Marketing Attribution Models Explained (And Which to Use)

A marketing attribution model is a rule that decides how credit for a conversion is divided among the touchpoints that led to it. Someone sees a Meta ad, searches your brand on Google a week later, ignores you for a while, then buys after an email. Three channels touched that sale. The attribution model is what says whether Meta gets all of it, email gets all of it, or each gets a slice. Every model is a simplification, and choosing one is really a choice about which simplification you can live with.

What is a marketing attribution model?

An attribution model assigns credit for a conversion across the marketing touchpoints in a customer's path to purchase. It answers a single question: when a sale happens, who gets paid for it in the report? Models fall into two camps. Single-touch models hand the entire conversion to one touchpoint (first or last). Multi-touch models spread fractional credit across several. Neither camp proves causation. They apply a rule to whatever touchpoints your tracking managed to capture, which, as we will get to, is a shrinking share of reality.

What are the 5 attribution models?

The five standard models are first-touch, last-touch, linear, time-decay, and position-based (U-shaped). Data-driven attribution is the algorithmic sixth. Here is how each one behaves in practice.

First-touch attribution

All credit goes to the first recorded interaction. If a prospect discovered you through a TikTok ad and bought four weeks later, TikTok gets 100 percent. First-touch is a demand-creation lens: it tells you which channels are filling the top of the funnel. Its weakness is obvious. It ignores everything that actually closed the sale, so it flatters awareness channels and makes your retargeting and branded search look worthless.

Last-touch attribution

All credit goes to the final interaction before the conversion. Last touch attribution (often last-click) is still the default in most ad platforms and in a lot of ecommerce reporting, mostly because it is simple and requires no judgment calls. It measures demand capture. The problem is that it systematically overpays the bottom of the funnel. Branded search and email look like heroes because they were standing nearest the cash register, while the paid social ad that created the demand three weeks earlier gets nothing.

Linear attribution model

Credit is split evenly across every touchpoint. Four touchpoints, 25 percent each. The linear attribution model is the fairest-feeling option and the easiest to explain to a skeptical CFO. It is also naive by design: it assumes a throwaway display impression and the demo call that closed the deal contributed equally. They did not.

Time-decay attribution

Recent touchpoints get more credit, older ones get less, usually on an exponential decay with a half-life you configure. Time decay attribution suits shorter consideration cycles and any business where recency genuinely correlates with influence. It still quietly discounts the top of the funnel, so if your awareness spend has a long lag, this model will make it look weaker than it is.

Position-based (U-shaped) attribution

The common split is 40 percent to the first touch, 40 percent to the last touch, and the remaining 20 percent divided among everything in between. Position based attribution is a reasonable compromise for teams that care about both discovery and closing. The weakness is that the 40/20/40 weighting is arbitrary. Nobody derived it from your data. It is a convention that feels balanced.

Data-driven attribution

Instead of a fixed rule, an algorithm compares converting and non-converting paths and assigns credit based on which touchpoints actually shift conversion probability. Google's data driven attribution is the best-known implementation and is now the default in GA4 and Google Ads. It is the most sophisticated of the six. It is also the least transparent: you generally cannot see the weights, you cannot easily audit them, and the model only sees the paths inside that platform's own tracking. Better math on partial data is still partial.

Attribution models compared

Model	How it assigns credit	Best for	Main weakness
First-touch	100% to the first interaction	Judging top-of-funnel and demand creation	Ignores everything that closed the sale
Last-touch	100% to the final interaction	Short cycles, impulse ecommerce, quick reads	Overpays branded search, email, and retargeting
Linear	Equal share to every touchpoint	Long B2B journeys, stakeholder buy-in	Treats a banner impression like a sales call
Time-decay	More credit the closer to conversion	Short to medium consideration cycles	Undervalues awareness spend with a long lag
Position-based (U-shaped)	40% first, 40% last, 20% split across the middle	Teams that need to value both discovery and closing	The weights are a convention, not a finding
Data-driven	Algorithmic weights from converting vs non-converting paths	High-volume accounts with enough conversion data	Black box, and blind to touchpoints it cannot see

What is the difference between first-touch and last-touch attribution?

First-touch attribution gives 100 percent of the credit to the first interaction, so it measures demand creation. Last-touch gives 100 percent to the final interaction before purchase, so it measures demand capture. Both are single-touch models, and both throw away every touchpoint in between. Run the same month through each and the channel rankings often invert completely.

What is multi-touch attribution?

Multi-touch attribution is any model that splits conversion credit across more than one touchpoint. Linear, time-decay, position-based, and data-driven are all multi touch attribution models. The premise is sound: real customers rarely convert on one interaction, so crediting one is a distortion. The catch is that a multi-touch model is only as good as the journey data feeding it, and journey data has degraded badly.

Why attribution broke

Attribution modeling was designed for a web where you could follow a single user across sites and devices. That web is gone.

