head of sales who checked pricing, the CFO who signed off. A person-level model sees three unrelated fragments and attaches the deal to whichever fragment contains the conversion event. The webinar that created the lead isn't even in that story.

Assembling the real journey required crediting the account, not a person. That single choice, before any weighting scheme, decides whether your attribution describes the deal or a third of it.

Three levers sit upstream of every model, and they move B2B numbers more than the model does:

• Who gets credited. B2B deals are bought by groups. Credit the account so every stakeholder's touches land on one journey, or accept that your model is scoring a group purchase by one member's browser history.
• Where credit anchors. Anchor at pipeline created, not closed-won. Anchoring at Won starves the top of the funnel (by the time a deal closes, early touches have aged out of windows and memory) and shrinks your sample to this quarter's handful of wins. Pipeline-created gives you more events, sooner, and it measures marketing's actual job: creating qualified opportunities.
• How long the window runs. Set it to your real sales cycle. The time-decay row above showed what a 7-day assumption does to a 90-day deal: the first ten weeks rounded to 1%.

W-shaped is what these levers look like as a model: its anchors are your pipeline milestones, pulled from your CRM's stage history. But notice that the levers apply to every model. A last-touch report at the account level anchored at pipeline created is a different, and dramatically more useful, report than the same "model" at the person level anchored at Won.

Those are the B2B answers to the three questions. A store answers the same three questions differently, and seeing both sets side by side is what makes the choice for your own motion straightforward.

#The same three levers, answered for B2C and ecommerce

The levers don't change for a store; the answers do. Who gets credited: usually one buyer, so the work isn't assembling a buying group, it's keeping one customer stitched together across devices, browsers, and the app. Where credit anchors: the purchase, though you still choose whether a first order or a repeat customer is the outcome that counts, and that choice reshapes which campaigns look good. The window: days rather than quarters, and in practice the ad platforms' own attribution windows decide a large share of what gets claimed as a conversion before your model ever sees it.

Position-based (U-shaped) is the standard multi-touch view for ecommerce because the pipeline milestones W-shaped anchors on don't exist; W-shaped implementations fall back to U for exactly this case. Data-driven attribution genuinely earns its keep here too: a store usually has the conversion volume the algorithm needs, which is the same reason it struggles on a forty-deals-a-year pipeline. And last non-direct matters more, not less, because returning customers typing the URL will otherwise absorb the acquisition story.

#"Which model should we use?" is the wrong first question

Whichever motion you run, three questions come first and do the actual work: what am I optimizing (ad spend, content, pipeline coverage), at what level of detail (a user clicking through a short funnel, or an account moving through stages), and over what window (days, or your real cycle)? Answer those and the model mostly falls out.

For a team standing up attribution from scratch, the sequence that works:

1. Start with last-touch as your baseline, on purpose. Every tool computes it, it's free, and its bias is the one you already understand: it flatters the close. A baseline whose lie you know beats a sophisticated model whose lie you don't.
2. Fix the level of detail before touching weights. Get journeys credited to the account and anchored at pipeline created. This is unglamorous plumbing (identity, CRM stage mapping) and it changes your numbers more than any model swap will.
3. Add a positional view once stage data is real. U-shaped for simpler motions, W-shaped when lead and opportunity milestones actually exist in your CRM. Compare it against your last-touch baseline; the gap between them is a map of what the baseline was hiding.
4. Add data-driven attribution only where one segment earns it. A self-serve funnel with real conversion volume, fine. A sales-led motion closing a dozen deals a quarter is not a training set; an algorithm fit to it is fitting noise, whatever the platform's requirements page says.
5. Before a model moves serious budget, test once. Hold out a region or audience, or pause the channel the model loves, and see what actually changes. One decent holdout will calibrate your trust better than a year of dashboard arguments.

And the multi-motion answer, since it's the question that starts most of these projects: run one setup per motion. A store gets order-level credit, platform-aware windows, and U-shaped or data-driven attribution, with last non-direct keeping the store-side view clean. Self-serve gets user-level credit, a short window, and data-driven attribution or a properly tracked last-touch. Sales-led gets account-level credit, a cycle-length window, and W-shaped or custom weights anchored at pipeline. Reconcile them where the money meets: revenue.

Reconciling matters because the motions leak. Running a hybrid motion ourselves, we see paid volume split three ways: self-serve conversions that never touch pipeline yet pay back a real share of the spend as revenue, self-serve users who raise a hand for sales months later, and direct pipeline. Judge those campaigns on pipeline alone and you undervalue the ones feeding both motions; keep the payback and the pipeline against the same spend, and keep the account window long enough that the original ad stays credited when a signup becomes an opportunity.

One company-wide model averaged across different motions is wrong for each of them, in different directions. The full walkthrough of this decision, including the on-ramp order for teams new to attribution, is its own guide.

#The model has to survive Monday morning

There's a second test for an attribution model, and it filters harder than any accuracy argument: can this model run your weekly report? It has two parts.

Legible: someone on your team can explain any number the model produces, in one sentence, without appealing to an algorithm. "LinkedIn gets 30% because it created the lead" is legible. "The model reweighted it" is not.

Stable: last week's numbers still mean the same thing this week. Rule-based models recompute only when you change the rule. Algorithmic models restate history on their own schedule, which quietly breaks the week-over-week comparison your budget meeting runs on.

A model can be theoretically superior and fail both tests, and a model that fails both can't do attribution's actual day job: telling you, every Monday, where to move budget this week, in numbers whose story you can tell. Prefer the simplest model whose bias you understand over the sophisticated one you can't; save the heavy machinery for the periodic questions it's built for.

