All White Papers & Case Studies
Doubling CTV Platform Engagement

What a Major Smart TV Platform Discovered When It Tested Watchworthy Against Its Own Personalization

2.1× CTR
vs. the platform’s production personalization
5× More Clicks
on personalized vs. trending content
60% of Clicks
home‑screen clicks, from one gallery
47% Engaged
of exposed users clicked a recommendation
Download the full white paper (PDF)Doubling CTV Platform Engagement · Watchworthy · PDF Contact Ranker to Discuss →
Watchworthy, Ranker’s enterprise personalization solution, powered recommendations on a major Smart TV platform to test its core premise: driving higher user engagement. This large‑scale test ran live in production with real users as a randomized, in‑production comparison, pitting Watchworthy directly against the platform’s native personalization on its own turf. The results were validated across two separate live tests under materially different seasonal and audience conditions. These are the findings.
Table of Contents
01 Key Findings 02 Personalization Drove 5× More Engagement 03 The Engagement Flywheel 04 Disambiguating Shared Devices 05 Background: Why Watchworthy? 06 Methodology 07 Unlocking More Gains 08 What This Means for CTV Platforms 09 Working With Ranker 10 Next Steps 11 Frequently Asked Questions
01Key Findings

Key Findings

Watchworthy increased the recommendation click‑rate over the platform’s native personalization in both tests, doubling it in the second. It continued to optimize and recalibrate while each test was live, and that optimization unlocked further gains.

A particularly notable finding was the degree to which the Watchworthy recommendation surface overperformed its footprint: the recommendation gallery held just 15 placements, only four of which were visible without scrolling. Yet this single gallery powered by Watchworthy accounted for more than half of every click on the entire home screen by the end of Test 1 and continued growing to over 60% before Test 2 concluded.

Crucially, these results were achieved with only a partial implementation of Watchworthy’s capabilities. The test deployed a small set of recommendations per user in a single gallery and omitted Watchworthy’s more advanced personalization stack capabilities like full slate optimization and context‑aware re‑ranking. Even under these operational constraints, Watchworthy outperformed the TV platform’s own production personalization.

The same personalization powering this gallery can be seamlessly applied to the rails, collections, promoted placements, and ad‑supported inventory already on the page. The return on a single gallery is a strong indicator of the engagement the rest of the screen is currently leaving unclaimed.

These results validate that there is a quantifiable opportunity for CTV platforms to optimize their user experience and achieve impactful performance gains leveraging Watchworthy’s personalization solution.

CTR Lift
1.3×
Test 1
2.1×
Test 2
User Recommendation Engagement
30%
Test 1
47%
Test 2

Test Data Summary

MetricWatchworthy · Test 1Control · Test 1Watchworthy · Test 2Control · Test 2
Daily CTR – Optimized1.5%1.1%1.3%0.7%
Peak daily CTR2.0%1.3%2.0%0.8%
CTR Lift1.3×—2.1×—
Recommendation Engagement30.0%—47.0%—
Recommendation Share of Home Activity36.4%—61.3%—
02Personalization Drove 5× More Engagement

Personalization Drove 5× More Engagement

Viewers strongly favor — and actively engage with — content that directly aligns with their personal taste profile. To measure this, we benchmarked Watchworthy’s personalized recommendations against non‑personalized, trending titles — content that was currently popular and topical, but not specifically personalized with individual preferences.

Users clicked on Watchworthy personalized recommendations 5× more often. This preference was consistent and validated across Test 1 (5.0×) and Test 2 (4.5×).

Users Strongly Prefer Personalized Content
Test 1
83.5%
16.5%
Test 2
81.7%
18.3%
Personalized Recommendations Trending (Non‑Personalized)

Personalization emerged as the primary driver of viewer action, and this preference was observed across all household viewing segments. Even users with a single historically recorded interaction clicked personalized recommendations 1.7× more often. The same preference for personalization shows up at the surface level, where Watchworthy’s single personalized gallery captured over 60% of all home‑screen clicks.

03The Engagement Flywheel

How Watchworthy Personalization Increased Engagement: The Flywheel

Personalization on a Smart TV platform is usually framed as a one‑way supply problem: insufficient user signal leads to low recommendation relevance. The data demonstrates that the feedback loop actually runs in both directions — relevance produces engagement, engagement generates higher‑quality signal, and that signal unlocks deeper relevance.

