--- title: OTT Recommendation Engine for Niche Streaming Services description: Specialty streaming has churned 6.6–9.2% monthly since 2023 (Antenna), and 19% of viewers give up without choosing anything (Gracenote). An OTT recommendation engine for the next-episode decision. url: https://preview.artificialpoets.com/use-cases/video-recommendations/ site: Artificial Poets type: page date: 2026-08-09T19:46:31+00:00 modified: 2026-08-18T02:10:14+00:00 image: https://preview.artificialpoets.com/wp-content/uploads/a13s-cards/1505-social-1e7742f4.png --- # Get Them to the Second Episode **The same engine, on new surfaces** "Subscribers churn before a second episode — how do I get them watching in the first session?" The fork between staying and cancelling opens in the first session: viewers spend 12 minutes looking for something to watch and 19% give up without choosing (Gracenote). The Platform makes the decision that closes that fork, what plays next per viewer, with the engine measured at +46% on publisher web surfaces: measured on web, not yet on video, and we say so first. [See it on my numbers](/request-a-demo/) Horse & Rider · Ride TV · Equine Network · Equus Magazine · The Score · Equine Network Lockup ## The problem, in your numbers Deloitte also finds 22% of consumers churned and returned to the same service. Serial churners are not rejecting your catalog. They are rejecting sessions that end with nothing chosen. - **6.6–9.2%** Specialty SVOD monthly churn since 2023, above premium SVOD throughout (Antenna, Q3 2025) - **12 min** Spent per session searching for something to watch; 19% give up (Gracenote, 2025) - **49%** Of viewers call difficulty finding something to watch grounds to cancel (Gracenote, 2025) - **41%** US consumers who cancelled a streaming service in the past six months (Deloitte) ## The first-session fork One signup, two paths — the engine's job is the decision in the middle. The churn figures are industry benchmarks (Antenna, Gracenote, Deloitte), not our measurements. - **The engine acts here** — The next-episode and continue-watching decision, made per viewer, in the first session. - **Found something to watch** — Comes back — the habit starts here. - **Browsed, closed the app** — Churn risk band: 6.6–9.2% a month for specialty SVOD since 2023 (Antenna). ## Your articles are your best trailers - **The highest-intent audience the matching series will ever have** — If you publish editorial alongside the service, your best promotion surface is already live: the reader four articles deep in a topic is the highest-intent audience the matching series will ever have — and today nobody introduces them. The same fingerprint that picks the next article can hand that reader the series, and hand the finished viewer the companion articles between releases. - **A loop you own end to end** — Deloitte finds 44% of streaming fans — about 60% of Gen Z fans — discover content on social media and watch it elsewhere, a loop you can neither attribute nor act on. The editorial-to-video bridge is a loop you own end to end, on first-party data. ## The evidence, at its honest tier - **Plainly: our measured results are on publisher web surfaces.** — This application runs the same mechanism — vectorization, fingerprinting, intent matching, orchestration — with no measured deployment on a video surface yet - **On web** — multi-page session share +46% on enabled titles (10.3% → 15.1%) vs −16% on comparison titles (12.0% → 10.1%) over nine months - **Continuation past the fourth served item ran 70–82%** — voluntary continuation, the exact behavior a next-episode decision has to produce - **A first video deployment starts with a four-week measurement-only baseline** — and an agreed comparison design — the first retention claim will be one your analyst can check The web deployment behind the mechanism, not a video result: [two titles grew multi-page sessions 46% while the comparison group fell 16%](/customers/equine-network-rollout/). **Introducing** ## Artificial Poets Platform One engine behind every solution on this site. It learns your archive and your readers, then acts inside your CMS, your templates and your ad stack. - It learns your archive: **Every story you have published, current again** — The engine understands each piece by what it is about, not when it ran or where it was filed. A feature from 2019 competes for the next slot on merit with one from this morning. - It reads the visit: **What a reader wants, without asking** — Interest builds from what someone actually does in the session. No login, no third-party cookies, nothing leaving your domain. Useful on the second pageview, not the tenth visit. - It chooses: **The right next read, not the popular one** — Someone comparing products and someone following a running story want different things. A most-read list gives both the same five links and serves neither. - It serves: **There before the reader leaves** — The feed, the recommendations, the search answer and the signup ask all run on the same engine, in your templates and your ad stack. Any slot that arrives with them is yours to sell. - It proves: **A lift you can defend, or we say so** — Every deployment runs beside titles that did not get it, plus a serving pause. That is how a result becomes a number you can take to a board instead of a vendor claim. ## FAQ ### We have a few hundred titles, not a Netflix-scale catalog — what can the model surface that our editors couldn't hand-pick? Editors can hand-pick a good home row. They cannot hand-pick per viewer, per session, at the moment of the decision — and a few hundred titles is thousands of episodes once the problem is per-viewer ordering. Between releases, the next-best archive episode is what carries a subscriber across the content gap. New episodes are recommendable from their transcripts before they have watch data — cold start is handled by the content, not the crowd. ### Our churn runs around 7% a month. What would this move it to? We will