The unbrowsable catalogue
An online casino carries thousands of games. A sportsbook carries thousands of markets across hundreds of events. Neither can be browsed meaningfully, particularly on a phone.
The consequence is that most of the catalogue is effectively invisible, and play concentrates heavily in whatever appears in the first screens. This makes discovery the primary product lever in casino, and a significant one in sportsbook.
It also makes it commercially contested, because what appears in prominent positions determines what earns, and suppliers know this.
Merchandising and its entanglements
Lobby design in gambling is not purely an editorial exercise. As covered in the iGaming Basics course, studios negotiate for placement, trade exclusivity windows for visibility, and sometimes pay for prominence directly.
This means the people designing discovery are working within commercial arrangements they may not control. A featured row may be contractually committed. A new release may require prominence for a defined period regardless of how it performs.
The practical implication for product work is to be clear about which surfaces are commercially committed and which are genuinely available for optimisation, and to argue for preserving enough of the latter that discovery can actually serve customers. An operator whose entire lobby is allocated commercially has no ability to improve outcomes, and the resulting experience is worse for everyone including the suppliers, because customers who cannot find games they enjoy play less overall.
Structuring the catalogue
Before personalisation, the basic structure has to work.
Categorisation needs to reflect how customers actually think about games rather than how suppliers classify them. Mechanic, theme, volatility, feature type and familiarity all matter differently to different customers, and a single taxonomy will not serve all of them.
Search is used more than product teams expect and is frequently poor. Customers search for partial names, misremembered names, supplier names and mechanics. Search that requires exact matching fails most of these.
Filtering allows customers to narrow by the attributes they care about, and depends on the underlying metadata being complete, which it often is not because it arrives from suppliers in inconsistent form.
Recently played and favourites serve returning customers, who are the majority of sessions and are frequently designed for last.
Curated collections created editorially, around themes, mechanics or occasions, give customers an entry point that is neither browsing everything nor accepting an algorithm's choice.
The metadata point deserves emphasis. Good discovery depends on knowing what each game actually is, and that information arrives from many suppliers in inconsistent formats with variable completeness. Operators that invest in normalising and enriching game metadata can build discovery experiences that operators without it simply cannot, and this is one of the less visible sources of product differentiation in this sector.
Recommendation approaches
Popularity ranking is the baseline and works better than it should, because popular games are popular for reasons. Its weakness is that it entrenches, since the games that are surfaced become popular and remain surfaced.
Content-based recommendation uses game attributes, suggesting titles similar to those a customer has played. It handles new games well, since attributes are known immediately, and produces narrow recommendations that can trap customers in a small region of the catalogue.
Collaborative filtering uses the behaviour of similar customers, which surfaces less obvious connections and handles the breadth problem better. It requires substantial data and struggles with new games and new customers.
Hybrid approaches combine these, which is what most mature implementations do.
The cold start problem, recommending to a customer about whom nothing is known, is handled through popularity, through any information gathered at registration, and through learning quickly within the first session. This matters commercially because the first session substantially determines whether a customer returns.
Where personalisation gets dangerous
Here the sector diverges from general product practice, and the divergence needs stating explicitly rather than being handled with a caveat.
A recommendation system optimised for engagement will find what increases play. Applied without constraint in gambling, it will learn several things that are effective and undesirable.
It will learn that higher-volatility content produces more engagement from customers who respond to it, and will surface it to them preferentially.
It will learn that customers who are chasing losses respond to particular content and particular prompts, because that responsiveness is visible in the data as engagement.
It will learn that certain customers respond to promotional prompts at particular moments, including moments that correspond to distress.
It will learn to increase spend among the customers most responsive to increases, which is a population overlapping with those experiencing harm.
None of this requires the system to be badly built. It requires only that it is optimised for engagement and given enough data, because these patterns are genuinely present and genuinely effective.
The response is guardrails: constraints applied to the system that prevent outcomes optimisation alone would produce.
Exclude customers under safer gambling restriction from promotional recommendation entirely, and ensure that suppression is reliable across every surface, including recommendation rails, notifications and email.
Do not use risk indicators as engagement signals. If a model has access to features that identify customers displaying markers of harm, it must not be permitted to use them to increase engagement. This is worth checking explicitly, because a model given a wide feature set will find them without anyone intending it.
