Content recommendation algorithms on streaming platforms have changed dramatically since the AI capability wave hit production. The platforms have all retooled their recommendation stacks at least once in the past three years, and the differences between platforms in May 2026 are more visible than they’ve been in a long time.

What’s actually working: cold-start recommendation for new subscribers has gotten meaningfully better. The combination of richer item embeddings — generated from content metadata, screenplay, dialogue, visual style — and few-shot learning on early viewing signals has compressed the time from sign-up to relevant recommendations from weeks to a couple of sessions. That’s a real product improvement, even if the platforms don’t talk about it loudly.

Within-session recommendation has also gotten better. The “what should I watch right now” question is being answered with more contextual awareness than it was even eighteen months ago. Time of day, recent viewing context, household viewing pattern, and stated mood all feed into the live recommendation in ways that produce a noticeably more useful homepage than the 2024 version.

What’s still not working: discovery beyond the existing taste profile. The recommendation systems are still optimised against engagement metrics that reward staying in the user’s current preferences rather than expanding them. The 2026 algorithms are excellent at telling you the next thing very similar to what you already watched. They’re poor at suggesting the thing that’s two steps removed but might genuinely interest you. This is a structural problem, not a technical one. The optimisation target is wrong, not the model architecture.

The “explainable recommendations” feature several platforms have rolled out is interesting. Showing why a particular show is being recommended — natural-language explanations generated by an LLM tied to the recommendation reasoning — has improved user trust in the system, even when the recommendations themselves haven’t substantively improved. The transparency is doing real work.

The other quiet trend is recommendation-system convergence. The platforms have largely solved the same engineering problems with similar architectures, and the recommendation-quality differentiation between platforms has narrowed. The platforms that win on user experience now win less on raw recommendation quality and more on UI, content depth, and the human curation layer that the better services have started investing in.

Human-curated lists are making a quiet comeback. After a decade of pure-algorithmic recommendation, several major streamers have rebuilt editorial teams. The combination of algorithmic surfacing with human-curated context appears to outperform pure algorithmic on subscriber retention metrics. That’s not a popular finding inside the platforms — it implies that the algorithm-only thesis was overblown — but the data is consistent enough that the editorial teams are growing rather than shrinking.

The privacy dimension is also worth noting. The 2026 recommendation systems are generally running on more limited per-user data than the systems they replaced. The combination of regulatory pressure and platform-level decisions to reduce identifiable personal data in recommendation pipelines has pushed the engineering teams toward methods that work with less rich user profiles. The recommendation quality has held up better under these constraints than expected, which is a credit to the engineering.

For Australian streaming subscribers in 2026, the practical takeaway is that the platform with the best recommendation algorithm for you is probably the platform whose content depth aligns with your taste. The recommendation-quality differential between major platforms has narrowed enough that catalog choice matters more than it used to. The AI improvements are real, but they aren’t large enough to overcome a content-catalog mismatch.

The next move in the recommendation space is likely to be agentic. Several platforms are quietly experimenting with recommendation experiences where you can have an actual conversation about what you want to watch and the system responds with reasoning rather than just a tile grid. The early implementations are uneven. The direction is clear.