Engagement Rates Compound as User Signal Increases
% Users % Rec Engagement Rate
48.8%
7%
1
21.7%
10%
2
10.4%
11.7%
3
9.3%
11.4%
4–5
5.9%
15.9%
6–10
1.8%
23.1%
11–20
0.4%
28.5%
21–50
<0.1%
59%
51+
Recommendation engagement rate: the share of users in each cohort who clicked at least one recommendation, across both tests. X‑axis: user interaction count on entering the test.

The distribution of user engagement signals follows a notable pattern that’s typical of many data‑sparse TV platforms. It also reveals how personalized recommendations can serve as a flywheel to drive higher levels of engagement across cohorts.

Across both tests, 7% of viewers who entered with a single recorded interaction clicked a recommendation. Among the most active viewers, 59% did — an eight‑fold spread in the recommendation engagement rate across the curve. The steepest step on that curve is the first: viewers with two prior interactions engaged at a nearly 50% higher rate than those with one, which puts the largest available gain in the largest cohort. Early relevance is what determines whether a user takes that second step, and it compounds from there into the retention and monetizable attention platforms are competing for.

Learning Without Onboarding: How to Solve the Smart TV Cold‑Start Problem

Half of all users in the partner’s randomized test groups had exactly one recorded interaction over the prior six months — a signal that could just as easily represent an accidental click as a genuine preference. This underscores the severe data sparsity CTV platforms face, and highlights the fragile foundation on which many native personalization engines attempt to build. Staking an entire personalized slate on a single interaction risks overfitting, while serving generic slates risks losing their attention on day one.

Watchworthy managed signal confidence across the user base — from new users with zero‑data to deeply engaged power users — by deploying adaptive learning strategies engineered to calibrate taste profiles and graduate viewers into higher‑yield cohorts:

Adaptive Taste Probing
For low‑confidence profiles, the engine strategically weaves exploratory candidate titles into the recommendation slate. This process is methodical and heavily optimized for signal clarity. Candidate titles are selected as pivotal taste‑cluster nodes within Watchworthy’s Taste Graph to efficiently measure and refine user preference without degrading overall slate quality.
Confidence‑Weighted Personalization
As signal confidence increases through confirmed interactions, adaptive probing automatically scales back, allowing deep, high‑precision personalization to take over the slate.
Powering the Engagement Flywheel
By resolving the cold‑start problem, Watchworthy efficiently converts passive, low‑signal viewers into active users with increased engagement.
04Disambiguating Shared Devices

How Watchworthy Disambiguates Multiple Viewers on Shared Devices

Smart TVs are often shared family devices, with distinctly different preferences among household members. Yet platform behavioral data inevitably collapses household interactions into a single profile, causing recommendations to degrade into a diluted average that serves no one (e.g., children’s programming placed beside prestige drama and sports). While most product teams recognize this problem, few have a tractable way to separate the personas because the behavioral signal that would distinguish them is exactly what a shared device obscures.

This is where Watchworthy has the advantage of being trained on rich psychographic viewer preference data over behavioral data. Because unique clusters are identifiable independently of who is holding the remote, distinct taste modes can be detected inside a single household stream and served separately rather than blended.

Household Preferences Vary by Daypart
% Distribution of Viewing by Daypart
Kids Segment Adults Segment
12.8%
5.2%
Early AM
6a–9a
36.3%
28.2%
Daytime
9a–4p
22.7%
22.4%
Early Fringe
4p–7p
19.1%
22.9%
Primetime
7p–10p
4.6%
6.5%
Late Fringe
10p–11p
2.2%
8.3%
Late Night
11p–2a
1.5%
5.6%
Overnight
2a–6a

The tests probed this capability in an unplanned way. The partner’s Test 2 randomized sample had an increased skew toward children’s content — 33% of consumption against 16% in Test 1. Because Watchworthy acts as an intelligence layer on top of raw behavioral data, this skew was identified early via the household segmentation instrumentation. The slate composition was dynamically rebalanced in response without requiring a full model retrain, and Watchworthy’s adaptive learning capabilities calibrated these profiles.

