Algorithmic targeting in betting marketing is the use of a customer’s own behaviour data to decide which promotion they see, when it arrives, and what it is worth. Stated that way it sounds like ordinary modern advertising, because it mostly is. The complication is that the same model which works out that a casual bettor likes a free bet on Sunday football can also work out which customers are losing more than they planned and still logging in.
That is the question now in front of the Massachusetts Gaming Commission, and it is worth being precise about what the regulator has and has not said. The commission has not published a finding of AI gambling marketing harm. It has asked the operators it licenses to explain what their systems actually do.
The Massachusetts letter, and what it does not allege
The Massachusetts Gaming Commission gave the state’s seven licensed mobile sports betting operators until 3 November to explain how they use artificial intelligence and machine learning in marketing, promotions and customer engagement. The request came in a letter dated 9 October from Executive Director Dean Serpa, setting out questions across eight operational areas.
The letter asks operators to disclose whether they rely on AI, algorithms, automated decision-making or comparable tools; to name the specific systems; to say what each one is for; and to state whether it was built in-house or bought from a third-party supplier. It also asks whether those technologies make decisions about individual customers, and how operators identify potentially harmful gambling behaviour.
The single question underneath all of it: can a customer’s losses, deposit patterns or riskier betting behaviour make them more likely to receive a promotion? Nothing in the request says that is happening. A regulatory probe of this kind is a mapping exercise. The commission is asking because it does not currently know, which is itself the most interesting fact in the story.
How AI targeting actually works in betting marketing
Targeted gambling ads are built from first-party behavioural and transactional data, fed into predictive models that sort customers into segments, which are then matched to pre-approved offers and delivery timings. There is no mystery layer. It is customer relationship management, the same discipline retailers and subscription services use, applied to an account that also happens to record every stake a person places.
Data signals operators use
A licensed sportsbook already holds, for compliance and payments reasons alone, a precise record of each account. The signals typically available to a marketing model include:
- Deposit frequency, size and payment method, including failed or reversed deposits
- Stake size over time, and whether it is rising, falling or spiking
- Session length, time of day, and how quickly bets follow one another
- Net position, meaning cumulative wins and losses on the account
- Market and sport preferences, including appetite for parlays and in-play betting
- Response history: which push notifications, emails and odds boosts were opened and acted on
- Dormancy, the gap since the last login or last bet
Models trained on that history predict things marketers care about: likely future value, probability of churn, probability of responding to a given incentive. Lookalike audiences extend the same logic outward, finding new prospects on ad platforms who resemble existing high-value customers.
From data to a personalised offer
Segmentation turns predictions into categories, and categories into campaigns. A customer who has not logged in for three weeks sits in a reactivation segment and receives a win-back incentive. A customer whose weekly stake has tripled sits in a high-value segment and may be routed to VIP or account-manager outreach. Both treatments are automated, both are triggered by behaviour, and neither requires anyone to have looked at the individual.
The table below describes the standard CRM playbook as the industry generally describes it, not the practice of any named operator.
| Behaviour signal | Typical segment | Typical marketing response |
|---|---|---|
| No login for several weeks | Lapsed / dormant | Reactivation push, free bet, deposit match |
| Rising stake and deposit size | High value / VIP candidate | Personal outreach, hospitality, enhanced limits |
| Declining activity after a loss run | Churn risk | Retention offer, odds boost on preferred market |
| Heavy in-play, short session gaps | High engagement | Real-time notifications, live-market promotions |
Why the harm risk is not the same as bad intent
The harm pathway regulators describe is a timing problem more than a content problem. The offers themselves are usually compliant; the concern is that behavioural triggers can deliver them precisely when a person is least able to turn them down.
Three pathways come up repeatedly. The first is re-engagement of people who stepped away. A dormancy signal cannot distinguish between a bettor who got bored and a bettor who stopped deliberately because they were losing control, and a self-excluded customer who is not correctly suppressed from every channel can be pulled back by an automated campaign.