Constrain volatility escalation. A system should not systematically move customers towards higher-volatility content, which is a pattern worth monitoring for directly.
Monitor outcomes by segment, checking whether recommendation is increasing spend disproportionately among customers who were already high-spending, which is the signature of the problem.
Include harm outcomes in evaluation. A recommendation change should be assessed not only on engagement but on whether it increased spend among customers displaying risk indicators. A change that improves aggregate metrics through that mechanism should not be shipped.
The general principle, developed further in the next lesson, is that optimisation systems in this sector require constraints that reflect what the operator should not do, because the optimiser will otherwise find the most effective route regardless of what it is.
Data protection considerations
Personalisation involves processing behavioural data, and in this sector that data is unusually sensitive in practical terms.
Records indicating what a person gambles on, when, how much and how often reveal a great deal, and inferences drawn from them may include indications that the person is experiencing difficulty. Handling this requires a lawful basis, transparency about what is being done, appropriate retention limits and security proportionate to the sensitivity.
There is also a specific caution about inferred characteristics. A system inferring that a customer is likely experiencing gambling problems has generated information about their health circumstances, and that inference has obligations attached regardless of whether it was sought. The appropriate use of such an inference is a safer gambling intervention, not a marketing input, and systems should be built so that it cannot become the latter.
Measuring discovery
Discovery measurement usually stops at click-through, which shows that a recommendation was noticed and nothing else.
Click-through indicates attention.
Session quality following the click indicates whether the recommendation was any good: did the customer play, for how long, did they return to that game.
Catalogue breadth indicates whether discovery is exposing customers to the catalogue or trapping them in a narrow region, and a system that maximises immediate engagement typically narrows.
Return rate indicates whether discovery contributed to the customer coming back.
Retention over months is the outcome that matters and is slow to observe, which is why proximate metrics dominate and mislead.
Spend consistency indicates whether recommendation moved customers outside their established pattern, which is the harm-relevant measure and is rarely tracked.
An honest discovery function measures the last three alongside the first, and accepts that a change improving click-through while reducing catalogue breadth and increasing spend variance is not an improvement.
What good looks like
Discovery done well in this sector has recognisable characteristics. Customers can find games they have played before quickly. Search works with imperfect input. Categorisation reflects how customers think. New customers reach something enjoyable within their first session. Returning customers see a lobby that reflects them without being narrow. Commercial placement exists without consuming every surface. And the personalisation system is constrained so that it cannot pursue engagement into places the operator should not go.
That last item is the one that distinguishes this sector, and it is the subject of the next lesson.
Sportsbook discovery
The lesson has focused on casino, where the catalogue problem is most acute. Sportsbook has its own version, structured differently.
The volume is comparable and organised around events rather than items. A major football weekend presents hundreds of fixtures, each with dozens or hundreds of markets, and the customer needs to reach the specific market they want quickly.
The design responses differ from casino. Sport and competition navigation provides the primary structure, since customers generally know which sport and often which fixture they want. Live and upcoming surfacing matters enormously, since time sensitivity is inherent. Bet builder entry points need to be reachable from the fixture rather than requiring separate navigation. Search must handle team names, player names and market types. And personalisation here means surfacing the sports, competitions and teams a customer follows, which is more tractable than casino recommendation because the signal is clearer.
The bet slip is a distinctive element with no casino equivalent, and it carries considerable design weight. It must support building complex combinations, show the effect of each addition on the price and potential return, handle price movements between selection and placement, and make the stake and potential return unambiguous. Bet slips that obscure the actual stake, or that present potential returns more prominently than the amount being risked, are a recognised concern and have attracted regulatory attention in some markets.
The general point is that discovery in sportsbook is a navigation and timeliness problem where casino is a curation and recommendation problem, and applying casino thinking to sportsbook produces a lobby of featured markets that nobody wanted.
Building the capability
A practical note on what personalisation actually requires, since the ambition frequently exceeds the foundation.
Metadata. Complete, normalised attributes for every item in the catalogue. Arriving from many suppliers in inconsistent form, this requires deliberate enrichment work, and without it content-based approaches cannot function.
