A retailer reads its dashboard and sees that branded search, people typing the company's own name into Google, drives a third of its online revenue at a glorious return on ad spend. So it pours more money in. Sales barely move. The dashboard wasn't lying; it was answering the wrong question. Those customers already wanted the brand. The ad just stood in the doorway taking credit for people who were walking in anyway. That gap, between credit and cause, is the whole subject of marketing analytics and attribution.

The quick version

  • Attribution assigns credit for a conversion across the touchpoints a customer saw, first click, last click, or a data-driven split. It describes the path; it does not prove the path caused the sale.
  • Incrementality answers the real question: would this conversion have happened without the ad? You can only measure it with a controlled experiment, a holdout group that doesn't see the ad.
  • Marketing-mix modelling (MMM) works at the aggregate level, using statistics to estimate how each channel moved total sales, no individual tracking required, which is why privacy changes have brought it back.
  • Use all three as a triangle. The trap is trusting one tidy attribution number and mistaking a description of the journey for proof of what worked.

The idea in depth: credit is not cause

The oldest complaint in marketing is a line usually attributed to the department-store magnate John Wanamaker: "Half the money I spend on advertising is wasted; the trouble is I don't know which half." Fittingly for a piece about attribution, the quote itself is mis-attributed, the Quote Investigator can't find Wanamaker ever saying it, and pins the same sentiment on the British industrialist Lord Leverhulme too. The point survives the uncertainty: knowing which half works is the job, and for a century it was guesswork.

Digital advertising promised to end the guessing and gave us attribution models instead. The default for years was last-click: whichever touchpoint immediately preceded the sale takes all the credit. It's simple and badly wrong, because it flatters whatever sits closest to the purchase, retargeting, branded search, discount-code sites, and starves the awareness work that created the demand. First-click has the opposite bias. Multi-touch models (linear, time-decay, position-based) share credit more evenly, and data-driven attribution (DDA) uses machine learning to weight each touchpoint by how much it shifts the modelled probability of conversion.

Here the ground has moved under everyone's feet. In 2023 Google retired four attribution models, first-click, linear, time-decay and position-based, from Ads and Analytics, leaving only last-click and data-driven, on the grounds that the deprecated models accounted for under 3% of conversions and that fragmented, privacy-restricted journeys had made them unreliable. So here's the practical habit worth building: stop treating "our attribution model" as a settled fact in your reporting. Know which model your dashboard uses by default (increasingly DDA or last-click), and ask what it systematically over-credits. If a channel only ever shows up at the end of journeys, suspect it's harvesting demand, not creating it.

The deeper problem is that every attribution model describes correlation, not causation. Avinash Kaushik, formerly Google's digital-marketing evangelist, puts the distinction sharply on his Occam's Razor blog: "Attribution is simply the science… of distributing credit for Conversions." Distributing credit is not the same as proving impact. To know whether an ad caused a sale, you need the counterfactual, what would have happened without it, and no attribution model can see a world that didn't occur.

"Attribution is simply the science… of distributing credit for Conversions.", Avinash Kaushik

Incrementality: the only test that proves cause

The fix is the same one medicine settled on a century ago: a controlled experiment. Incrementality testing splits your audience at random into a group that sees the ad and a holdout group that doesn't, then compares conversions between them. The difference, the lift, is the genuinely incremental result, the sales that wouldn't have existed otherwise. The major platforms run versions of this: Meta's Conversion Lift studies do it at the user level, and geo experiments (Meta's GeoLift, Google's geo tests) do it by region when you can't split individuals. Google now calls incrementality testing "the industry's gold standard for understanding advertising's true impact in a privacy-first way."

The results are often humbling. Branded search, retargeting and discount affiliates, the heroes of a last-click dashboard, frequently show low incrementality, because they catch people who were already converting. So before you scale a channel that "performs," run a holdout. Turn it off in a few regions or for a random slice of the audience for a few weeks and watch whether total conversions actually fall. If they don't, you've been paying to take credit for free sales. This is the single most decision-changing test in the kit, and it usually costs nothing but the nerve to switch something off.

flowchart TD
  Q("A customer converts after seeing an ad") --> A("Attribution: which touchpoints get the credit?")
  Q --> I("Incrementality: would they have converted anyway?")
  A --> Desc(["Describes the journey (correlation)"])
  I --> Exp(["Needs a holdout / experiment"])
  Exp --> Cause(["Proves causal lift"])
  Desc --> Risk(["Risk: over-credit demand-harvesting channels"])
Attribution and incrementality answer different questions. Only the experiment with a holdout group establishes cause. Leaders Loop

The honest caveat: experiments are bounded. A holdout tells you the lift of this campaign, in this window, at this spend level, it won't capture long-run brand effects that build over months, and small tests can be statistically noisy. Treat each test as one reliable data point, not a universal law about a channel.

