Casino Personalization in 2026: Lift, Cold Start and Explainability
Published personalization lift in iGaming is real but mostly vendor-reported; the first session is still the hard part; regulators now expect tailoring to be explained.
TL;DR
Personalization in casino lobbies and CRM produces measurable lift, but nearly every published number comes from the vendor that sold the system, and only a few disclose a control group. The hardest part is still the first session, where a new player has no history: the problem recommender research calls cold start. Meanwhile regulators in Great Britain, the Netherlands and Sweden have moved from "do not harm" to "monitor, intervene and justify", which turns explainability from a feature into a working requirement. This post reports the evidence Adkuu could open in October 2026, separates measured from vendor-reported from unknown, and gives a checklist for evaluating a personalization vendor.
Key takeaways
- The best-documented 2026 result Adkuu could open is a two-month A/B test at one operator: a 2.2 percentage-point lift in first-time-deposit conversion from personalised onboarding video (iGB, 29 April 2026). Other published figures, such as a 35% turnover uplift, a 3.5x return on bonus spend or "up to 20%" of turnover redirected, are vendor-published without a described control.
- Cold start remains the open problem. A January 2025 survey by Weizhi Zhang and 17 co-authors defines it as "accurately modeling new or interaction-limited users or items" and sorts the field into content features, graph relations, domain information and the world knowledge of large language models.
- The Gambling Commission's chief executive told regulators on 20 October 2025 that "the general trend towards hyper personalisation" is something "we need to watch out for"; the Commission's AI page, updated 5 January 2026, commits to "Understanding and regulating (where appropriate) the use of AI by the gambling industry".
- The Netherlands requires continuous monitoring with intervention within an hour; a Swedish appeals court held in May 2024 that offering bonuses to players showing clear signs of risky gambling breached the duty of care. Both make "why did this player get this offer" an auditable question.
- Adkuu found no rule, in the documents it opened, that requires an operator to explain an AI-generated offer to a player in those words. The pressure is indirect, through consent, monitoring and duty-of-care rules, and it is growing.
What the published lift numbers show
Operators asking "how much does personalization lift" get answers that sound precise. The table lists every 2024 to 2026 figure Adkuu opened for this post, who reports it, and what the source discloses about how it was measured.
| Result | Who reports it | Design disclosed | Classification |
|---|---|---|---|
| A 2.2 percentage-point uplift in first-time-deposit conversion and an 86% higher Net Promoter Score among recipients of a personalised onboarding video, at Superbet | Seen.io and Xtremepush, via iGB (29 April 2026) | A two-month A/B test against a standard email onboarding sequence | Measured, vendor-run |
| "Up to 20%" of turnover and GGR redirected to more profitable titles across four brands; published ranges of 5–15% uplift in GGR and turnover, 10–15% more games played, 5–10% more active days | ZingBrain AI, via iGaming Future (9 July 2026) | A/B tests are mentioned; no duration, sample size or control is described | Vendor-reported |
| A 3.5x return on bonus spend, plus higher session length and spin volume | EveryMatrix, on its own site (23 December 2025), describing results with Future Anthem | No baseline, period or method | Vendor-reported |
| A 35% turnover uplift from personalised game recommendations, a 7% uplift in new games and a 300% increase in conversions from personalised campaigns, at Gala Bingo | Optimove, on its own blog (24 July 2024) | No control, timeframe or definition of "uplift" | Vendor-reported |
| The share of any lift that is incremental GGR rather than redistributed turnover; decay after the novelty period; the effect on players already showing risk markers | Nobody Adkuu could find | None | Unknown |
Three things stand out. The one result with a described control is also the smallest: 2.2 percentage points on a conversion rate, over two months, at one operator. That is a real and useful number, and it is nothing like 35%. The ZingBrain figure is honest about what a lobby recommender mostly does: it "redirected" turnover to titles that earn the operator more, which is margin management rather than new money. And the figures quoted most often in sales decks, the 71% of customers who "expect personalised interactions" and the 76% who are frustrated without it, cited by EveryMatrix from a McKinsey survey, are cross-industry consumer surveys and say nothing about gambling outcomes.
Adkuu's reading is that lift is real, bounded, and highly dependent on the baseline. An operator whose lobby was ranked by hand will see a bigger first-month change than one already running a decent popularity model; neither number transfers to the next operator.
