Incrementality vs Attribution: Which Question Each Answers

It is written for performance marketers, marketing leaders and agencies who have been told that attribution is dead and incrementality is the answer, or the other way round. Neither is true, and the useful work is knowing which question you are asking.

Attribution looks backwards at conversions that happened and divides them among touchpoints. Incrementality runs an experiment and compares the world with the ad against the world without it.

• Attribution

  Of the leads and deals you got, which ads, campaigns and channels were on their path, and how much credit each gets under a chosen attribution model. Every conversion is shared out, so the credits add up to what happened.

• Incrementality

  How many conversions exist because of the ad. Measured by comparing a group that could see the ads with a holdout that could not, by people or by region. The difference is the incremental conversions; the rest would have happened anyway.

The two are not rivals. Attribution is a map of what happened, detailed enough to act on every day. Incrementality is a measurement of cause, coarse and slow but the only way to know whether a campaign is creating demand or collecting it.

Every ad platform attributes conversions to its own ads, and none deducts the people who would have converted anyway. Four places are where that difference gets large.

• RetargetingA retargeting audience is people who already visited your site, many of them already deciding. Any who come back and convert after seeing an ad get credited to it, whether the ad changed anything or not.
• Brand searchSomebody who types your company name already knows you. A brand ad sits on top of the organic result they would have clicked, and collects the conversion. In a large field experiment at eBay, published in Econometrica, brand keyword ads showed no measurable short-term benefit.
• View-through conversionsA platform credits an impression that was never clicked if the person converts inside its view window. Exposure is not cause, and the people most often shown an ad are the ones most likely to convert. See click-through vs view-through conversions.
• Every platform claims the same leadMeta, Google and LinkedIn each count the lead their ads touched, and none deduplicates against the others, so the platform totals add up to more leads than the CRM holds.

The last of the four is an attribution problem that better attribution fixes: count each lead once, in your own data, joined to the CRM. The first three are not. Even perfectly deduplicated, first-party, CRM-joined attribution still credits a retargeting click on a lead who was coming back anyway, because attribution describes the path, not the cause. That is the gap an incrementality test measures.

The research agrees. Gordon and colleagues, in Marketing Science, compared the observational methods advertisers use with large randomised experiments run at Facebook and found that the observational estimates often failed to recover what the experiments measured. The eBay study above reached the same conclusion for paid search: returns were a fraction of what the non-experimental numbers suggested.

One illustrative month for a lead generation account, with every figure adding up. The retargeting audience is 200,000 site visitors, split at random: half can see the ads, half are held out.

On attribution, retargeting is the best campaign in the account at half the cost per lead. Now the holdout. The 100,000 people in the test group produced 800 leads from every source in the month, a rate of 0.8%. The 100,000 held out produced 700, a rate of 0.7%, without seeing a single retargeting ad.

The campaign caused 100 leads, not 500. The other 400 attributed to it were people who would have come back anyway, through a bookmark, a brand search or an email. Is the difference real? A two-proportion test on 800 of 100,000 against 700 of 100,000 puts it at about 99% confidence, so yes. With a 10% holdout of the same audience it would not have been: the difference would sit inside the noise, which is the most common way these tests end. The incrementality calculator does the significance read for your own numbers.

Suppose a geo test on prospecting later finds 240 of its 360 attributed leads incremental, two thirds. Its incremental cost per lead is €18,000 ÷ 240 = €75. The ranking has flipped: prospecting creates a lead for €75, retargeting for €120. Attribution was not wrong about what happened; it was answering a different question.

ROAS divides attributed revenue by spend. Incremental ROAS, iROAS, divides only the revenue the ads caused by spend, so it is the figure that tells you whether a euro in produced more than a euro out.

In the example, the retargeting leads closed into deals worth €6,000 each. With won deals sent back to the platform, it attributes 25 deals to the campaign: €150,000, a ROAS of 12.5. In the CRM, the test group closed 40 deals (€240,000) and the holdout 35 (€210,000). The ads caused €30,000 of revenue, an iROAS of 2.5.

An iROAS of 2.5 is still a campaign worth running if your margin clears it. The point is not that ROAS lies, it is that ROAS cannot tell you what happens if you cut the budget. iROAS can. The ROAS calculator covers the attributed side and the break-even point.

Most decisions a marketing team makes in a week are too small, too fast or too granular for an experiment. Attribution is built for those.

• Daily optimisation inside a channelWhich campaign, ad set, keyword or creative brings leads that turn into deals. No test can answer that for 400 keywords every week; a record per lead can.
• Lead quality per sourceTwo campaigns at the same cost per lead can produce very different pipelines. Only attribution joined to the CRM shows which source brings qualified leads and which brings form fills.
• Understanding the journeyWhich channels open journeys and which close them, how long a B2B buyer takes, which content sits in the middle. Multi-touch attribution is a map, and a test cannot draw it.
• Sending revenue back to the platformsBidding algorithms learn from the conversions you send them. Sending qualified leads and won deals with their value, via offline conversion tracking, needs each deal tied to its click.

In all four, the question is relative: this campaign against that one, inside a channel whose overall value nobody is questioning today. Attribution's bias towards people already converting is roughly the same for two prospecting ad sets, so the ranking between them holds even when the absolute numbers are generous.

An incrementality test earns its cost when the question is absolute: is this channel or campaign creating demand at all, and how much budget does it deserve?

