Measurement: Attribution, Incrementality and Payback

The question measurement is meant to answer

Every gambling marketing team can say how many first-time depositors each channel "delivered" last month. Very few can say how many would have arrived anyway. The first number comes free from a dashboard; the second is the only one that matters when deciding whether to spend another pound, and no tracking system will ever report it.

That gap is the subject of this lesson. It covers why the default numbers overstate, what consent rules and browser and device changes have done to attribution, how incrementality tests and marketing mix models work and where each one breaks, and how acquisition cost, customer value and payback fit together in the economics operators actually report. The CRM and Player Lifecycle course covers what happens to a customer after acquisition; this lesson stops at the point where marketing spend has to be justified.

Gambling makes the problem harder in three ways. Customer value is highly skewed, so averages hide most of what is going on. Promotions are a large part of acquisition cost and are frequently booked outside the marketing line, so "cost per acquisition" depends on what you count. And the responsible gambling obligations from the Marketing Compliance course mean that the most "valuable" customers in a model can be precisely the ones a licensee should be protecting, which puts a constraint on optimisation that other sectors do not face.

Why last-click lies

Last-click attribution gives all the credit for a conversion to the final tracked touchpoint before it. It is the default in most affiliate tracking, it was the default in most web analytics for years, and it systematically rewards the channels that sit closest to a decision the customer had already made. Brand search and retargeting look extraordinary under last-click; television and sponsorship look worthless.

The best-known evidence comes from outside gambling. In a large field experiment at eBay, Blake, Nosko and Tadelis switched off paid search in randomly chosen regions and measured what happened. Their working paper found that brand-keyword ads "have no measurable short-term benefits", because almost all of the traffic that the paid brand clicks had been receiving simply arrived through the organic listing instead. The same paper showed how badly the conventional measurement misled: observational methods produced returns in the thousands of per cent, while the experiment found a return that was negative. The study was published in Econometrica in 2015; the specific percentages quoted here come from the 2014 working paper version.

A second study generalised the point. Gordon, Zettelmeyer, Bhargava and Chapsky compared the results of fifteen large randomised advertising experiments at Facebook with the answers that standard observational methods would have given on the same campaigns, and concluded that observational methods "often fail to produce the same effects as the randomized experiments", even after controlling for extensive demographic and behavioural data. The bias is systematic: ad systems show ads to the people most likely to convert anyway.

For a gambling operator the implication is direct. A brand that already has strong awareness, and that bids on its own name, will see paid brand search report a very low cost per first-time depositor. Some of those customers were reached by the ad. Many of them typed the brand name because of a sponsorship, a television campaign, a friend or an affiliate review, and would have clicked the organic result or the app icon anyway. The same applies to affiliates permitted to bid on the brand. The question is not "is the CPA low?" but "what happens to total depositors when we stop?". Only a test answers it.

What consent and privacy changes did to attribution

Even last-click attribution depends on following a person from an ad to a conversion, which has become steadily harder, especially in gambling, where the conversion usually happens inside an app after identity verification.

Consent. In the UK, storing or reading information on a device for advertising needs consent. The Information Commissioner's Office guidance on storage and access technologies, finalised in April 2026, states that using these technologies for online advertising purposes "requires consent", and that measuring campaign effectiveness is treated as part of that advertising purpose. The Data (Use and Access) Act 2025 introduced an exception for purely statistical analytics, but the ICO is explicit that the statistical purposes exception "does not apply to purposes related to online advertising". Every visitor who declines consent is a conversion that ad-level attribution cannot see.

Mobile operating systems. On Apple devices, apps must obtain permission through App Tracking Transparency before tracking a user across other companies' apps and websites, and Apple's definition of tracking expressly includes linking data "for targeted advertising or advertising measurement purposes". Most gambling acquisition now ends in an app, so a large share of mobile conversions arrive without a user-level path back to the ad.

Browsers. The long-predicted end of third-party cookies in Chrome did not happen in the form expected. Google abandoned the plan to deprecate them in 2024, and in October 2025 announced it would retire most Privacy Sandbox technologies, including the Attribution Reporting API, citing "low levels of adoption". For a planner, browser-level tracking survives in Chrome but is overridden by the consent layer in any case.