Cold start: the part nobody has solved
A returning player with six months of sessions is easy to serve. A player who registered 40 seconds ago is not, and much of an operator's acquisition spend is decided in that first session. Recommender-systems research calls this cold start. The most complete recent treatment Adkuu could open is "Cold-Start Recommendation towards the Era of Large Language Models (LLMs): A Comprehensive Survey and Roadmap", submitted to arXiv on 3 January 2025 and revised on 16 January 2025 by Weizhi Zhang of the University of Illinois Chicago and 17 co-authors from universities and companies including Google DeepMind and Netflix. The authors define the problem as "accurately modeling new or interaction-limited users or items to provide better recommendations" and organise the field by what a system can lean on when it has no interactions: content features, graph relations, domain information, and the world knowledge held by large language models. They also note that the research community "still lacks a comprehensive review and reflection in this field".
For a casino lobby the asymmetry is useful. The item side is rich: a catalogue of thousands of slots and tables comes with volatility, theme, provider, mechanics and release date, so a new game is not really cold. The user side is the problem. The practical approaches map onto the survey's categories: content features (game metadata matched against whatever is known at registration), domain information (behaviour transferred from similar players or a sister brand) and, since 2024, language-model world knowledge used to connect a thin signal such as an acquisition channel to a plausible first screen. Adkuu's earlier post on the cold-start problem in iGaming describes the layered method. This post's point is narrower: when a vendor claims a lift figure, ask what share of it came from players in their first session, because that is where most systems are weakest and where most churn happens.
Explainability: from feature to expectation
Three regulatory threads converged in 2025 and 2026. None says "explainable AI". All require an operator to show why a player was treated the way they were.
Great Britain. In his keynote to the International Association of Gaming Regulators on 20 October 2025, published by the Gambling Commission on 22 October 2025, chief executive Andrew Rhodes said: "One of the other things that I think we need to watch out for is the general trend towards hyper personalisation." He accepted that serving more meaningful content raises engagement, then warned of the risk that players feel "an increased intensity that nobody necessarily intended". He also observed that "in the last 12 months or so" operators were "increasingly now using generative AI" to make customer interactions more consistent. The Commission's page "The Commission's approach to Artificial Intelligence", last updated 5 January 2026, states a "low and/or minimal risk appetite in relation to the use of AI" and lists "Understanding and regulating (where appropriate) the use of AI by the gambling industry" among its aims. On offers specifically, the Commission's consultation response of 1 May 2024 (updated 27 August 2024) confirmed licence condition 5.1.12, in force from 1 May 2025, under which remote licensees must give customers "options to opt-in to the product type they are interested in" for direct marketing. A model that generates an offer must respect that consent before it ranks anything.
The Netherlands. The Dutch government's announcement of 21 December 2023 set out that operators must contact a player who wants a deposit limit above 350 euros a month (150 euros for young adults), must actively check affordability when a player deposits more than 700 euros (350 euros for young adults) to their player account, must notify players about their limits and session duration every 30 minutes, must log them out when a limit is reached, and must monitor players continuously in order to intervene in time ("spelers voortdurend gaan monitoren om tijdig in te kunnen grijpen"). In a speech published by the Kansspelautoriteit, René Jansen said on 6 June 2024: "The real-time monitoring obligation has also been introduced and requires operators to intervene within an hour in the event of potentially excessive gambling behaviour." The regulator enforces it: a binding instruction dated 13 November 2025 against Hillside New Media Malta Plc, the company behind bet365, stated that an important part of the duty of care is actively monitoring the play of every player ("het actief monitoren van het speelgedrag van elke speler"), with four weeks to comply (CasinoBeats, 23 January 2026).
Sweden. The Swedish Gambling Act imposes a duty of care to counter excessive gambling at the individual level ("motverka överdrivet spelande på individnivå"); since 1 October 2024 Spelinspektionen has had the power to set specific requirements for the content of licensees' action plans (Ministry of Finance, 29 August 2024). In judgments of 27 May 2024, reported by Realtid on 28 May 2024, the Administrative Court of Appeal in Jönköping sided with Spelinspektionen in cases against five operators and found that they had breached the duty of care by offering bonuses to players who showed clear signs of risky gambling ("tydliga tecken på riskfyllt spelande"), according to Realtid. A personalised bonus engine that cannot show which risk signals it checked is exposed to exactly that finding.
The European Union. The AI Act entered into force on 1 August 2024 and became applicable on 2 August 2026 with exceptions; the AI Omnibus that entered into force on 27 July 2026 moved the obligations for Annex III high-risk use cases to 2 December 2027 (European Commission, page updated 3 August 2026). The Commission's page does not mention gambling, and Adkuu does not read a lobby recommender as a listed high-risk system. The relevant principle is the Act's transparency logic, under which people "should be made aware that they are interacting with a machine".
