Ask a CRM manager what last month’s reload campaign delivered and the answer arrives in about 30 seconds. Deposits from claimers, turnover, NGR over the following four weeks, set against the bonus cost. A ratio. Usually a flattering one.
Ask what the campaign changed and the room goes quiet.
The two questions sound like the same question. They are not, and the gap between them is where a large share of the promotional budget in this industry goes to die.
Here is the mechanic. Players who claim a reload offer had to open the message, read it and act on it. That is already a filtered group, filtered on precisely the trait that predicts revenue: they were paying attention. Compare them with the players who ignored the email and you have measured attention, not the bonus. The offer’s real contribution is buried somewhere inside that difference and there is no way to dig it out after the fact.
Every operator we have worked with knows this in the abstract. Most still run the comparison, because it is the one the platform reports natively and because the alternative involves telling the commercial director that five percent of the eligible base will receive nothing this month.
eBay switched off its ads and nothing happened
The most quoted evidence for this problem comes from outside gambling. A team of economists working with eBay, whose results were published in Econometrica in 2015, ran a series of experiments in which paid search advertising was switched off in selected markets to see what would actually change. Brand keyword advertising produced no measurable short-term benefit, and because the bulk of the spend landed on loyal users whose behaviour did not move, average returns came out negative.
Attribution had been reporting the opposite for years.
Swap paid search for a Thursday free spins push and the structure of the error is identical. Bonus spend concentrates on the most active segment because that segment responds, and response gets read as causation. Nobody checks whether those players were going to deposit anyway, because checking requires deliberately not sending them the offer.
Research from inside the sector points the same way, and one finding in particular should worry anyone running a CRM calendar. Australian work on wagering inducements, carried out by the Experimental Gambling Research Laboratory at CQUniversity, found that direct messages carrying no incentive at all prompted betting roughly as effectively as messages carrying one. If the notification is doing the work, the bonus stapled to it is a rebate.
The regulator and the tax authority turn up together
Sloppy measurement here used to be survivable. Two things happened.
From 19 December 2025, under changes confirmed by the Gambling Commission, operators licensed in Great Britain must cap wagering requirements at 10 times and can no longer combine more than one gambling product inside a single incentive. During the consultation it emerged that some licensees had been running requirements as high as 60 times. Wagering requirements are what made a headline bonus cheap in practice, since a meaningful share of the money never leaves the account. Cap them and the real cost of every surviving offer rises.
The tax side is starker. Analysis by the Tax Policy Center found that in American states publishing the data, promotional deductions have accounted for roughly a third of taxable gross gaming revenue, and in heavy campaign months have exceeded 100 percent of it, leaving nothing to tax at all. The Tax Foundation put Colorado’s 2022 collections at 5.6 percent of pre-deduction GGR against a headline rate of 10 percent. Legislators read those reports. Since 1 July 2026, Colorado sportsbooks can no longer deduct the value of promotional free bets from taxable revenue.
For years part of every free bet was quietly subsidised by the tax base. That subsidy is being withdrawn market by market, and the bonus line is about to start behaving like a real cost centre.
What a holdout actually costs, and what it buys
Incrementality asks one thing. What would have happened if this offer had never been sent.
There is exactly one dependable way to find out, and it is to randomly withhold the offer from part of the eligible population and compare. Everything else is a proxy that breaks under pressure.
The arithmetic is trivial. Incremental NGR is the treated group’s NGR minus the holdout group’s, scaled to equal size, and the return is that figure divided by total bonus cost. All of the difficulty sits in the design, and most of the design failures are organisational rather than statistical.
Randomise inside the eligible segment, never across it. If the campaign goes to players who deposited in the past 14 days, the holdout comes from that same pool, on the same day, picked at random. Five to 10 percent is enough for a large base. Freeze the membership before the campaign runs and then leave it alone, because the moment somebody pulls three complaining VIPs out of the control group the comparison is finished and no amount of later analysis will repair it. Fix the measurement window in advance too. Bonus effects decay fast, and a seven day window and a 60 day window can hand you opposite verdicts on the same campaign, which makes choosing the window afterwards a way of choosing the answer.
One more, and it is the one that catches finance-literate operators off guard: measure NGR net of the bonus cost rather than GGR. A campaign can lift turnover, lift bonus cost by more, and still look like a triumph on every dashboard that stops at the gross line.