This bar only rises from here, because the second reader of your weekly report is increasingly an AI agent, and an agent can only explain, or act on, credit logic that has stateable rules. An illegible model can't brief anyone, human or machine.

This is the reporting job Buron is built around. Its datasets (the modeled tables people and agents query) run one credit rule across every channel you advertise on, credit the buying group at the account level, anchor at whichever CRM stage you pick (most B2B teams pick pipeline created), and read the same way every week. Teams switch on only the models they use, with first-touch, last-touch, W-shaped, and U-shaped on by default, so the field picker and the weekly report keep one legible story instead of eighty metric columns. No model in that lineup claims to be the true one. That claim isn't available, as the next section explains.

#No model is ground truth, so triangulate

Every model in this article, including the algorithmic ones, reads correlation inside the paths you managed to track. None of them can tell you what would have happened without the spend, which is the question a budget decision actually asks. Effectiveness researcher Les Binet put the gap bluntly in 2023: "Attribution – quickly and cheaply – will give you an answer that is precise and wrong. Econometrics – slowly, laboriously and expensively – will give you an answer that is right."

The gap isn't hypothetical. In verified cases from Tom Roach's effectiveness work (2022), activity reporting a ROAS around £4 measured only 1 to 2% incremental sales when tested properly: the ads were collecting credit for sales that were happening anyway. And the distortion compounds when the model steers spend: Ron Berman's peer-reviewed analysis in Marketing Science (2018) found that last-touch attribution doesn't just misreport, it over-incentivizes ad exposure and lowers advertiser profit when used to allocate budget, with Shapley-value approaches (the principle behind today's data-driven models) the more sound benchmark.

So the econometrics camp is right about truth. What that camp's quarterly models can't do is run your Monday: an incrementality test takes weeks and a media mix model (MMM) recalibrates a few times a year, and neither will tell you which campaign to trim tomorrow.

The working arrangement is triangulation. Attribution, on a legible model at the right level of detail, runs the weekly optimization loop. Incrementality tests and MMM audit the big directional calls (kill the channel, double the brand budget) a few times a year. Where the two disagree, the experiment wins, and the disagreement itself tells you which bias your model carries.

One reading habit ties it together: no tool computes truth, and any pitch that a particular model is "the accurate one" deserves the question of what its maker sells. Attribution advice tracks product lines with remarkable consistency. This page is no exception, which is why its recommendation is a procedure and a pair of tests, not a model.

#Where to go deeper

This page is the map; the territory gets its own guides. As this library grows, each of these becomes a full article:

The models. Multi-touch attribution explained · data-driven attribution (how Shapley and Markov approaches actually work, and GA4's version) · attribution models in Google Ads and GA4 · last-click attribution and why it misleads · first-touch attribution · cross-channel attribution · view-through attribution · Meta's attribution settings.

The B2B layer. Revenue attribution for B2B · B2B marketing attribution at the account level · why who-gets-credited beats how-credit-is-split.

The decisions. Which attribution model should you use (the self-serve vs sales-led decision guide) · attribution vs incrementality · why "direct" and "organic" aren't channels · how to read attribution advice by what its author sells.

#The model is a choice. Make it on purpose.

Attribution models disagree because they're supposed to: each one is an instrument with a known bias, built for a different question. The move is to stop asking which instrument is correct and start asking three questions the instruments can't answer for you: what you're optimizing, who should get credited, and over what window. Then pick the simplest rule whose bias you understand, run it at the level your motion demands (the account and the pipeline in B2B, the customer and the order in ecommerce), and audit it with a real test before the big calls. That's not a compromise position; it's what the evidence-heavy end of this field has converged on, from Berman's game theory (2018) to Binet's econometrics (2023) to Google's own consolidation of its model lineup (2023).

The same lens reads every measurement debate you'll meet after this one: every method is an instrument, every instrument has a bias, and owning a bias you understand beats renting a precision you can't explain. The hard part isn't the theory; it's keeping a credit table that stays consistent, at the right level of detail, week after week, whether three stakeholders touch one deal or a thousand orders arrive through five channels. That table is the thing worth handing off, and building one that behaves is what Buron is for.

#Frequently asked questions

#What is a marketing attribution model?

A marketing attribution model is the rule that decides how credit for a conversion is divided across the marketing touches that preceded it. Single-touch models (first-touch, last-touch) give one touch everything; multi-touch models (linear, time-decay, U-shaped, W-shaped) split credit by position and timing; data-driven models compute the split from historical paths.

#What is the best attribution model for B2B?

There is no single best model. Match the model to your motion: sales-led teams get the most from W-shaped or custom weights, credited at the account level and anchored at pipeline created; self-serve funnels with real volume can trust data-driven attribution or a properly tracked last-touch baseline. Teams running both motions should run both setups and reconcile at revenue.

#Which attribution model is best for ecommerce and B2C?

For most stores, position-based (U-shaped) or data-driven attribution fits best: journeys are short, conversion volume is high, and the pipeline milestones W-shaped needs don't exist. Keep last non-direct in the store-side view so returning customers typing the URL don't absorb the credit, and read each platform's claimed conversions against its own attribution window.

#How do first-touch, last-touch, U-shaped, and W-shaped attribution differ?

First-touch gives 100% of conversion credit to the first recorded touch; last-touch gives it to the final touch. U-shaped splits 40/40/20 across first touch, last touch, and the middle. W-shaped ties credit to pipeline: 30% each to the first touch, the lead milestone, and the opportunity milestone, with 10% across the rest.

---