The partner’s own data shows how cleanly these audiences separate. Half of all children’s viewing concentrates in the morning and daytime, and after primetime it all but disappears. Nearly two‑thirds of adult viewing takes place in the evenings and later. This is a stable, learnable pattern sitting unexploited in the data the platform already collects. Layering dayparting logic allows platforms to automatically align recommendations with the active viewer. Watchworthy has these capabilities to disambiguate household taste and optimize the slate contextualized around temporal patterns like time‑of‑day.

05Background: Why Watchworthy?

Background: Why Watchworthy?

This Smart TV platform approached Ranker with the goal of improving their on‑platform engagement. Like many OEMs in the CTV space, they own the glass but struggle to own the experience. Viewers routinely bypass native platform discovery to jump straight into third‑party streaming apps. That engagement gap becomes a retention problem, which ultimately affects platform monetization.

The partner cited a principal challenge impacting the effectiveness of their current native personalization solution: the platform’s native interaction data was sparse, with large knowledge gaps about individual user taste preferences, and this was compounded by the quality of the data signal, as behavioral interaction data is largely inferential to actual user taste.

The partner had already diagnosed that their feedback loop was deadlocked: to drive higher engagement they needed better personalization. However, in order to improve personalization, they needed better engagement to train their models and realize improvement.

This platform’s situation is representative, rather than unusual. Ranker hears this across the TV and streaming ecosystem, and many CTV platforms are trying to crack this issue.

ChallengeConsequence
Sparse first‑party interaction dataToo little signal per viewer to model individual taste; collaborative filtering degrades
Behavioral signal is inferential, not declaredA click or app launch cannot distinguish genuine interest from a misclick or idle browsing
No history at device activationCold start hits precisely when the platform has one chance to establish itself as the discovery hub
Discovery leaks into third‑party appsViewers bypass native surfaces, moving engagement, retention, and monetizable inventory off‑platform
Shared‑device viewingOne profile blends several household members; output converges on an average that serves no one
Fragmented catalog identityThe same title exists as multiple entities across providers, splitting preference signal and hiding long‑tail inventory

Watchworthy is specifically designed and optimized for fast cold‑starts in lean, sparse, and zero‑signal environments, providing the necessary jumpstart to initiate engagement and accelerate the feedback loop with the partner’s own signals.

Ranker’s taste graph of over 1.5 billion explicit preference signals from millions of real consumers powers Watchworthy. Its deep psychographic signal helps identify audience‑title relationships that are difficult to infer from metadata, aggregated ratings, or behavioral data alone — particularly to bridge the gap with platform viewer data. This enables CTV platforms to rapidly pinpoint viewer preferences, and is particularly effective with cold‑starts and jumpstarting new viewer engagement.

06Methodology

Methodology

The test was deployed in production across randomly selected US devices, running for four weeks in August–September 2025 and again for five weeks in November–December 2025 to validate the results.

The study was fully anonymized; Watchworthy had no visibility into individual test participants or household composition. The only inputs provided to Watchworthy were a sparse set of interactions (clicks and searches) as an implicit signal of household preference. These signals were used exclusively to retrieve recommendation slates from the Watchworthy service.

The test variants shared identical placement, occupying a single rail within the grid on the Smart TV’s home screen to ensure test parity and validity.

Partner technical constraints limited the test to a bare‑minimum deployment footprint. While these choices enabled the partner to deploy tests faster within their platform, they came at the expense of maximum performance: personalization slate generation was restricted to once daily (rather than real‑time generation), and engagement metrics were delivered by the partner two days in arrears. This reporting lag delayed the feedback loop, slowing the adaptive learning that typically occurs in real time under standard deployments. Additionally, no partner data was used in Watchworthy model training, restricting the scope of typical deployment optimizations.

Even under these operational constraints and a bare‑minimum deployment footprint, Watchworthy outperformed the platform’s native personalization at the conclusion of both tests, and without leveraging its most powerful capabilities.

07Unlocking More Gains

Unlocking More Gains

Watchworthy’s complete platform encompasses a full suite of personalization capabilities designed to maximize CTV user engagement and retention. While the live test deployments on this Smart TV operated under a technically constrained footprint, deploying these additional capabilities may unlock significant future lift:

01

Preference Signals (Expanding Signal Quality & Depth)

Move beyond basic click history by incorporating recency‑weighted implicit interaction signals and explicit user thumbs/ratings. Watchworthy can also power an optional, low‑friction onboarding experience during device setup that immediately establishes high‑confidence taste profiles and eliminates the cold start on day one.

