--- title: "Build vs Buy: A Recommendation Engine for Publishers, Both Costs" description: "Build vs buy a publisher recommendation engine, both costs printed: 18 to 24 months and a standing ML team, or eight weeks to a measured +46% in sessions." url: https://www.artificialpoets.com/platform/recommendation-engine/ site: Artificial Poets type: page date: 2026-08-09T06:31:06+00:00 modified: 2026-08-22T08:10:09+00:00 --- # You Could Build It. Should You? # Buy the recommendation engine. Build what makes you different. Every publisher's engine has the same four parts. We ship all four in eight weeks, so your team stays on your product. [See it on my numbers](/request-a-demo/) · [Get the pilot design template](/resources/pilot-design-template/) Horse & Rider · Ride TV · Equine Network · Equus Magazine · The Score · Equine Network Lockup ## The model is one part of four What you are actually buying Teams budget for the part they can picture. The model is real work and it is the smallest piece. Underneath all four parts sits the same foundation: scaling, uptime, monitoring, and one person who owns it for years. - **The model** — The part you costed. Training, evaluation, tuning against your own archive. Two engineers can do this. It is the only piece of the four that feels like the interesting problem. - **The event pipeline** — Every view, scroll and click, collected, cleaned and kept in order. When it breaks on a Sunday the recommendations go stale quietly. Nobody notices for a week, and by then you cannot tell which days of training data to throw away. - **The serving layer** — An answer in under 100ms, on every article page, on your worst traffic day. This is the part that pages someone at night. It has to stay up when a story breaks, which is exactly when it is hardest and matters most. - **The editorial surface** — Where editors pin, exclude and set rules. Skip it and your newsroom has no way to correct the machine, so the first bad recommendation ends their trust in it. This is the part teams cut for v1 and pay for later. ## Who owns this in year three? The ownership test - **A model gets worse on its own** — It is trained on the archive as it was. Your archive keeps growing and your readers keep changing. Skip retraining for a quarter and the recommendations quietly get worse. There is no alert for this. - **Hiring is the long pole** — A machine learning hire takes three to four months to ship anything, on top of however long the search takes. That clock starts after you get the headcount approved, not when you decide. - **There is no finish line** — Retraining, drift monitoring and pipeline upkeep are permanent jobs. You are not approving a project with an end date. You are adding a standing line to payroll. ## What it costs over five years - **$500k a year** Salary alone for the smallest team that can run this. One ML engineer, one data engineer, and someone who owns the pipeline when it breaks - **$2.5m** Five years of that team. Before infrastructure, before the rebuild when the person who built it leaves and takes the reasoning with them - **8 weeks** Authorisation to a result you can read in your own analytics. Four weeks of measuring only, then four weeks live ## What the eight weeks buys Build vs buy With Artificial Poets Platform you can see the results in just eight weeks. - **Your engineers stay on your product** — This is the real argument. Your team is good enough to build it, which is exactly why their time is worth more somewhere else. - **A date you can put in a board deck** — Eight weeks from authorisation to a measurable effect. We have run this on live publisher traffic and we will show you the ones that went slowly too. - **The upkeep is ours** — Retraining, drift monitoring and pipeline repair sit with us. If the model starts to degrade, noticing it is our job, and you never had to hire anyone to do it. - **Numbers before you commit, not after** — A build gives you nothing to look at until it ships. The first four weeks with us measure your baseline with nothing served, so the comparison is against your own traffic, not a model of it. - **Editors keep the last word** — Pin, exclude and set rules per section. When an editor overrules the system, the system stays overruled. You are not buying a black box you have to argue with. - **Your data and your content stay yours** — Reader profiles are anonymous and your behavioural data leaves with you. Ask us for the exit clause before you sign, not after. We will send it. ## Build it yourself if any of these are true - **Your product** The ranking is what customers pay for — If recommendation is the thing your customers pay for, you cannot buy it from anyone. That is your roadmap, not your infrastructure. Build it. - **Ownership** You have one named owner for three years — Not a team that could do it. One person whose job this is, who will still be here in 2029. If that is you, build it. - **Data residency** The data cannot leave your building — Some contracts and some jurisdictions settle this for you. We host in your region, but if it must be your own hardware, that is a build. - **Catalogue** Your content does not behave like articles — If your content is not articles, video or audio, a general model may not fit it. Test that before you buy anything, including from us. **The cost of deciding slowly** ## What does another quarter cost you? ## FAQ ### Is a recommendation engine really not a differentiator? The logic on top of it is. The four parts underneath are not. Every