Marketing-mix modelling, the long view, and why it's back

Attribution and incrementality both lean on tracking individuals. Privacy regulation, the death of third-party cookies and walled-garden data have shredded that, practitioners now estimate usable identity coverage at roughly 30–60%, down from the 90%-plus of the cookie era. So the industry has dusted off a pre-cookie technique: marketing-mix modelling. MMM ignores individuals entirely, taking aggregate time-series data, weekly sales, spend per channel, price, seasonality, even weather, and using regression to estimate how much each input moved total sales. Needing no personal data, it survives the privacy reckoning intact.

The idea isn't new. In Harvard Business Review in 2013, Wes Nichols framed "Advertising Analytics 2.0" as three linked activities, attribution, optimisation and allocation, arguing that judging each touchpoint in isolation misrepresents how a real mix drives sales. What's changed is access: Meta open-sourced its MMM tool Robyn in 2020, and Google launched Meridian (built on its earlier LightweightMMM code) in 2025, so a method that once needed a seven-figure budget is now within reach of a competent in-house team. The rule of thumb: if you spend across more than two or three channels and run regular brand campaigns, put an MMM in place for the strategic split and use incrementality tests to calibrate it. MMM sets the direction; experiments check the model isn't fooling itself.

There's a strategic reason this matters beyond measurement plumbing. Les Binet and Peter Field, analysing the IPA's databank of 996 campaigns in The Long and the Short of It (2013), found that the most effective marketing splits its budget roughly 60/40 between long-term brand building and short-term sales activation. Last-click attribution is structurally blind to the 60, brand effects are slow, diffuse and don't end in a trackable click, so a business that optimises purely to its dashboard keeps defunding the very work that drives its long-run growth. That's the deepest danger of measuring only what's easy to measure: the right answer is to run the three methods as a triangle, each covering the others' blind spots.

flowchart LR
  MMM(["Marketing-mix modelling"]) -->|"sets the strategic budget split"| Decision(["Where the money goes"])
  Inc(["Incrementality tests"]) -->|"prove causal lift, calibrate the model"| Decision
  Attr(["Attribution"]) -->|"granular, in-flight optimisation"| Decision
  Decision -->|"each method covers the others' blind spots"| MMM
No single method is enough. MMM, incrementality and attribution form a measurement triangle. Leaders Loop

A worked example

Illustrative figures, a composite scenario, not a real company's data.

Meridian Coffee Co., a direct-to-consumer roaster, spends £40,000 a month: £18k on Meta and Google awareness ads, £10k on branded search, £8k on retargeting, £4k on a discount-affiliate network. Its last-click dashboard is emphatic: branded search and retargeting together "drive" 70% of online revenue at a 9:1 return, while the awareness ads look like a money pit at barely 2:1. The obvious call is to cut awareness and feed the winners.

Before doing that, the team runs a four-week geo holdout: in a third of postcodes, they switch off branded search and retargeting. Total sales there fall by only about 6%, not the 70% the dashboard implied, most of those buyers simply converted through other paths. A separate Meta Conversion Lift study on the awareness ads shows clear positive lift: the "money pit" was creating the demand branded search was harvesting downstream. An MMM across two years of weekly data agrees, and adds that price promotions had been borrowing sales from future months.

The decision reverses. Meridian holds awareness spend, trims branded search to a defensive minimum, cuts the discount affiliate that showed near-zero incremental lift, and reinvests the saving into brand work, closer to the 60/40 split the evidence favours. Same £40k, materially more incremental revenue, and a reporting habit that now asks "what changed?" rather than "who got the last click?"

Frequently asked questions

Is data-driven attribution good enough on its own?

It's better than last-click, because it weights touchpoints by their modelled contribution rather than just recency. But it's still an attribution model, it distributes credit across observed journeys and can't see the counterfactual. It will still tend to over-credit channels that sit close to conversion. Use DDA for day-to-day channel reporting, but validate the big budget calls with incrementality tests, not DDA alone.

What's the difference between attribution and incrementality in one line?

Attribution tells you where a conversion came from along the journey; incrementality tells you whether your marketing caused it to happen at all. One describes the path, the other proves the impact.

We're a small business with no data-science team. Where do we start?

Start with the cheapest causal test you have: a holdout. Turn off one "high-performing" channel, branded search or retargeting are the usual suspects, in a region or for a random audience slice for two to four weeks, and watch whether total orders actually drop. You don't need a model to learn whether a channel is incremental; you need the discipline to switch it off and look.

Has the cookie's decline killed attribution?

It has badly degraded user-level tracking, which is why attribution numbers are noisier and why marketing-mix modelling, which needs no personal data, has come roaring back. The practical answer isn't one method but the triangle above: MMM for the strategic split, incrementality for causal proof, attribution for granular, in-flight optimisation.

Why does my dashboard say branded search is our best channel?

Because branded search sits right before the purchase and catches people who already decided to buy you, so last-click and even data-driven models hand it disproportionate credit. It's often the least incremental spend in the account. Run a holdout before you scale it.

Related in the Toolkit

Where to go next