What the global benchmark actually does
Netflix attributes more than 80 percent of viewing to its recommendation engine. The engine is not a third-party product Netflix bought and configured. It is in-house, owned end-to-end, refined continuously, and treated as the central product asset that determines the streaming experience. The recommendation runs against the viewer's history, the content catalogue, the contextual signals — device type, time of day, network quality, geographic location — and the personalised artwork that the viewer sees against the same title.
This is the global benchmark. The Middle East OTT operators — Shahid, OSN+, Starzplay, the various regional services and the international platforms that operate in the region — are operating against this benchmark. The investment to compete is real, and the institutional default is shifting toward owning the recommendation surface rather than renting it from a vendor.
The vendor pitch in the recommendation category is well-developed. Salesforce sells personalisation surfaces. Adobe sells personalisation surfaces. The dedicated recommendation specialists — Recombee, Algolia Recommend, the SaaS-positioned start-ups in the category — each have a working product. Each works to a point. None of them, structurally, deliver the differentiation Netflix gets from owning the engine. The reason is architectural. The recommendation engine is too tightly coupled to the content catalogue, the viewer data, the streaming-experience design, and the business strategy to live cleanly inside a vendor's commercial perimeter.
This piece is the architecture Middle East OTT platforms are building, on infrastructure they operate, against the catalogue and the viewer base they own.
What the recommendation agent actually does
The owned recommendation agent operates on four loops in parallel.
Catalogue understanding. The agent reads the content catalogue, the metadata, the genre taxonomy, the cast and crew graph, the language and dialect of the original audio, the subtitle language coverage, the rights window for each title in the relevant market, and the content-similarity graph the platform has built across the catalogue. The understanding is the substrate every other loop reads from.
Viewer understanding. The agent reads the viewer's history, the explicit preferences captured during onboarding, the implicit signals from the watching behaviour — completion rate, episode-to-episode pacing, time-of-day pattern, device pattern, search history, the abandoned-viewing sessions, the rewatch behaviour — and the contextual signals at the moment of the recommendation. The viewer model is updated continuously against the live signal.
Recommendation generation. The agent generates the recommendation set against the viewer model, the catalogue, and the platform's business strategy — the title-promotion priorities, the rights-window expiry urgency, the licence-economics-driven exposure curves, the content-acquisition campaign objectives. The recommendation is the ordered list the viewer sees on the home surface, the post-watch surface, the search-result surface, and the email-and-push channels.
Artwork and presentation. The agent selects the artwork variant the viewer is most likely to respond to against the title. The same title can present different artwork to different viewers. The selection is informed by the viewer's historical response to artwork variants, the title's overall artwork performance, and the cohort patterns in the viewer base.
What stays human
The content strategy. The acquisition decisions, the original-production commissioning, the title-promotion priorities, the editorial curation that the platform's content team performs. The agent does not make the content strategy. The agent operationalises it.
The taste-and-tone judgement that the editorial team applies. The Ramadan programming, the festive-season programming, the cultural-moment programming, the family-appropriate content surface that the regional audience expects — each is editorial work the platform's team performs. The agent reads the editorial intent and respects it.
The recommendation-quality oversight. The platform's product team reviews the recommendation behaviour against the platform's strategic objectives. The agent's eval suite captures the quality metrics. The product team owns the decisions about model refinement, training-data adjustment, and feedback-loop tuning.
The architecture
The data layer reads from the platform's own catalogue, the viewer-data layer, the streaming-telemetry pipeline, the campaign-and-marketing surfaces, and the editorial-intent system. The data lives on infrastructure the platform operates, in open formats the platform's data team can query from any engine. The platform is not dependent on a vendor's customer-data platform for the substrate the recommendation runs against.
The model layer has multiple components. The catalogue-embedding model that produces the content-similarity graph. The viewer-embedding model that produces the viewer-similarity and the viewer-content-affinity scoring. The contextual-bandit components that drive the live recommendation against the viewer's current session. The artwork-variant selection model. Each component is trained against the platform's own data, evaluated against the platform's own eval suite, and deployed on infrastructure the platform operates.
The serving layer delivers the recommendation at the latency the streaming experience requires. The home-surface recommendation is generated against the live viewer session. The post-watch recommendation is generated against the just-completed title and the viewer's pre-watch state. The cold-start surface for new viewers handles the onboarding sequence against the explicit preferences captured at sign-up.
The observability and evaluation surfaces capture every recommendation decision against the viewer's subsequent behaviour. The product team can examine why a particular viewer received a particular recommendation, the influence of each model component, and the response the viewer exhibited.
What the vendor product cannot match
Three structural advantages accrue to the platform that owns the recommendation engine.
The feedback-loop latency. The owned engine retrains, redeploys, and refines against the live data at the cadence the platform's data team operates. The vendor product retrains at the cadence the vendor's platform supports, which is structurally slower.
The catalogue-specific tuning. The Middle East catalogue, the Arabic-language content lifecycle, the regional-dialect distinctions, the cultural and editorial conventions — these are first-class features in the owned engine. The vendor product handles them through configuration, with the configuration cost structural.
The strategy alignment. The recommendation engine is the operational expression of the content strategy. The content team and the data team work the same feedback loop. The vendor product creates a structural separation between the strategy and the operational expression.
The Middle East dimension
Three dimensions matter at a Middle East OTT platform.
The Arabic and English bilingual experience is the operating baseline. The right-to-left presentation, the metadata in both languages, the search behaviour, and the subtitle and dubbing logic all operate as first-class features.
The Ramadan content cycle is the highest-stakes annual programming moment. The recommendation engine carries the Ramadan editorial intent for the duration of the month and reverts cleanly afterward.
The household-account dynamic is more pronounced than at most global services. The shared viewing across family members on a single account requires the recommendation to differentiate intelligently without forcing explicit profile switching at every session.
The saasinator perspective
The vendor pitch for the recommendation surface is rational when the platform is not competing on personalisation. The moment the platform competes on personalisation, the recommendation engine is the product, and owning the product is the only architecture that scales.
Middle East OTT platforms that have made the move are in a different commercial conversation. The vendor relationships they keep are the ones where the vendor adds value the platform does not generate internally. The recommendation engine is not one of those.
What to bring to the diagnostic
Bring the platform's current recommendation footprint, the viewer-data architecture, the catalogue-and-rights inventory, and the product team's strategic objectives for the current planning cycle. The diagnostic is ten working days. The output is the agent recommendation, the architecture sketch, and the first-quarter scope. Book a diagnostic at /diagnostic.