Analytics defaults. Google Analytics 4 now uses data-driven attribution by default, and the first click, linear, time decay and position-based models were removed in November 2023. Data-driven attribution distributes credit across observed touchpoints, which is better than last click, but it can only distribute credit among touchpoints it can see. A sponsorship, a television spot, a podcast read or a poster never appears in the path, so it can never be credited.

The result is that attribution in 2026 is best understood as a partial, biased sample of journeys. It is useful for optimising within a channel (which keyword, which creative, which audience) and unreliable for deciding between channels. Decisions between channels need methods that do not depend on following individuals at all.

Incrementality testing: holdouts, conversion lift and geo-lift

An incrementality test creates a credible version of the world without the advertising and compares it with the world with it. There are three common designs.

User-level holdouts. The platform randomly withholds ads from a share of the target audience and compares conversions between the exposed and unexposed groups. Google describes its Conversion Lift product as comparing people who see your ads with people who do not, and measuring "the increase in conversions that are caused by the presence of the ad", reporting incremental conversions and incremental cost per action. Meta's equivalent creates a randomised test group and a control group "who don't see your ads". They are cheap and statistically clean, but they are run by the seller of the media, on the seller's conversion data.

Geo-lift tests. Advertising is switched on, off or up in a set of regions while comparable regions continue as before, and the difference is measured in the operator's own first-party data: registrations, first deposits, net revenue. Meta's open-source GeoLift package describes this as measuring "the true incremental effect (Lift) of ad campaign" using synthetic control methods, which build a weighted combination of untreated regions that tracked the test regions closely before the test began. Geo tests suit gambling well. The conversion is recorded by the operator, not the platform; the method works for television, out-of-home and radio as well as digital; and regulated markets are already organised by geography, which in the United States means state lines provide natural test boundaries. The weaknesses are spill-over (national television and sponsorship cannot be switched off by region), small numbers of regions in small markets, and events: a local derby or a regional holiday in a test cell can swamp the effect.

Planned on-off tests. The bluntest design: stop a channel entirely for a period and watch total outcomes. It suits brand search and affiliate placements, where the question is cannibalisation, and misleads for anything with a long lag, such as brand advertising.

A worked readout

The figures below are hypothetical and illustrate how to read a result, not what any channel will deliver.

An operator runs a six-week geo-lift test on paid social. Twelve regions receive an extra £400,000 of spend; a synthetic control built from twenty untreated regions predicts what the test regions would have done without it.

Three readings follow. First, the platform's reported cost per depositor of £250 understates the true incremental cost by more than three times. Second, the interval is wide: the true incremental cost could plausibly be anywhere from under £500 to over £2,000, which is normal. Lewis and Rao, analysing twenty-five large field experiments, found that the "median confidence interval on return on investment is over 100 percentage points wide" because individual purchasing is so volatile relative to the effect of any one campaign. Third, the factor of 0.31 is the useful output. It can be applied to the platform's day-to-day reporting until the next test, turning a biased but timely number into an approximately calibrated one.

What a readout does not tell you is whether the incremental depositors are as valuable as the ones who would have come anyway. Measure net revenue per incremental depositor over at least three months before declaring the channel a success, because a channel can drive cheap, low-value registrations that never pay back.

Marketing mix modelling

A marketing mix model is a statistical model of total outcomes (registrations, deposits, revenue) over time, explained by spend in each channel together with everything else that moves the numbers: the sports calendar, seasonality, pricing and promotions, competitor activity, regulatory changes. It needs no user-level data at all, which is why it has returned to the centre of measurement as tracking has degraded.

Two open-source frameworks have made it accessible to in-house teams. Google's Meridian, made generally available in January 2025, is described on its repository as a framework "based on Bayesian causal inference" that supports regional data. Meta's Robyn is described as a "semi-automated and open-sourced Marketing Mix Modeling (MMM) package", suited to advertisers with rich digital data. Neither is a substitute for a modeller who understands the business.

Three features of a good gambling mix model are worth insisting on.

Adstock and saturation. Advertising effects persist after the spend (adstock) and diminish as spend rises (saturation). A model that ignores either will overvalue the latest channel and undervalue the long-lived ones, notably sponsorship and television.

The calendar. Major tournaments, the start of domestic seasons and the big racing festivals drive both spend and demand. A model that does not control for the fixtures will attribute to the advertising what the fixtures did.