What Adkuu did not find matters too. No document opened for this post requires an operator to explain an AI-generated offer to the player in those words. The requirement is indirect: consent by product, continuous monitoring, intervention within a fixed time, and courts asking whether a bonus went to someone it should not have. An operator can answer those questions only if, for every recommendation or offer, it can retrieve the inputs, the gating rules and the risk flags that were checked. That is what explainability means operationally, and it is why a black-box ranking model has become a licensing liability rather than a technical preference. Adkuu's answer page on explainable AI in gambling sets out the definitions.
A checklist for evaluating a personalization vendor
| Area | What to ask | What a good answer looks like |
|---|---|---|
| Data needed | Which events and fields are required on day one, which are optional, and whether any personal data leaves the operator | Works from game metadata plus session events; an opaque player id, no personal data; a documented event schema |
| Time-to-value | When the first measurable change appears and how it will be measured | A dated plan: integration, holdout, first read-out; the vendor insists on a holdout from the first day |
| Cold-start method | What a player registered 40 seconds ago sees, and from which signals | A named method: content features, acquisition context, declared preference, first-session learning; not "the model learns" |
| Explainability | Whether the system can state why item X was shown to player Y, in a form a compliance officer can read | Per-decision reasons stored and retrievable; written rules for how risk flags override ranking |
| Measurement design | Holdout or A/B, duration, metric, and who owns the analysis | An operator-owned holdout, at least one full promotional cycle, a pre-registered metric, incremental GGR separated from redistributed turnover |
| RG safeguards | How limits, self-exclusion, affordability flags and marketing consent gate the model | Hard gates upstream of ranking; consent by product honoured; an audit log of every suppressed recommendation |
| Latency and coverage | Response time at peak and the share of the catalogue with usable metadata | Both stated and testable in a sandbox before contract |
| Exit terms | Who owns models, features and logs at termination | The operator keeps its data and logs; export is documented |
Adkuu's answer page on choosing an iGaming personalization vendor expands each row.
Where AI Sphere sits
AI Sphere, by Adkuu, is built around the two gaps above. It personalizes from the first click rather than waiting for a history, which is the cold-start case; it draws on 20,000+ games with rich metadata, so the item side is never cold; it returns explainable recommendations, so each recommendation carries its reason; and it responds in under 100 milliseconds, integrated through an SDK. Details are on the personalized lobby page. Whether any personalization system is acceptable in a given market is a question for the operator's compliance team; the rules above are the ones to read first.
FAQ
How much lift does casino personalization deliver? The only 2026 figure with a disclosed control that Adkuu could open is a 2.2 percentage-point uplift in first-time-deposit conversion over a two-month A/B test at one operator. Vendor-published figures run from "up to 20%" of turnover redirected to a 35% turnover uplift, without a described control. Treat anything above single digits as a claim to be tested on your own player base.
What is the cold-start problem in a casino lobby? It is the task of recommending well to a player with no interaction history, defined in the January 2025 survey by Zhang and co-authors as "accurately modeling new or interaction-limited users or items". In casinos the user side is the hard side; game metadata usually makes the item side warm.
Do regulators require explainable AI in gambling? Not in those words, as of October 2026, in the documents Adkuu opened. They require consent by product (Great Britain, from 1 May 2025), continuous monitoring with intervention within an hour (the Netherlands) and a duty of care that courts have applied to bonus targeting (Sweden, May 2024). Meeting those requires per-decision records, which is explainability in practice.
How should an operator measure a personalization vendor? With an operator-owned holdout, a pre-registered metric, at least one full promotional cycle, and a split between incremental GGR and redistributed turnover. Ask for first-session results separately, because that is where cold-start methods are tested.
Does personalization conflict with responsible gambling rules? It can, if the model optimises intensity. The Gambling Commission's chief executive warned on 20 October 2025 of "an increased intensity that nobody necessarily intended". The fix is architectural: risk flags, limits and consent sit upstream of the ranking model, not after it.
Sources
- IAGR 2025 Conference keynote - Andrew Rhodes speech — Gambling Commission
- The Commission's approach to Artificial Intelligence — Gambling Commission
- Summer 2023 consultation – Proposed changes to LCCP and RTS: Consultation Response — Gambling Commission
- Maatregelen voor bewuster speelgedrag en betere bescherming spelers van kansspelen — Rijksoverheid
- Speech René Jansen, 6 juni 2024 — Kansspelautoriteit
- Ett skarpare verktyg för att motverka överdrivet spelande — Regeringen
- AI Act | Shaping Europe's digital future — European Commission
- Cold-Start Recommendation towards the Era of Large Language Models (LLMs): A Comprehensive Survey and Roadmap — arXiv
- Seen.io and Xtremepush partner to bring AI-powered personalised video into iGaming CRM journeys — iGB
- Vegangster integrates ZingBrain AI for smart game recommendations across operator lobbies — iGaming Future
Last verified: October 10, 2026.