02

Slate Optimization (Curating for Household Realities)

Refine the balance of the recommendation grid to match specific household environments. This includes dynamically adjusting format blends (balancing movies vs. TV series based on user habits), filtering slates by active streaming platform subscriptions, and balancing slate composition to match the preferences of individual household members and their co‑viewing patterns.

03

Context‑Aware Re‑Ranking (Matching Viewer Mindset by Time)

Dynamically re‑orders recommendation slates on the fly based on dayparting (time‑of‑day) and day‑of‑week viewing patterns. By adjusting candidates to reflect natural household rhythms — such as prioritizing kids’ titles on weekend mornings and prestige dramas during weekday primetime — the platform consistently surfaces the right content at the moment of intent.

08What This Means for CTV Platforms

What This Means for CTV Platforms

Across two live tests, four months apart, on different audiences and in different seasons, the same pattern held: users engaged far more with personalized content, and engagement compounded as user signal increased. Five key learnings for platform teams:

01

Viewers Demand Relevance Over Popularity

Users engaged with personalized recommendations 5× more often than trending content. Delivering true taste alignment — rather than broad, platform‑wide popularity — is what captures and retains viewer attention.

02

Cold Start is Solvable Now

A Taste Graph built on explicit human preference achieved 2.1× the click‑rate of native personalization and reached a 2.0% peak CTR. The primary bottleneck to recommendation relevance is signal quality, not years of accumulated, noisy behavioral logs.

03

The Home Screen is Under‑Claimed

A single gallery of 15 placements (with only four visible) captured over 60% of home‑screen clicks when optimized by Watchworthy. Applying this same predictive signal to the rails, collections, and promoted inventory represents an immediate opportunity to recover lost engagement.

04

The Engagement Flywheel

Graduating passive users up the engagement curve is where long‑term retention and monetizable attention are won. Across both tests, the engagement rate scaled from 7% for single‑interaction users to 59% in the top cohort — an eight‑fold spread. The steepest step on that curve is the first, worth a nearly 50% higher engagement rate on its own. Most of a platform’s install base is standing on that step.

05

Performance is Operated

Active, continuous optimization yields far higher returns than static, “set‑and‑forget” deployments. Watchworthy recommendation engagement increased 53% after Test 1 calibration, while native personalization declined by 5% during this time. In both tests, Watchworthy continued to unlock new performance gains as it adapted to the platform trends and seasonal shifts.

09Working With Ranker

Working With Ranker

Every platform is different. So is every audience.

While today’s modern tech‑stacks make it easy to integrate personalization, every CTV platform is unique. Install bases differ in composition and taste. Guides differ in structure and emphasis. Catalogs differ in depth, licensing, and how titles are identified. A recommender deployed once and left alone runs on assumptions imported from somewhere else.

Ranker begins with the partner’s platform leads: objectives, solution specifications, integration points, delivery, measurement, and reporting. With this Smart TV platform, that produced a solution plan aligned to their specification and operational requirements before any code moved. This enabled the partner to deploy two consecutive tests in short order.

Most of the work happens before launch — and it happens in sequence, because each stage is what makes the next one possible.

Pre‑Launch Preparation

Catalog Mapping and Resolution

Ranker’s automated mapping resolved 242,055 titles across the platform’s addressable catalog, drawn from nearly 50 separate data providers, collapsing duplicate entities onto single identifiers. The accuracy of this step is critical, as everything downstream inherits its errors, and the failures are quiet ones.

Addressability
Popular and current titles are usually well mapped by platforms. Mid‑tail and long‑tail content often isn’t. Unmapped content is unreachable content — it introduces popularity bias, degrades relevance for viewers with specific tastes, and suppresses deep‑catalog discovery. Platforms license substantial catalogs and then never surface them to the viewers most likely to watch them.
Deduplication
Master catalogs are often assembled from disparate third‑party sources, frequently without a unifying identifier. Without one, a recommender treats identical titles as distinct — fragmenting the preference signal across copies. The duplication survives into model training, where multiple versions of the same title can be recommended to the same viewer simultaneously. In this partner’s catalog, A Minecraft Movie appeared as 20 separate entries, each with its own ID, none tied to a common identifier.

