Content Relevance – The Invisible Product Series

When you open any social or entertainment app, the feed that greets you feels almost magical. It’s there, it’s moving, and you just keep scrolling. Yet every single item you see – a friend’s photo, a viral short, a suggested song – is the result of a quiet, relentless decision‑making engine. That engine is the ranking system, and for most of us it’s the product we never see.

In this series I call that hidden engine the “invisible product.” It’s not a bug; it’s a feature. It translates a platform’s values – what it thinks matters – into millions of tiny choices each day. Those choices, in turn, define the experience for the user, the creator, the brand, and even the broader culture that grows around the app.

My own path to this perspective started in the trenches: I’ve built ranking models for Instagram comments, Stories, Notes, sharing suggestions, ads, marketplace listings, and the main discovery feeds. The tech stack often looked the same – similar data pipelines, shared infrastructure, comparable loss functions – but the definition of “value” was never the same. A comment ranker should surface thoughts that help you join a conversation. A Stories tray should keep you in touch with the people you care about. A sharing suggestion must respect relationships, not just content similarity. And a streaming homepage has to bridge the gap between opening the app and committing an hour of attention.

Calling all of that “engagement optimization” is a cop‑out. The real work is translating each surface’s unique promise into three things: the pool of candidates it can ever consider, the objective it tries to optimize, and the safeguards that keep the system honest.

Below is a quick roadmap of the seven essays that make up the series. Each one peels back a layer of the invisible product, from the high‑level product promise down to the nitty‑gritty of generative AI recommendations.

1. The Feed Is a Product Constitution – Ranking starts with a contract, not a model. We explore how a clear statement of intent shapes everything that follows.

2. The Candidate Set Is Destiny – Retrieval quietly decides what the ranker can ever pick. If the right items never make it into the pool, the ranker is blind.

3. The Objective Function Is the Product – Metrics are not neutral; they encode values. Choosing a loss function is a moral decision dressed in math.

4. Your User Is Not a Vector – Intent, context, relationships, and evolving tastes make personalization far richer (and more complicated) than a set of numbers.

5. Serendipity Is a Systems Problem – Exploration isn’t a nice‑to‑have; it’s essential for keeping creators discoverable and users curious.

6. Every Ranker Creates Its Own Reality – Recommendation loops shape future data, steering the whole ecosystem in subtle ways.

7. After the Ranker – Sequence models, foundation models, and generative AI are redefining the unit of recommendation itself.

---

Part 1: The Feed Is a Product Constitution

TL;DR – Ranking is the invisible operating system of modern apps. It turns a platform’s abstract values into concrete, repeatable decisions about whose voice is heard and what content gets a spotlight.

Open any social app and pull down to refresh. In the fraction of a second it takes for the UI to spin, the backend has already chosen: which friend’s post to surface first, which creator deserves a chance, which conversation feels timely, which song matches your mood, and even which brand‑new idea should interrupt your routine. The feed appears seamless, as if it has always been arranged that way. But every position, every omission, is a deliberate choice.

Those choices become the product when they happen billions of times a day. The feed isn’t just a list; it’s a constitution that encodes the platform’s promise to the user. Is the promise “keep me close to the people I know”? “Show me the next binge‑worthy show”? “Introduce me to fresh perspectives”? The answer determines everything that follows – from how candidates are fetched to what metrics are rewarded.

When I moved from ranking comments to ranking Stories, the most striking lesson was that the same model architecture didn’t magically solve both problems. A comment ranker needs to surface relevance that helps you understand a thread or join a debate. A Stories tray, by contrast, should surface moments you’ll want to peek at in the next few minutes – it’s about recency, relationships, and emotional tone, not just topical relevance.

So the first step is to write the product promise in plain language. Only then can you design a retrieval pipeline that pulls in the right candidates, define an objective that aligns with that promise, and build feedback loops that keep the system honest over time.

In the next essay we’ll look at the hidden gatekeeper – the retrieval stage – and see why “what you don’t see” matters just as much as “what you do see.”