The second is loss chasing. A model optimising for response rate may discover, without anyone designing it to, that customers who have just lost are unusually likely to accept a deposit match. Optimisation does not know what it is optimising into.
The third is VIP escalation. High-value segments attract the most attention because they generate the most revenue, and heavy spending is also the clearest observable marker of possible harm. The same signal flags both, which is why regulators want to know whether the flag routes to marketing, to safer-gambling staff, or to both.
Where the evidence actually runs out
Nobody has published a reliable figure for how much algorithmic targeting increases gambling harm, and anyone quoting one should be asked where it came from. The mechanisms above are plausible and well described in regulatory discussion. What is missing is the operational detail: which models are in use, what they optimise for, how suppression lists are enforced across email, SMS, push and paid social, and whether risk flags are shared between marketing and compliance teams.
That is exactly the gap the Massachusetts request is written to fill. Treating the probe as proof of wrongdoing gets ahead of the evidence in one direction; treating it as a formality gets ahead of it in the other. The honest reading is that a regulator has noticed it cannot audit something it does not have visibility into.
A wider pattern in gambling regulator AI scrutiny
The Massachusetts request fits a broader supervisory shift from reviewing advertising content to reviewing the systems that decide who sees it. Regulators have spent years policing what a betting ad says. The newer question is how the distribution list was built.
Several strands point the same way. Affordability and source-of-funds checks in some jurisdictions already require operators to act on spending data, which means the data is demonstrably there and usable for protective purposes as well as commercial ones. Licensing bodies such as the Malta Gaming Authority supervise operator conduct toward customers as a condition of licence, and general-purpose AI transparency rules increasingly apply to automated decisions affecting individuals regardless of sector. Whether any specific authority opens its own inquiry is a separate matter, and not something to assume. The underlying argument about whether gambling needs AI-specific rules at all is one the industry has not settled, as covered in our look at AI in online casinos and the case for gambling-specific rules.
What responsible use looks like in practice
Responsible AI use in betting marketing is mostly a matter of suppression, caps and escalation routes, none of which are technically hard. The difference between protective and exploitative targeting is usually governance, not capability.
Concretely, that means self-excluded and cooling-off customers are removed from every marketing channel, including paid social and lookalike seed lists, not just the in-house email file. It means frequency caps on how often any individual can be contacted, and intensity caps on offer value. It means affordability or risk triggers that suppress promotional contact rather than adding to it, so that a customer showing harm markers stops receiving offers automatically. And it means human review before anyone is enrolled in a VIP programme, with problem gambling prevention staff able to override a commercial decision.
The test is simple to state: does a harm flag ever increase marketing pressure? If the architecture makes that possible, the safeguards are decoration.
Recognising and limiting targeted betting marketing
You can usually tell you are in a behavioural segment by the timing of what arrives. Offers that appear within hours of a losing session, a reactivation bonus that lands shortly after you stopped logging in, or personal contact from an account manager after your stakes rose are all signs of automated triggers rather than a general campaign.
What you can do about it:
- Switch off marketing consent in account settings, separately for email, SMS and push. Opting out of one channel rarely covers the others.
- Set deposit, loss and session limits before you need them. Limits applied in advance are harder to argue yourself out of than limits applied mid-session.
- Use cool-off or self-exclusion if the pull is persistent, and check that it covers marketing as well as account access.
- Decline VIP enrolment if your spending has recently increased. Hospitality and enhanced limits are a commercial response to that increase, not a reward for skill.
- Reset ad identifiers on your phone and clear tracking permissions to reduce retargeting outside the operator’s own channels.
If betting has stopped feeling optional, independent support exists in most regulated markets, and it is free. Verify any service before sharing details with it, as our guide to checking a gambling support service explains in the context of how to spot a legitimate problem gambling helpline.
By the November deadline, Massachusetts will know more about what these systems do than the public does now. Whether that becomes a rulebook, a disclosure standard, or nothing at all is still open. What has already changed is the question being asked of operators: not whether an ad was accurate, but whether the decision to send it was made by a model that knew the customer was struggling.