Ranker’s resolution system runs against a taste graph of more than 30 million entertainment entities and supports the industry’s identifier standards, including Gracenote TMS and IMDb. It also processes catalogs carrying no deterministic identifiers at all. Partners inherit this capability rather than building it — catalog resolution is otherwise slow, manual, and easy to get wrong.

Audience Analysis and Segmentation

Mapping is what makes interaction data interpretable. Once every interaction resolves to a known entity, behavioral data can be analyzed by content property rather than by opaque ID.

From there we run compositional analysis against the platform’s own audience: distributions, preferences, and structural biases specific to that install base — household disambiguation, taste clustering, viewing patterns, interactions per user, data sparsity, and signal strength. Platform interaction data describes a device, not a person. This is the step that makes device‑level data usable at the viewer level, and that separates genuine preference from artifacts of how the guide is laid out.

These outputs set the recommender’s initial operating parameters. Signal strength in particular varies widely between platforms and cannot be assumed. Having measured the distribution of interactions per user and the confidence levels it implied, we set exploration rates accordingly — weighting broader results more heavily where confidence was low in order to learn tastes faster, then adaptively tightening recommendation precision as confidence rose.

Watchworthy was configured to the Smart TV’s audience before it served a single recommendation.

Operational Telemetry: Continuous Measurement and Optimization

With mapping resolved and the pipelines running, building the analytical layer on top was straightforward. Daily dashboards enabled Ranker to track mapping health, catalog performance, recommendation engagement, and outlier detection across every segment.

Outlier detection carries more weight than it sounds. It converts a live deployment into a controlled system, so anomalies surface in days rather than at the end of a test.

Operating the Deployment

Ranker Managed Operations
OperationWhat Ranker Does
Performance MonitoringDaily
  • Monitor core vitals + KPIs
  • Detect demand spikes (new releases, trending, seasonal)
  • Investigate day‑to‑day performance swings
  • Identify titles driving engagement gains and declines
  • Verify content availability
  • Flag issues requiring intervention
Content, Data & AlgorithmDaily / Weekly
  • Check data freshness and monitor data pipelines
  • Review engagement metrics across user segments and content catalog
  • Identify missing, mis‑mapped, or unservable titles
  • Review recommendation outputs for coherence and coverage
  • Adjust business rules and model inputs as needed
Operational StewardshipWeekly / Monthly
  • Align on client requirements and success KPIs
  • Conduct Engineering and Data Science reviews
  • Plan enhancements and priorities with stakeholders
  • Manage releases, experiments, and enhancements
  • Monitor latency, delivery, and SLAs
  • Produce client‑facing performance summaries and insights
  • Own recommendation outcomes

This instrumentation is what makes optimization proactive rather than reactive. With this platform, the analytics produced the insights that guided Ranker’s in‑flight adjustments while the test was running.

Test 1 reveals the result of these optimizations. Watchworthy’s average daily click‑rate rose 53% between the calibration weeks and the final week. The lift came from diagnosing where recommendations were being lost between generation and platform display, then re‑optimizing the slate accordingly. By comparison, the platform’s native personalization, running concurrently on the same home screen across the same days, moved −5%.

Had everything risen together, the gain would read as seasonal. Only the actively managed surface moved, which locates the improvement in the work rather than the calendar.

Recommendation performance is operated, not installed.

Consider this concrete example of how active management maximizes results: when partner impression data was delayed and incomplete, it meant the platform’s primary success metric could not be observed directly. Rather than wait on reporting cycles, Ranker built a two‑stage model: the first stage estimates click‑rate from observable inputs, the second projects those inputs forward and applies the first. Validated against the days the platform did report figures, the modeled series tracked actual CTR at a correlation of 0.86, with the two means differing by 0.014 percentage points. This enabled Ranker to continue to push optimizations and measure performance even in blind spots.

Why Ranker Works This Way

Ranker builds and operates its own consumer entertainment products, including the Watchworthy application. Watchworthy exists in two forms: a consumer recommendation app with real users, and an enterprise personalization system for TV platforms.

The consumer product serves as the live testbed — engagement, onboarding, and retention optimizations are validated against real viewer behavior before they reach a partner deployment. This is a significant operational advantage in managing personalization systems: it enables experimentation and data‑backed decision‑making that drives continuous optimization.

10Next Steps

Next Steps

Watchworthy is an enterprise‑grade personalization solution that can fit in any CTV personalization stack. All aspects of the system are customizable to the specific goals of the platform, including configuration and delivery. It can operate as a preference data layer inside an existing recommender, a candidate‑generation and ranking service, a grounding layer for AI‑powered discovery, or a fully managed personalization system.