Event data. Detailed behavioural records at sufficient granularity, retained long enough to model, and joined correctly to customer identity across devices and sessions.
Identity resolution. The same customer recognised across devices and channels, without which the behavioural picture is fragmented.
Serving infrastructure. The ability to compute and deliver recommendations within the latency a lobby load allows, which constrains model complexity.
Evaluation framework. The means to test recommendation changes properly, including the harm-relevant measures described above.
Governance. Defined constraints on what features a model may use and what outcomes it may optimise for, with someone accountable for them.
Most operators attempting personalisation discover the constraint is the first three rather than the modelling. An organisation with excellent data science and poor metadata will produce recommendations no better than popularity ranking, and the unglamorous foundational work is where the return actually is.
The related caution is against over-personalising. A lobby that shows a returning customer only what they have played before is convenient and traps them, reducing catalogue exposure and eventually engagement. Deliberate exploration, surfacing content outside the established pattern, is necessary and is penalised by any evaluation framework measuring only immediate response, which is why it needs to be a designed constraint rather than an emergent behaviour.
A closing principle
Discovery in this sector sits at an unusual intersection. It is a genuine customer service, because a customer who cannot find something they enjoy among thousands of options has a worse experience and plays less. It is a commercial instrument, because placement determines revenue and is negotiated with suppliers. And it is a harm-relevant system, because what is surfaced to whom affects spend among people for whom spend is consequential.
Those three roles are frequently held by different teams with different objectives, and the friction between them is where discovery quality is decided.
The principle worth holding is that the customer service role should constrain the other two rather than being subordinate to them. A lobby optimised entirely for commercial placement serves nobody well, including the suppliers who paid for the placement, because customers who cannot find content they enjoy play less overall. A recommendation system optimised entirely for engagement pursues outcomes the operator should not want.
A discovery function that starts from what would actually help this customer find something they enjoy, and then accommodates commercial and optimisation objectives within that, produces better outcomes on every dimension including revenue. That ordering is easy to state, difficult to maintain under commercial pressure, and is what distinguishes the operators whose lobbies work from those whose lobbies are a negotiated allocation of screen space.
Practical starting points
For an operator whose discovery is essentially a supplier-ordered game list, the sequence that produces improvement fastest.
Fix the metadata. Normalise game attributes across suppliers and fill the gaps. Everything else depends on it.
Make search work. Partial matching, tolerance for misspelling, supplier and mechanic search. This is used more than most teams assume and is frequently the worst part of a lobby.
Surface recently played and favourites prominently. Returning customers are the majority of sessions and are frequently designed for last.
Build simple popularity and category rails before attempting personalisation, since they establish the surfaces and the measurement.
Instrument properly. Impressions, clicks and downstream session outcomes by rail and position, without which no improvement can be evaluated.
Then personalise, starting with straightforward approaches and adding sophistication as the data and evaluation framework support it.
Set the guardrails before the model ships, not afterwards, because retrofitting constraints to a system already optimised without them is considerably harder than building with them.
Most operators attempting this start at the personalisation step and discover the first two were the constraint. The unglamorous order is faster.
Explaining recommendations
One design question worth raising because it is becoming a regulatory as well as a product matter.
Customers generally do not know why particular content is being shown to them. A rail labelled as recommended could reflect their behaviour, a commercial arrangement, a popularity ranking or an algorithmic prediction, and the label conveys none of that.
Several arguments favour greater transparency. Consumer protection frameworks in a number of jurisdictions increasingly require disclosure where content ordering is influenced by commercial arrangements, which applies directly to paid placement in a lobby. Data protection frameworks give individuals rights concerning automated processing that affects them. And there is a straightforward fairness argument that a customer choosing what to play is entitled to know whether the ordering serves them or someone else.
The design responses are modest. Labelling commercially sponsored placement as such. Distinguishing personalised rails from editorial ones. Explaining, in general terms, what a recommendation is based on. Allowing customers to see and adjust what the system believes about their preferences.
None of these is technically difficult. All of them reduce the effectiveness of the surfaces they apply to, which is why they are uncommon. The direction of regulatory travel suggests operators will be doing some of this regardless, and doing it deliberately produces a better result than doing it under instruction.