This is largely a plumbing problem, which is why it fails so often in practice. The eligible population needs to be a rule rather than a spreadsheet somebody pulled in March, which in Gamblitude means a Dynamic List built from Metrics and Attributes and reused every cycle. The holdout itself is better handled as a Static List with a snapshot label, frozen at the moment of assignment, with segment history recording who entered and left. That sounds like housekeeping until the quarterly review, when somebody asks whether the control group was tampered with halfway through and the answer needs to be a log rather than a recollection. Pushing the holdout to the CRM as a suppression list through a webhook is the step manual processes get wrong most often, usually about six days into the campaign.
The statistics are worse than you want them to be
Now the part vendors tend to skip.
Player-level revenue is brutally skewed. A handful of accounts drive most of the outcome, and in sportsbook you are also reading through betting variance, which means a single settled accumulator can outweigh the entire treatment effect you are trying to detect.
Digital advertising hit this wall a decade ago. Randall Lewis and Justin Rao, writing in the Quarterly Journal of Economics, examined 25 large field experiments and found the median confidence interval on return on investment ran more than 100 percentage points wide, with a coefficient of variation around 10 in individual sales. Detecting a real effect under those conditions can require populations far larger than any single campaign will ever reach.
So some campaigns will never be measurable on their own. Accept that early and the rest of the programme gets easier.
Three things help, and they compound. Use pre-period behaviour as a covariate: the method known as CUPED, published by Alex Deng and colleagues at Microsoft in 2013 and now standard practice at Netflix, Booking.com and Airbnb, strips out the predictable part of each player’s outcome and can cut the sample you need substantially. Deposits in the month before the campaign predict deposits during it, so remove that and what remains is closer to signal. Second, cap extreme outcomes before comparing and report the result both ways, because one jackpot should not decide a budget. Third, and most usefully, stop measuring individual sends. Twelve midweek free spins pushes, pooled as a single recurring mechanic against a persistent holdout, will tell you something. One of them will not.
The persistent holdout deserves more attention than it gets. Keep five percent of the active base out of all non-mandatory promotional contact for a full quarter and you get a clean read on what the entire CRM bonus programme is worth. That number is more useful than any individual campaign result, and considerably easier to obtain. It is also the number a CFO will ask for, sooner or later.
Spend where the effect is, not where the money is
Once you can measure the effect, targeting changes shape.
The players with the highest predicted value are rarely the ones a bonus moves most. The players with the highest churn risk are frequently already gone. Eva Ascarza of Harvard Business School demonstrated this in the Journal of Marketing Research, combining two field experiments with a set of machine learning models to show that the customers at highest risk of leaving are not necessarily the best people to aim a retention programme at. The quantity worth modelling is the difference between treating a player and leaving them alone, estimated per player.
That is uplift modelling, and it has a hard prerequisite. It needs experimental data. Without a holdout that has been running for a while, a model can only learn who takes bonuses, which is a question nobody needed answered.
This is where churn probability and early VIP scoring stop being separate exercises. Prediction tells you who is worth an intervention. Incrementality measurement tells you whether the intervention works on them. Run only the first and you will spend efficiently on people who did not need spending on.
Meanwhile the slow drift needs watching. Bonus cost as a share of GGR, wagering completion rates, redemption by segment. These move gradually and get noticed at quarter end, which is roughly two months too late. Continuous monitoring through Insight Radar catches the campaign whose cost per incremental player has quietly doubled, and holding the bonus cost ratio against plan in Targets turns overspend into a variance somebody has to explain rather than a surprise in the management accounts.
What you should expect to find
The first honest result usually stings.
In our experience a substantial share of a mature bonus programme turns out to be paying for behaviour that was already coming, and it clusters exactly where the eBay experiments said it would, in the most engaged segment. Welcome offers hold up reasonably well. Reactivation offers aimed at players dormant past 90 days hold up worst, which is awkward, because those are the campaigns with the best-looking response rates.
None of this is an argument against bonusing. Bonuses work. They just do not work uniformly, and until now there has been no commercial penalty for failing to know which half was which.
That penalty has arrived. The wagering mechanics that made offers cheap are being capped, the tax deductions that subsidised them are being withdrawn, and the operators who come through the next two years in decent shape will be the ones who can point at the bonus line and say which portion of it is an investment.
Nobody gets that answer from a better dashboard. You get it by holding a group of players out, writing the rules down before you start, and then living with whatever comes back.