To discuss a test on your platform, request a data sample, or review integration options, contact Ranker at business.ranker.com.

Contact Ranker to Discuss →
11Frequently Asked Questions

Frequently Asked Questions

Common questions about personalizing recommendations on Smart TV and CTV platforms, and about the two live tests described above.

Results & Validation

Do personalized recommendations outperform trending or popular content on a TV home screen?
Yes. Across two live tests, viewers clicked personalized recommendations roughly 5× more often than non‑personalized trending titles — 5.0× in the first test and 4.5× in the second. The advantage held among low‑signal viewers as well: users with a single prior recorded interaction still clicked personalized recommendations 1.7× more often, so the result is not an artifact of trending content being served mainly to less engaged viewers.
How much engagement lift can personalization deliver on a Smart TV home screen?
In the second of two live tests on a major Smart TV platform, Watchworthy reached 2.1× the recommendation click‑rate of the platform's in‑production personalization, with a 2.0% peak daily click‑rate, and 47% of exposed viewers clicked a recommendation. A single gallery — 15 placements, only four of them visible without scrolling — accounted for more than 60% of all clicks on the entire home screen.
How do you know the lift came from the recommender and not from seasonality or placement?
Through the controls built into the test. Assignment was randomized across production devices; every variant occupied identical placement, a single rail within the same home‑screen grid; and the platform's own personalization ran concurrently, on the same home screen, across the same days. The clearest evidence is divergence between the two: in the first test, Watchworthy's daily click‑rate rose 53% between the calibration weeks and the final week while the concurrently running native personalization moved −5%. Had the gain been seasonal, both surfaces would have risen together. The results were then validated in a second test under materially different seasonal and audience conditions.

Personalization & Cold Start

How can a TV platform personalize recommendations when it has almost no interaction data per viewer?
By bringing preference signal in from outside the platform. In two live tests on a major Smart TV platform, half of all users had exactly one recorded interaction over the prior six months — far too little to model individual taste from behavior alone. Watchworthy supplied external taste signal drawn from Ranker's graph of more than 1.5 billion explicit consumer preferences, and reached 2.1× the recommendation click‑rate of the platform's own in‑production personalization in the second test.
What is the cold‑start problem on a Smart TV, and can it be solved without an onboarding flow?
Cold start is the condition where a platform has no viewing history for a device or viewer, leaving its recommender nothing to personalize from. It can be solved without onboarding: because an explicit preference graph already encodes how tastes cluster across millions of consumers, a new viewer's earliest interactions can be matched against known clusters rather than waiting for behavioral history to accumulate. In the tests described above, viewers with only a single prior recorded interaction still clicked personalized recommendations 1.7× more often than trending titles, with no onboarding step of any kind. Onboarding remains available and highly effective where a platform wants it — in Watchworthy's own consumer application, a low‑friction taste‑capture flow has users watchlisting an average of 4.7 shows within three minutes of starting a session, as described in the Watchworthy technical whitepaper. The distinction that matters is that onboarding is an accelerant, not a prerequisite.
How do you personalize a shared device that several household members use?
Behavioral data collapses several household members into a single profile, producing a blended average that serves none of them well. Watchworthy makes those members separable by grounding the platform's behavioral stream against an explicit preference graph: because taste clusters are identifiable independently of who is holding the remote, distinct taste modes can be detected psychographically inside a single household stream rather than averaged together. Time of day adds a second axis: in testing, half of children's viewing concentrated in the morning and daytime and all but disappeared after primetime, while nearly two‑thirds of adult viewing fell in the evening and later.
What is dayparting, and how does it improve TV recommendations?
Dayparting re‑ranks recommendations by time of day and day of week to match the household member most likely to be watching at that hour. Household viewing follows stable, learnable daily patterns: in testing, half of children's viewing concentrated in the morning and daytime and all but disappeared after primetime, while nearly two‑thirds of adult viewing fell in the evening and later. Aligning slate composition to those rhythms surfaces relevant content at the moment of intent, using patterns already present in data the platform collects today.