• Sizing a channel's budgetBefore doubling or halving a channel, measure what it adds. Attribution tells you what it touched, and the two can differ by a factor of five, as in the example.
• Brand searchThe classic candidate. If competitors bid on your name, brand ads may protect real leads; if nobody does, a test may show you are paying for clicks you would get for free.
• RetargetingThe campaign attribution flatters most. A holdout on the retargeting audience is usually the cheapest, clearest test an account can run.
• Upper funnel, video and CTVChannels that are seen rather than clicked get little credit on click-based attribution and too much on view-through. A test is the fair judge in both directions.

The practical answer is calibration. A test gives you a ratio of incremental to attributed results per channel, and you apply that ratio to the attributed numbers you get every day until the next test.

1. Run attribution continuously, on one record per lead joined to the CRM, so every channel is counted once and on the same outcome.
2. Test the channels where the budget question is biggest: retargeting, brand search and whichever channel you are about to scale. One test at a time per channel.
3. Turn each result into a ratio: incremental leads, deals or revenue divided by the attributed figure for the same period and the same outcome.
4. Apply the ratio to the attributed figures when you compare channels and set budgets. Keep the raw attributed figures for decisions inside a channel.
5. Retest when something big changes: the creative, the audience, the budget level or the market. A ratio is a snapshot, not a constant.

Two cautions. The ratio is only valid against the attribution it was measured against: a ratio from Meta's own reported conversions cannot be applied to CRM-attributed leads, so compute it on the same numbers you will apply it to. And a ratio measured at €12,000 a month does not tell you what happens at €40,000, because returns usually fall as spend rises. Test again before a big step.

Both are randomised experiments, which is why they get confused. The difference is what the control group sees.

Marketing mix modelling estimates each channel's contribution from years of weekly spend and outcome data with a statistical model, including offline channels. An incrementality test measures one channel directly, in one period, with a controlled experiment.

They support each other more than they compete: a model is only as good as its assumptions, and test results are the standard way to check and calibrate it. The catch for most lead generation teams is that an MMM needs a lot of history and varied spend to learn from, more than many accounts have. MMM vs MTA covers what each needs and which a lead gen team can feed, and incrementality testing how to run the test.

Most incrementality content is written for e-commerce, where the conversion the pixel sees is a purchase with a value. In lead generation it sees a form fill, and the outcome that pays happens weeks later in the CRM.

• The outcome is in the CRMA campaign can cause more form fills and no more customers, if the extra leads are students, competitors and spam. A lift test read on form fills measures the wrong thing; read it on qualified leads and deals.
• Long sales cyclesA test that runs four weeks ends before most of its deals close. Read it first on qualified leads or pipeline created, then again on revenue once the deals from the test period have had time to close.
• Lower volumesA B2B account may produce a few hundred leads a month and a few dozen deals. A test needs enough conversions in the holdout to read, which pushes lead gen tests towards larger holdouts, longer runs and earlier funnel outcomes.
• Sales capacityIf the sales team cannot call more leads, extra leads in the test group close at a lower rate. Check that the test period did not change how leads were worked.

Platform lift studies read the conversions the platform receives. If the platform only receives form fills, that is what the study measures. Send qualified leads and won deals back with their value and a study can be read on outcomes closer to revenue; Google's Conversion Lift overview says its geography-based studies support offline data, and for any other study type, check whether it accepts offline events before you plan around it. For a do-it-yourself geo test, export deals from the CRM with their source and the customer's region, and compare regions.

LeadJourney is an attribution platform, and it sits on the attribution side of this post. One script on the site or in Google Tag Manager records every visit server-side on your own domain, first-party, at 95%+ tracking accuracy, and captures the click IDs (gclid, gbraid, wbraid, fbclid, li_fat_id, msclkid) at the click. The anonymous history is joined to the person at the form fill, the call or the booking.

Native CRM integrations (HubSpot, Salesforce, Pipedrive, Zoho and more) write the source, campaign and journey onto the record and read stage changes and won deals back. Spend from the ad platforms is joined to leads, deals and revenue per channel, campaign and ad, under five switchable models: first click, last click, linear, time decay and position based. Won deals go back to Meta through the Conversions API, to Google Ads through offline conversion import, to LinkedIn through its Conversions API and to Microsoft Ads, with the deal value.

What it does not do, stated plainly: LeadJourney does not run incrementality tests, lift studies, geo experiments or MMM, has no holdout feature, does not report by region and counts no view-through conversions. Visitors who decline consent are not measured. Its place in an incrementality programme is the outcome: the qualified leads, deals and revenue per campaign that a test should be read on, the deals sent back so a platform's own study or Meta's incremental attribution can learn from revenue rather than form fills, and the attributed figures you apply the test's ratio to.

Setup takes about 21 minutes, with a 14-day free trial, and the reports can be clicked through on demo data in the live demo first.

One line per question a marketing team asks most often, and the measurement built for it.

The cluster: incrementality defined, incrementality testing for running a test on CRM revenue, Meta's incremental attribution, MMM vs MTA and the incrementality calculator. On attribution: attribution model, click-through vs view-through conversions, cross-channel attribution and B2B marketing attribution. On sending outcomes back: offline conversion tracking and sending CRM data back to Meta and Google. Tools that do run incrementality: Rockerbox and Northbeam compared with LeadJourney.