Why 2026 is the Year Sportsbooks Move to Session-Aware Architectures

While market leaders are shifting to "Netflix-style" personalization, most sportsbooks still rely on static, popularity-based odds that fail to capture real-time intent. This engineering deep dive explores why rules-based engines break down during live events and outlines a modern architecture, using collaborative filtering and sequential prediction, to build session-aware discovery that compounds handle per user.

January 9, 2026
 | 
5
 min read
by 
Nic Scheltema

It is early 2026. For years, heavyweights like DraftKings and FanDuel have optimized their funnels to perfection, yet their homepage discovery remains conservative.

Most sportsbooks are still using popularity-based odds with almost no user context.

They rely on hard-coded rules (Time of Day, Seasonality) and generic "Featured" carousels. Teams generally accept that this approach gets them 60–70% of the way there. It avoids bad outcomes and handles the "happy path" well. But the remaining 30%, the long-tail discovery, the cross-sport upsells, and the in-play retention, is where the value leaks.

The industry has realized that User Acquisition is expensive, but Churn is free. The next battleground isn’t just offering more markets; it’s Session-Aware Discovery.

Here is the engineering deep dive on why traditional rules engines are failing to close that 30% gap, and how you can build a real-time sports betting personalization engine using Shaped.

The Current Friction: Where Rules Break Down

Most betting and fantasy apps are still early on personalization, even if they talk about it constantly on earnings calls. The dominant pattern is a disconnect between the user's history and their current intent.

1. The "Saturday vs. Tuesday" Problem

A user’s intent shifts wildly based on context.

  • The Scenario: An NFL bettor opens the app on a Saturday.
  • The Rule: "Show popular leagues." Global popularity pushes NBA or Premier League to the top.
  • The Reality: The user is looking for College Football because they are watching GameDay.
  • The Failure: Rules struggle to decide when short-term context (Saturday afternoon) should outweigh long-term history (NFL bias).

2. The "Ad-Click" Disconnect

This is rampant in DFS (PrizePicks, Underdog).

  • The Scenario: A user clicks an Instagram ad for a specific UFC fighter prop.
  • The Rule: Redirect to the "MMA Lobby."
  • The Reality: The user lands on a generic list of fights. They have to scroll to find the specific prop that brought them there.
  • The Failure: The session is treated as a "new session" rather than a continuation of the ad intent.

The "Data Readiness" Trap

Why haven't teams solved this yet? It’s rarely a lack of model sophistication. It is Data Readiness.

Across the board, we see Product teams that don't trust their analytics tools, and ML engineers spending 90% of their time on plumbing, cleaning inconsistent event schemas and reconciling bet_placed events, rather than training models.

Shaped.ai was built to bypass this bottleneck. We connect directly to your raw data streams (Segment, Amplitude, Snowflake), handle the schema unification, and provide the architecture out of the box.

The Emerging Solution: Context-First Architectures

Leading engineering teams are moving toward Session-Aware Ranking. In the last 12 months, three distinct patterns have emerged from the market leaders:

1. Fanatics Betting: The "FanGraph" (Unified Identity)
Fanatics feeds commerce data (Jersey purchases) into their betting model to signal high intent for specific entities (e.g., Shohei Ohtani) even if the user has never placed a bet.

  • The lesson: Identity graphs beat generic popularity.

2. ESPN Bet: The "FanCenter" (Fantasy-to-Bet Context)
ESPN links Fantasy rosters to Betting slips. When a user logs in on Sunday, the model prioritizes props for their fantasy players, overriding historical team biases.

  • The lesson: Short-term context (Fantasy Lineup) > Long-term history.

3. Polymarket: Velocity-Based Ranking
Polymarket uses content-based embeddings (vectorizing news headlines) to map real-time trends to markets instantly, allowing them to recommend markets before they have significant user interaction data.

  • The lesson: In fast-moving markets, semantic relevance beats collaborative filtering.

How to Build This (Without the 50-Person Team)

You need an architecture that handles Hybrid Approaches: using models for ranking within the guardrails of business rules. Here is how you execute this with Shaped.

1. Solving Context with Collaborative Filtering (Two-Tower)

To handle the "Saturday vs. Tuesday" user, you cannot rely on a static user profile. Shaped uses a trained collaborative filtering model like Two-Tower:

  • User Tower: Encodes history plus real-time context (e.g., source: instagram_ad, time_of_week: Saturday_Afternoon).
  • Item Tower: Encodes the metadata of the bet (e.g., "League: NCAAF", "Type: Player Prop").

Because inference happens in real-time, the model adapts instantly. If the user comes from a UFC ad, the User Vector shifts, and the ranking engine promotes UFC props immediately, even if that user usually bets on the NFL.

2. Mastering the Baseline with ShapedQL (The Hybrid Approach)

One of the biggest blockers for teams in 2026 is defining "What is Good?" Many teams ship a homepage without knowing if "Trending" actually lifts conversion over "Popular."

Shaped allows you to experiment safely using ShapedQL, a query language that lets you mix Neural Scores with Hard Logic. This is the "Hybrid" approach that Product and Compliance teams love:

‍

SELECT *
-- start with a pool of 200 personalized bets
FROM similarity(
embedding_ref = 'als',
limit = 200,
encoder = 'interaction_round_robin',
input_user_id = '$user_id'
),
-- enrich the pool with a few promoted bets
filter(where='promoted=true', name='promoted_bets', limit=50)

-- score results by predicted ctr and trendiness
ORDER BY score(
expression="0.7 * predicted_ctr * 0.3 log(bet_velocity_1h)"
) desc

-- Risk: filter out results with low odds
WHERE odds > -500

-- Compliance: never show high liability bets for this user ID
AND prebuilt(
'high_liability_bets_for_user',
input_user_id="$user_id"
)

-- Promotion: Ensure 5% of the results are promoted
REORDER BY boosted(retriever='promoted_bets', strength=0.05)

‍

This allows you to establish a clear baseline. You can A/B test a query that is 100% Popularity against a query that is 50/50 Personalization, giving you definitive proof of lift.

3. Handling "In-Play" Velocity with Sequential Prediction

For markets that decay quickly (In-Play, Prediction Markets), Shaped implements a sequential prediction model like SASRec.

  • Instead of looking at "All bets ever placed," SASRec looks at the sequence of user actions.
  • If a user views "Lakers vs. Warriors" -> Clicks "Live Odds" -> Bets "Lakers ML", the model detects the sequence intent and immediately serves correlated props (e.g., "LeBron James Next Point").

Conclusion: Differentation is Open for the Taking

The competitive takeaway for 2026 is simple: There is very little true differentiation right now. Most apps feel the same once you get past the branding.

The teams that solve the "Saturday vs. Tuesday" problem and operationalize Session-Aware Ranking will compound gains in handle per user, while the rest fight over expensive acquisition costs.

Shaped gives you the infrastructure to beat the baseline today.

Ready to move beyond the static odds board?

Get up and running with one engineer in one sprint

Guaranteed lift within your first 30 days or your money back

Talk to us
100M+
Users and items
1000+
Queries per second
1B+
Requests

Related Posts

Nic Scheltema
 | 
August 19, 2025

The 10 Best Semantic Search APIs in 2025

Tullie Murrell
 | 
November 19, 2024

Shaped Launches Semantic Search with Behavioral Re-ranking

Nic Scheltema
 | 
September 18, 2025

The PM's Guide to Personalization: A 'Build vs. Buy' Framework for 2025