Data, Taste Graph & About Watchworthy

What is a taste graph, and how does it differ from behavioral data?
A taste graph maps relationships between audiences and titles using declared preferences rather than inferred ones. Behavioral data records that a viewer clicked something, but cannot distinguish genuine interest from a misclick or idle browsing. Ranker's taste graph holds more than 1.5 billion explicit preference signals across more than 30 million entertainment entities, which lets a platform identify audience‑title relationships that metadata, aggregated ratings, and behavioral logs do not reveal on their own.
Where does Ranker's preference data come from, and why is it difficult to replicate?
It comes from real consumers stating what they like — ranking and voting on titles across Ranker's own consumer entertainment properties, where the graph has been accumulating since the company was founded in 2009. Declared, rather than licensed or inferred. That provenance is what makes it hard to reproduce. Behavioral logs can be collected quickly by anyone with traffic, but declared preference at scale requires an audience with a reason to express taste in the first place — which is why a platform's own logs describe what viewers did without ever describing what they like. The resulting graph spans more than 1.5 billion explicit preference signals across more than 30 million entertainment entities, and it is the asset a platform borrows to cover the viewers its own data cannot yet describe. The Watchworthy technical whitepaper covers the data model in more depth.
Why does catalog mapping matter for recommendation quality?
Because unmapped and duplicated titles are effectively invisible to a recommender. Master catalogs assembled from many providers often lack a unifying identifier, so identical titles are treated as distinct and preference signal fragments across copies. In one platform catalog, a single major film release appeared as 20 separate entries, none tied to a common ID. Unmapped mid‑tail and long‑tail content also introduces popularity bias and suppresses deep‑catalog discovery. Watchworthy's automated mapping resolved 242,055 titles drawn from nearly 50 separate data providers before either test launched.
What is Watchworthy?
Watchworthy is the entertainment personalization system built by Ranker, a consumer entertainment company founded in 2009. Its recommender models have been in continuous development since 2018 and live with consumers since 2020, drawing on the preference graph Ranker has built over that longer history. It exists in two forms: a consumer recommendation application, and an enterprise personalization service for TV platforms, streaming services, and connected‑device manufacturers. The consumer product serves as a live testbed, so engagement, onboarding, and retention optimizations are validated against real viewer behavior before they reach a partner deployment.

Deployment & Integration

Is Watchworthy a complete recommendation system, or a data layer that feeds one?
Either, depending on what a platform needs. Watchworthy can operate as a preference data layer inside an existing recommender, as a candidate‑generation and ranking service, as a grounding layer for AI‑powered discovery, or as a fully managed personalization system. The full stack spans catalog resolution and entity mapping, audience analysis and segmentation, cold‑start handling and adaptive exploration, household disambiguation, slate optimization, context‑aware re‑ranking, and the monitoring and managed operations that keep a live deployment tuned. In the tests described above it ran as a managed service powering a single rail, with the more advanced capabilities outside the scope of that deployment.
Does a platform have to replace its existing recommender to add explicit preference data?
No. The most common starting point is to leave the existing recommender in place and supply preference signal to it, which addresses sparsity and cold start without a replatform. In the tests described above, Watchworthy ran as a single rail within the home‑screen grid, in the same placement as the platform's own personalization and alongside it, rather than replacing anything.
What data does a platform have to share, and is that data used to train the vendor's models?
The only inputs required are a sparse set of interactions — clicks and searches — used as an implicit signal of household preference. In the tests described above the study was fully anonymized: Watchworthy had no visibility into individual test participants or household composition, those signals were used exclusively to retrieve recommendation slates, and no partner data was used in Watchworthy model training. Because those signals are anonymized and carry no viewer identity, a platform can evaluate personalization without the personal‑data handling that often slows security and privacy review.
Can the same personalization be applied to promoted or ad‑supported placements?
Yes. The ranking signal is placement‑agnostic, so the same personalization that powered a single organic gallery can be applied to the rails, collections, promoted placements, and ad‑supported inventory already present on a home screen. The scale of the opportunity is visible in the test result: one gallery of 15 placements, only four visible without scrolling, captured more than 60% of all home‑screen clicks. The engagement the rest of the screen is currently leaving unclaimed is the reason to extend personalization beyond a single surface.
Where else can Watchworthy recommendations be deployed besides the home screen?
Anywhere a platform engages a viewer. The same ranked output that powers a home‑screen gallery can drive the streaming guide, search and browse surfaces, and lifecycle channels including personalized email and messaging, by delivering into the CRM and messaging systems a platform already runs. Watchworthy's own consumer application does exactly this today, using one recommendation service for in‑app surfaces and for outbound messaging. The practical benefit is a single preference layer behind a platform's entire engagement footprint, rather than a separate personalization effort for every surface.
Can Watchworthy operate on platforms that do not support real‑time personalization?
Yes. Watchworthy runs in both real‑time and batch environments, and which one applies is a platform architecture decision rather than a limit of the system — Watchworthy's own consumer application has operated fully in real time since 2020. In the tests described above, partner constraints restricted slate generation to once daily rather than real time, and delivered engagement metrics two days in arrears; Watchworthy operated within both limits and still outperformed the platform's in‑production personalization. Where reporting lags, measurement is modeled rather than postponed: Ranker built a two‑stage model that estimates click‑rate from observable inputs and projects them forward, which tracked actual click‑through rate at a correlation of 0.86 against the days the platform reported full figures. Real‑time delivery is supported where a platform can provide it, and generally improves adaptive learning further.