• iOS 14.5 and App Tracking Transparency. Apple made cross-app tracking opt-in. A large share of iOS users declined, and Meta in particular lost visibility into conversions it used to observe directly. Platforms filled the gap with modeled conversions, which are estimates, not receipts.
• Third-party cookie deprecation and shorter cookie lifetimes. Safari and Firefox have restricted cross-site tracking for years, and ITP caps the lifetime of many client-side cookies. Long consideration windows quietly fall off the end of the journey.
• Cross-device journeys. A phone discovery and a desktop purchase are two anonymous users unless the customer logs in. That path never gets stitched.
• Walled gardens. Meta, Google, and TikTok each report only what happened inside their own ecosystem, using their own attribution windows (a view-through window on one platform, a click window on another). None of them see each other.

The result is that every platform reports a plausible number, and every platform's number is different. This is the exact problem marketing attribution software exists to reconcile, and it is why multi-channel attribution has shifted from a tracking problem to a math problem.

Platform-reported ROAS double-counts: a worked example

Here is what walled gardens do to your reporting. One month, one DTC brand:

• Meta Ads spend: $20,000. Meta reports $80,000 in attributed revenue, so a 4.0x ROAS.
• Google Ads spend: $15,000. Google reports $60,000 in attributed revenue, so a 4.0x ROAS.
• Email and other marketing costs: $5,000.
• Actual revenue in Shopify and Stripe: $100,000.

Add the platform claims and you get $140,000 of revenue from a business that took in $100,000. Both platforms claimed the same $40,000 of overlapping sales, because the same customer clicked a Meta ad and later a Google brand ad, and each platform counted the conversion in full. Nobody lied. They just each answered a question about their own garden.

Now do the blended math. Total marketing cost is $40,000. True revenue is $100,000. Blended ROAS is $100,000 / $40,000 = 2.5x, not the 4.0x each platform advertised. If those sales came from 500 new customers, blended CAC is $40,000 / 500 = $80, regardless of who claims credit. The marketing efficiency ratio (MER) is the same ratio expressed against total revenue, and it is the number a CFO will actually accept, because it reconciles to the bank account.

Blended metrics do not tell you which channel worked. That is their limitation and you should say so out loud. What they do is put a ceiling on the story: if the sum of your channel-level ROAS implies more revenue than you actually booked, your channel numbers are inflated and you now know by how much. Start from the blended ROAS figure, then use attribution models underneath it for directional budget decisions. If you want to sanity-check the platforms against ground truth, it also helps to query your raw conversion data in plain English and count the orders yourself.

Media mix modeling and incrementality testing, in plain English

Media mix modeling (MMM) ignores individual users entirely. It uses regression on aggregate weekly spend and revenue, across channels and over a long history, to estimate how much each channel contributed. It is privacy-proof by construction, since it never touches user-level data. It also needs a couple of years of history, real variation in spend, and someone who knows what they are doing. MMM answers "how should we split the annual budget," not "should we pause this ad set today."

Incrementality testing is the only method that actually establishes causation. You turn a channel off in some geos and leave it on in others, or hold out a randomized audience, then measure the difference in revenue. It is the closest thing to truth in marketing measurement. It is also slow, costs you real revenue while the test runs, and answers one question at a time. Run a few geo holdouts a year rather than pretending a click-path model settles the question.

Which attribution model is the most accurate?

None of them, in the strict sense. Every model is a rule applied to partial data, and the rule was chosen before the data arrived. Data-driven attribution is the most defensible of the six, since its weights come from observed paths rather than a convention, but it still only sees what its own platform tracked.

The practical stance most good growth teams have landed on: pick a directional model and stay consistent, so week-over-week comparisons mean something. Use last-touch if your cycle is short, position-based or data-driven if it is not. Then hold every channel number against the blended reality check. When Meta says 4.0x and blended says 2.5x, blended is the one that matches your bank account.

How to run this without a spreadsheet ritual

The reconciliation is not intellectually hard. It is just tedious: pull spend from each ad platform, pull true revenue from Shopify and Stripe, align the windows, compute blended ROAS, blended CAC, and MER, then compare against what each platform claimed. Do it by hand and it is stale by Wednesday.

MixedMetrics connects read-only to GA4, Google Ads, Meta Ads, TikTok Ads, Search Console, Shopify, Stripe, Klaviyo, and HubSpot, and blends it into one live view: blended ROAS, blended CAC, MER, LTV, LTV:CAC, and revenue by channel, with an AI layer that flags what changed and where spend is leaking. We are honest about what we are: attribution-lite. We give you cross-channel blended truth and the gap between claimed and actual revenue, not a black-box multi-touch model promising a precision nobody in this industry can deliver. Start with the blended ROAS dashboard and work down from there.

Pick a model, understand exactly how it lies to you, and keep a blended number next to it. That is the whole discipline.