publisher who builds one builds the same event pipeline, the same serving tier and the same editorial screen, and none of those are why readers choose you. Your archive and your editorial judgement are the differentiator, and you keep both either way. ### Our engineers say they could build this. Could they? Almost certainly yes. Capability is rarely the blocker. The blocker is that it takes 18 to 24 months to a first production ship, it needs three people you have to hire and keep, and it never finishes. Ask them a different question. Not can we build it, but what do we stop building if we do. ### What does building it actually cost? Salary is the floor. About $500,000 a year at 2026 US medians for the smallest team that can run it, which is roughly $2.5m over five years before infrastructure. That figure assumes nobody leaves. When the person who built it does leave, budget for a rebuild, because the reasoning behind the thresholds was never written down. ### What do we give up by buying? Less than the usual answer, and we will be specific. Editors pin, exclude and set rules, and an override stays an override. You do give up control of our roadmap and our pricing, which is a real cost and worth naming. If exact fit to an unusual data model matters more than time, build it. ### We have been burned by recommendation widgets before. How is this different? Those networks are paid to send readers off your site. We are paid to keep them on it. The engine serves your articles only, inside your templates, and the ad slots it creates are yours to sell. If a reader leaves, we did not do our job. ### Your models train on our first-party data. What do we keep when the contract ends? Your behavioural data is yours and leaves with you in a documented format. Ask for the exit clause during evaluation rather than at renewal. We would rather you read it early. ### Adtech keeps consolidating. Will you exist in three years? A fair thing to ask of anything you buy instead of build. The honest answer is that no vendor can promise it, so ask for the thing that survives the answer being no. That is your data in a portable format and a notice period long enough to move. Both are in the contract. ## Related - [You Can’t Get the Traffic Back. You Can Get the Session Back.](https://www.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. - [Horse & Rider Grew Pages Per Session 19% Without Growing Traffic](https://www.artificialpoets.com/customers/horse-and-rider/) — In the first quarter after enabling the Artificial Poets 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. - [Make Every Visit Worth More](https://www.artificialpoets.com/solutions/session-depth/) — "Traffic keeps falling and I can't stop it — how do I make the visits I still get worth more?" ## See it running on your own numbers We'll walk you through what the Platform does on titles like yours. [Book my demo](/request-a-demo/) · Calculate ROI ## Questions this page answers ### Should we build or buy a recommendation engine? Buy it unless the recommender is the product you sell. A first in-house production ship runs 18 to 24 months and needs ML, data and MLOps roles hired and retained before a reader sees anything, then permanently for retraining and monitoring. A deployment runs eight weeks from authorization: four weeks of measurement-only baseline, then four of serving and tuning, run by our team. ### What does it cost to build a recommendation engine in house? Salary is the floor. The smallest viable team is three people, an ML engineer, a data engineer and someone who owns MLOps, which is roughly $500,000 a year at 2026 US medians before benefits, recruiters or infrastructure. Budget three to four months before a new ML hire ships anything, and treat retraining, drift monitoring and pipeline upkeep as permanent roles rather than a project phase. ### How long does buying one take instead? Eight weeks from authorization on the Platform. Four weeks are measurement only while the engine indexes the archive and the baseline is frozen, then enablement on a day you pick, then four weeks of serving and tuning. The effect is measurable at week eight in your own analytics against that frozen baseline. ### What do you give up by buying rather than building? The ability to change the core ranking yourself; deeper tuning is run with our team. Everything publishers usually fear losing stays: editors pin, exclude and constrain what the system may serve, selection runs inside those rules, the optimization target is voluntary continuation rather than click-through, and your behavioural data and content stay yours and are never pooled across customers. ### Does a bought recommendation engine actually work on publisher titles? On the measured network, multi-page sessions rose 46% on enabled titles (10.3% to 15.1%) while comparison titles fell 16% (12.0% to 10.1%) over nine months. Verified three ways: a comparison group of the entire eligible set, a window-sensitivity sweep, and a four-week pause in serving that returned the metric to baseline and recovered on resume. Sessions and users were flat, so this is a depth result and we say so. ### How is this different from Taboola-style widgets? Those networks rent your page to send readers somewhere else, and the pennies come with a trust cost. The Platform serves only your own library on your own domain, is measured on voluntary continuation rather than click-through, and your behavioural data stays yours. Editors can pin, exclude, and constrain what it may serve.