not give you a number, because we have not measured one on a video catalogue. Specialty streaming has run 6.6 to 9.2% monthly churn since 2023 (Antenna), and the mechanism is proven on publisher web surfaces at +46% multi-page sessions. A retention claim would have to come from your deployment, measured against a comparison design agreed first. ### On day one a new title has zero watch history. Does it sit at the back of the rail? No, because matching starts from what a title is about rather than from what it has accumulated. Titles are embedded from their own metadata, description and transcript, so a release competes on subject from the first session and blends in engagement signal as it earns one. ### Our retention moves with our release calendar — how do you separate your lift from our slate? Measurement first, the way the web deployments ran: a four-week baseline before any serving, frozen metrics, and a comparison design agreed before enablement. We read cohort survival curves by signup week, so slate effects appear in both arms. On web, a four-week pause in serving made the dependency visible: +12.2% the quarter before, −2.7% during the pause, +6.9% after serving resumed. If the lift is not there, the measurement will say so. ### Our metadata is thin. What do you actually need before this works? Text is the requirement: titles, descriptions and transcripts where they exist. Thin metadata is common; roughly a third of programmes submitted to Gracenote's FAST database lacked genre information. The honest sequence is to see what your catalogue carries before scoping anything, because that is where the real project cost usually sits. ### We run a white-label OTT stack with no real dev team — and whose data is it once it's in your platform? Integration is carried out by our team through the event exports and APIs your stack already exposes — weeks of our work, not a six-month project of yours. The data stays yours: profiles are anonymous, no PII is stored, and your audience's behavior is not pooled into models served to other customers. ### Why should we take a bet on a surface you have not measured? You should not, on our word. The measured record is on publisher web pages and we lead with that limit rather than bury it. What a first video deployment gets is the same design that produced the web numbers: four-week baseline, frozen metrics and a comparison group, so the first claim is one your analyst can check. ## Related - [Your Reader, Viewer and Listener Are One Person.](https://preview.artificialpoets.com/use-cases/cross-media-graph/) — "Why does my podcast app have no idea what this person reads?" - [One Audience. Every Property.](https://preview.artificialpoets.com/solutions/media-networks/) — "We own a dozen properties — why does our audience behave like it belongs to strangers?" - [Horse & Rider Grew Pages Per Session 19% Without Growing Traffic](https://preview.artificialpoets.com/customers/horse-and-rider/) — In the first quarter after enabling the Platform, one in six Horse & Rider readers went past the first page, up from one in ten. Three comparable titles moved less than a quarter of a point. - [You Can’t Get the Traffic Back. You Can Get the Session Back.](https://preview.artificialpoets.com/resources/traffic-back/) — The economics of session depth, and the causal evidence behind it: +53% pageviews per user, 95% CI [+48%, +58%], with the modelling assumptions written down. ## See how this works on your titles Thirty minutes on your analytics. We tell you what share of your sessions stop at the first page, and what this would realistically move first for a setup like yours. You leave with the annotated read. No deck. [See it on my numbers](/request-a-demo/) How these numbers are made: [the measurement method](/solutions/measurement/). Figures from a multi-title publisher network measured continuously on publisher web surfaces, August 2025 to May 2026. Last updated August 15, 2026. ## Questions this page answers ### What is an OTT recommendation engine for niche streaming? It is the decision about what plays next, made per viewer rather than from a fixed carousel. Specialty streamers have churned 6.6 to 9.2% of subscribers a month since 2023 (Antenna), and the fork opens early: viewers spend 12 minutes looking for something to watch and 19% give up without choosing (Gracenote). ### How would it work on our video catalogue? The same mechanism as the web deployments: titles embedded by what they are about, an anonymous interest profile built from what the viewer does in the session, and the next title chosen on merit. Stated plainly, this surface carries no measured claim yet; our measured results are on publisher web pages. ### How much does it cost? There is no public price list; pricing follows network size and formats. Because this surface is unproven, a first video deployment is structured as a measured pilot rather than a rollout: a four-week baseline, an agreed comparison design, and a retention claim only after it survives that. ### Who is it for? Specialty and enthusiast streaming services whose subscribers churn before a second episode, particularly publishers who already run editorial sites alongside video. If you also publish articles, the reader who is deep in a topic is the most likely viewer of the series about it. ### What are the alternatives? Editorially curated rows work until the catalogue outgrows the team maintaining them. Platform-native recommendations optimise for the platform. Building in-house means an ML team and a multi-year project. The honest comparison here is that our alternative is proven on web and not yet measured on video. ### How do I get started with video recommendations? A first video deployment starts the way the web ones did: four weeks of measurement-only baseline, frozen metric definitions, and an agreed comparison design before anything is served. The first retention claim will be one your analyst can check, because it will be produced the same way.