Calibration with experiments. A mix model infers effects from correlations over time, and when every channel rises and falls together in the same weeks it cannot separate them. Feeding the results of geo-lift tests into the model as priors, or checking the model's channel estimates against them, is now the standard way to keep it honest. The Interactive Advertising Bureau's November 2025 guidelines on incremental measurement set out the same hierarchy, grouping methods into experiments, model-based counterfactuals, econometric models and hybrid proxies and emphasising credible counterfactuals and control of bias.

The working arrangement most mature operators converge on is a triangle: attribution for daily optimisation within channels, experiments to establish causal truth for the channels that matter most, and a mix model refreshed quarterly to allocate budget across everything, calibrated by the experiments.

CPA, LTV and payback

Acquisition economics in gambling are usually expressed through four numbers: the cost per acquisition of a first-time depositor, the lifetime value of that customer, the payback period before cumulative value covers the cost, and the ratio between value and cost.

Operators that report on these do so in their own defined terms. DraftKings, at its November 2023 investor day, defined customer acquisition cost as external marketing spend divided by new customers, including "national spending, team and league deals", and said it had historically acquired customers at less than three-year paybacks on adjusted gross profit. Flutter, reporting its 2024 results in March 2025, said its US business was investing with payback periods "of less than 18 months and remaining well under our 24-month threshold", defining payback as the projected time for players to generate enough adjusted gross profit to repay the average cost of acquiring them. Two lessons follow. Payback is the metric investors are given, and the definitions differ, so comparisons between companies need care. The Reading iGaming Financial Results course covers how to read these disclosures.

A worked calculation

Again the figures are hypothetical. An operator acquires a cohort of customers with these averages per first-time depositor:

At twelve months the cohort has repaid £247 of £360. If net revenue holds at £18 a month (£12.60 of contribution), the remaining £113 takes about nine more months, so payback arrives at roughly month 21. On paper, that is inside a 24-month threshold.

Now apply the incrementality factor. If a test shows that only 60% of the attributed depositors were incremental, the true acquisition cost of each additional customer is not £360 but £600 (£360 divided by 0.6). On the same value curve, payback moves from month 21 to beyond month 40, and the channel fails any reasonable threshold. The spreadsheet did not change; the honesty of the denominator did.

Two cautions specific to gambling apply to the value side. First, averages conceal skew. In most gambling cohorts a small share of customers generates a large share of net revenue, so the "average" customer barely exists, and a lifetime value model can be dominated by a few accounts. Second, some of the highest-value customers in any cohort will be experiencing harm. A model that rewards channels for finding customers who resemble the top of the value distribution is, in effect, rewarding them for finding customers who gamble most heavily. The Responsible Gambling course covers markers of harm; the measurement discipline is to calculate value with restricted and excluded customers removed, and to treat any channel whose economics depend on the top few per cent as a risk, not a success.

Ratios that appear in results

Investors see marketing efficiency mainly through marketing as a share of revenue. Flutter reported that its US sales and marketing expense fell by 300 basis points to 20.2% of revenue in the fourth quarter of 2024. A falling ratio can mean better efficiency, a maturing customer base or simply a cut that will show up in next year's growth; lesson seven covers how to read it. The ratio the board watches is an output of the incrementality and payback work in this lesson, not a substitute for it.

A measurement plan that survives scrutiny

A credible measurement programme for a gambling operator has five parts, and none of them is a dashboard.

1. Define the conversion in first-party data. First deposit and net revenue at 30, 90 and 365 days, recorded by the operator, with restricted and self-excluded customers excluded from value.
2. Keep attribution for in-channel optimisation only. Use it to choose creative, keywords and audiences, never to allocate budget between channels.
3. Run a standing test calendar. At least one geo or holdout test a quarter, prioritising the biggest budgets and the channels most likely to be cannibalising (brand search, retargeting, affiliate review placements).
4. Refresh a calibrated mix model quarterly. Feed it the tests, control for the sports calendar, and use it for cross-channel allocation.
5. Report payback on incremental cost. Every acquisition report shows the attributed cost, the incrementality factor used and its date, and the resulting incremental cost and payback.

The test of the programme: when finance asks what would happen if a channel's budget were cut by a quarter, the team answers with evidence rather than the platform's own report.