Operations & Evaluation

Why not build a personalization system like this in‑house?
Because there are three barriers, and the first is larger than it looks. Building a production recommender is a feat in itself — it sits at the intersection of data science, systems architecture, and engineering, and a model that performs in evaluation is a different thing from one that is performant and cost‑effective at production scale. Reaching that first deployment is roughly the halfway point: the remaining gains come from sustained tuning against real traffic, and early versions of any recommender behave very differently from mature ones. Both phases carry substantial cost and time, and both carry the risk that after the investment the system does not deliver the lift it was built for. Watchworthy's recommender models have been in continuous development since 2018 and in production against real users since its consumer application launched in 2020, so a platform integrating it starts from performance that has already been measured rather than projected. The data is the second barrier: 1.5 billion declared preferences across 30 million entertainment entities, accumulated from real consumers since 2009, cannot be engineered from a platform's own logs — precisely the constraint that creates cold start. Catalog resolution is the third, and the one platforms most often underestimate: it is manual, quietly error‑prone, and every downstream system inherits its mistakes. Partners inherit all three rather than rebuilding them.
How long does it take to deploy and test personalization on a TV platform?
With Watchworthy it is weeks rather than quarters, because it integrates into an existing personalization stack rather than replacing it. Catalog mapping and resolution, audience analysis and segmentation, and monitoring infrastructure are all completed before launch — Watchworthy was configured to the platform's audience before it served a single recommendation. That preparation is what allowed one Smart TV platform to run two consecutive live tests in short order: four weeks in August–September 2025 and five weeks in November–December 2025.
Why do recommender deployments require ongoing operation rather than one‑time installation?
Because performance depends on how closely a system is matched to a specific audience, catalog, and season, and none of those hold still. In the first live test, Watchworthy's average daily click‑rate rose 53% between the calibration weeks and the final week following in‑flight re‑optimization, while the platform's own personalization — running concurrently, on the same home screen, across the same days — moved −5%. Only the actively managed surface moved, which locates the improvement in the operating work rather than in seasonality.
How is a personalization deployment managed once it is live?
Watchworthy deployments are fully managed, and operated in alignment with the platform's own team. Daily work covers performance monitoring against the platform's KPIs, demand‑spike detection for new releases and seasonal shifts, investigation of day‑to‑day swings, and content availability checks. On a daily‑to‑weekly cadence, data freshness and pipelines are monitored, engagement is reviewed across user segments and the content catalog, mis‑mapped or unservable titles are identified, and business rules and model inputs are adjusted. Weekly and monthly stewardship covers engineering and data science reviews, release and experiment management, latency, delivery and SLA monitoring, and client‑facing performance reporting. The work sits within Ranker's purview but runs in alignment with the platform: success KPIs are agreed with the partner's team, enhancements and priorities are planned with their stakeholders, and reporting goes to them throughout, so the platform is never looking at a system it cannot see into.
How can a platform evaluate Watchworthy on its own audience?
Ranker runs production tests with platform partners, comparing Watchworthy against the personalization a platform already ships, in the same placement and on the same home screen. Engagement is measured against the platform's own KPIs, with reporting delivered throughout the test. To discuss a test, request a data sample, or review integration options, contact Ranker at business.ranker.com/contact‑us.
All White Papers & Case Studies