The Product Manager's Guide to Personalisation Engines

Showing everyone different things doesn’t mean showing everyone better things. Personalisation without strategy is just - wait for it - chaos at scale.

Why Personalisation Is Harder Than It Looks

Everyone wants Netflix-level personalisation. “Show each user exactly what they want.” Yes, it sounds simple. “Just do it like Netflix did.” Oh my. This requires infrastructure most product teams don’t have and creates problems most product teams haven’t considered.

I read a post about a team building personalised recommendations. These guys used collaborative filtering - if user A likes items 1, 2, 3 and user B likes items 1, 2, recommend item 3 to user B. It clearly is a standard approach.

But the results were terrible. Why were they terrible? The content had a long tail, most items had few interactions. Collaborative filtering was optimised for popular content. You know what happened. They built a popularity engine, not a personalisation engine.

Even when it worked, it created new problems. Editorial team couldn’t curate homepage anymore - it was different for everyone. A/B testing became complicated - control and variant groups weren’t seeing the same baseline. Support couldn’t debug issues because “but it looks different for me” (famous iteration of “it works for me…”).

Personalisation introduces system complexity that quickly compounds. It is fun, until it stops being fun - and that happens very early in the process.

Technology Overview

Current State: What’s Actually Possible

Skip the hype. Here’s what personalisation engines can realistically do today:

Content recommendation: If you have explicit signals (clicks, likes, purchases) and reasonable volume, you can build decent recommendations. Netflix, Spotify, Amazon do this well because they have millions of interaction events.

Your B2B SaaS with 500 users probably doesn’t have enough data. You’ll end up recommending what’s already popular, which isn’t personalisation.

Behavioral targeting: Show different content/features based on user attributes (role, company size, usage patterns). This works well and doesn’t require ML. Technically not personalisation - it’s segmentation - but it worked.

Predictive models: Churn prediction, upsell scoring, feature recommendations. Need historical data and clear outcome metrics. If you have both, these can work. If you’re pre-PMF or changing rapidly, models become stale fast.

Key Capabilities: What Matters vs What’s Shiny

The capabilities that actually move metrics:

Segment-based personalisation: Not true 1:1 personalisation. Show different experiences to different customer segments. Low complexity, high impact. At one company, we had three segments (trial users, paying SMBs, enterprise). Different homepage, different feature prominence, different messaging. Conversion improved 31%.

Search ranking personalisation: Adjust search results based on user context. Same query, different users, slightly different ranking. Spotify does this well - search “workout” and results depend on your listening history. Requires good search infrastructure first.

Email/notification personalisation: Send different messages based on user behavior. Someone who uses feature A daily gets different product updates than someone who hasn’t logged in for a week. High impact, relatively simple to implement.

The shiny capabilities that usually don’t matter:

Real-time personalisation: Adjusting experience based on in-session behavior. Technically impressive, rarely moves metrics unless you’re a high-frequency product like news or social media.

Cross-platform personalisation: Connecting behavior across web/mobile/email. Useful in theory, complex in practice, marginal impact unless you have significant cross-platform usage.

Hyper-personalised UI: Every user sees a completely custom interface. Sounds great, breaks everything else. Can’t do QA, can’t troubleshoot, can’t curate, can’t test effectively.

Product Applications

Use Cases That Actually Work

Onboarding paths based on user intent: Ask why they signed up, show relevant features first.

Feature discovery based on role: Show different “getting started” tips for different user types. Engineers see API docs, product managers see dashboards, sales sees CRM integration. Basic segmentation, but it works.

Content recommendations in specific contexts: Not homepage personalisation. Contextual recommendations. User viewing a document? Suggest related documents. User on a settings page? Suggest relevant integrations. Context matters more than user history.

Integration Approaches: Build vs Buy vs Skip

Build if: You have unique data, specific requirements, engineering resources, and personalisation is core to your value prop. Netflix builds because recommendations ARE the product.

Buy if: You need standard personalisation (email, recommendations, search), want fast implementation, don’t have ML expertise. Tools like Algolia, Segment, or Amplitude Recommend work well.

Skip if: You’re pre-PMF, have limited user data, or haven’t nailed your core product experience. Fix the base product first. Personalisation won’t fix a broken experience.

Future Implications

Trends to Watch

LLM-powered personalisation: Using language models to generate personalised content/recommendations. Early days. Expensive. Prone to errors. But potential is real.

Privacy-first personalisation: With cookie deprecation and privacy regulations, personalisation without tracking is the challenge. Techniques like federated learning, on-device personalisation, and differential privacy matter more.

Apple does this well - personalisation happens on device, data never leaves. Most companies don’t have Apple’s resources.

Explainable personalisation: Users want to know why they’re seeing something. “Because you liked X” is basic explainability. Future is better transparency about personalisation logic.

Preparing Your Team

Don’t hire ML engineers for personalisation until you’ve proven you need them. Start with:

Product thinking about segments: Who are your user types? What do they need? Can you serve them better with segmentation before you need personalisation?

Data infrastructure: You need clean event tracking, user attributes, and outcome metrics before personalisation makes sense. Fix data quality first.

A/B testing capability: Personalisation needs measurement. If you can’t run controlled experiments, don’t build personalisation. You won’t know if it works.

Key Takeaways

Right, let’s be concrete:

• Segmentation beats personalisation for most products - Group users into segments and optimize for segments. Simpler, easier to test, often better results.
• Fix your core product before adding personalisation - Personalisation won’t save a mediocre experience. It’ll just show people mediocre content faster.
• Start with high-context, low-complexity personalisation - Onboarding flows based on intent. Feature tips based on role. Recommendations in specific contexts, not everywhere.
• Build data infrastructure before ML - Clean event tracking, user attributes, outcome metrics. You need this regardless. Personalisation can wait.
• Measure everything - Personalisation adds complexity. Only worth it if you can prove it moves metrics that matter.
• Consider buy over build - Unless personalisation is your core differentiator, use existing tools. Your engineering time is more valuable elsewhere.

Final Thoughts

Personalisation is seductive. It sounds sophisticated. It promises magic - show each user exactly what they want. Investors love it. Engineers get excited about ML.

But most products don’t need personalisation. They need better core experiences, clearer value props, simpler workflows.

If you’re considering personalisation, ask: Is our core experience good enough that showing it to the right people at the right time would make a difference? If not, fix that first.

Personalisation is an amplifier. It makes good products better. It makes bad products worse, just in personalised ways.

Have questions or thoughts? Get in touch - I’d love to hear from you!