十個事件、三個數字：以工程師方式為 MVP 設計數據追蹤

Most MVPs launch in one of two states. Either there is no analytics at all, or someone dropped in the default snippet and it has been dutifully recording page views and button clicks that nobody has opened since launch week.

Both end the same way. The founder asks "is it working?" and the honest answer is a feeling.

That question is an engineering problem even if nobody frames it as one. Analytics breaks for the same reasons other systems break: vague contracts, unreliable clients, identity bugs and data nobody validates. Here is how I treat instrumentation on a first version.

Start from decisions, not events

Do not open an analytics tool and ask "what should we track?" Ask what you will decide in the next three months, and what data would change your mind. For a typical MVP:

• Is the core loop working? Do people who sign up reach the moment where the product does its job?
• Where do they drop off? Which step between install and that moment loses the most people?
• Do they come back? Of the people who got value once, how many return without a nudge?
• Which channel is worth doubling down on? Do users from one source behave differently?

If an event maps to no decision, it is noise you pay for later in cleanup, event-volume pricing and confusion. Same instinct as logging: a log line nobody reads is just cost.

Activation is a spec, so write it like one

Activation is the moment a new user first gets the value the product promises. A few from products we build and run at Null Studio:

• LectureNotes AI records lectures and turns them into summaries and outlines. Signing up is not activation. Activation is a student opening their first generated summary of a real lecture.
• Lifemaxxing AI builds personalized self-improvement plans. Activation is closer to receiving the plan and logging the first piece of progress against it, not finishing onboarding.
• Fyuel, an accounting platform for the fuel trade, is a business tool. Activation there is a company getting real customers, suppliers and transactions into the system, because an empty ledger has delivered nothing.

None of these are "created an account." Write the moment in one sentence, get the team to agree, then instrument every step that leads to it. If the team cannot agree on the sentence, analytics just found you a product problem for free.

The tracking plan is your event contract

A tracking plan lists every event the product sends, when it fires and what it carries. I treat it as an API contract between the product and whoever reads the data later. Ten to twenty events is plenty. Each row needs:

• A name in one consistent style: past tense, snake_case, like summary_viewed or plan_created.
• When it fires, precisely. "When the summary screen finishes loading with content" beats "when the user sees the summary." The second one will be implemented three different ways by three different people.
• Properties, where it fires (device or server) and which decision it supports.

If your stack is TypeScript, you can make the plan enforce itself instead of living only in a spreadsheet. This is illustrative, not lifted from a specific codebase, but it is the shape I reach for:

// Illustrative: the tracking plan as a type, so a typo is a compile error
type Events = {
  signup_completed: { source: "organic" | "paid" | "referral" };
  recording_uploaded: { durationSec: number; platform: "ios" | "android" | "web" };
  summary_viewed: { lectureId: string; isFirst: boolean };
  plan_created: { planType: string };
};

export function track<E extends keyof Events>(
  event: E,
  props: Events[E],
  userId: string
) {
  analytics.capture({ distinctId: userId, event, properties: props });
}

// track("summary_veiwed", ...) fails to compile.
// track("summary_viewed", { lectureId }) fails too: isFirst is missing.

The point is not the wrapper. The point is that the event list now lives in the repo, shows up in code review and cannot drift silently from what the dashboard expects.

Send outcomes from the server

Events that record something the product did, like a payment succeeding, a summary being generated or an invoice being saved, belong on the backend. Device-side tracking gets blocked, lost offline and duplicated when a mobile client retries. If your activation event fires from the client, your activation rate is partly a measurement of how many users run an ad blocker. The server knows the summary was generated. Let the server say so.

Keep device events for what only the interface knows, like which screen someone abandoned. And name everything for the person who joins in six months: btn_click_3 tells them nothing, checkout_started with a plan property tells them everything.

Identity is the bug that corrupts your best data

Analytics breaks quietly when one person looks like three users. Someone browses anonymously, signs up on the web, then logs in on their phone. Without care, that is three distinct IDs and a funnel that says your conversion is a third of what it really is.

Decide early how you will:

• Merge anonymous activity into the account at sign-up, so the steps before registration still count toward that person's journey.
• Recognize one person across web and mobile, using your own user ID rather than a device ID.
• Carry an account or organization ID if you sell to businesses. For a B2B tool like Fyuel, the real question is often whether a company activated, not whether one person did.

The nasty part is timing. Fix identity after launch and the data you cannot trust is from your launch weeks, exactly the period you most wanted to read.

Three numbers, reviewed every week

Early on, three numbers do almost all the work.

Activation rate. Of the people who signed up this week, how many reached the activation moment? If it is low, walk the funnel to find the step where people stop.

Retention by cohort. Group users by signup week and track how many come back in the following weeks. A growing total user count can hide the fact that every new cohort leaves as fast as the last one. The cohort chart cannot hide that.

The core value action. The one repeated action that means the product is delivering: summaries viewed, plans updated, transactions recorded. Usually a better health signal than DAU, because people open apps without getting anything out of them.

Review those three every week with a named owner. A metric nobody looks at is not a metric. It is storage. And pair them with a few conversations with users who activated and users who dropped off: with dozens or hundreds of users, the why matters more than the what.

Privacy is part of the schema

Decide before launch what you collect, how long you keep it and how users opt out or delete it. Never send sensitive content, like the text of someone's notes or financial details, into third-party analytics. The typed schema helps again: if there is no noteText property in the plan, nobody casually adds one. On mobile, this also has to match the privacy disclosures you give Apple and Google. For AI features, the same care applies to prompts and outputs, which I covered in Nobody Opened Your Product to Chat With It.

Do not build your own pipeline yet

I feel the pull to build a custom event pipeline too. Resist it until you know the product has users. The effort belongs in the plan, server-side events and identity, not charts. A few rules:

• One tool, not three. Three tools with slightly different numbers produce arguments, not decisions.
• Own the data. The account should be in the company's name, with raw event export available. Same principle as Your README Is a Lie Until Someone Else Runs It: if you cannot hand it over, you do not really own it.
• Check the pricing curve. Many tools bill by event volume, one more reason not to auto-capture everything.
• Test events like features. They break when screens change and nothing fails loudly.

It is a few days, not a project

Instrumentation is a few days of work inside the MVP build, far cheaper then than retrofitted later. When we scope a first version, the tracking plan is written with the feature list, so activation is agreed before development starts.

My own product taught me why. When LectureNotes AI went viral on Instagram, the reach was the easy part. The real work was the path from view to paid: getting a student to their first good set of notes fast, and keeping them through the next exam season. I wrote that up in 15 Million Views, 10k Users, $1k MRR.

The original, longer version of this guide is on the Null Studio blog: Product analytics for an MVP.

What is the one event in your current project you would be most afraid to discover was firing wrong?