Installing analytics is not getting out of the building

Lean Startup said get out of the building. The launches I watch treat a survey or a dashboard as enough. The customer already walked a path through the product: signup, first value, the action they repeat, the stop they drop off at. None of those stops have a name in the telemetry.

Installing Mixpanel is not knowing the customer. Tagging the journey is. Until each stop carries a name, a time, and a count, the warehouse is a pile. It's not a story about how they use the product.

I haven't wired this as a discipline on a product I own. I've watched enough launches skip it to know the shape of the miss. I fall in love with the solution, ship the thing I designed, and reach for feedback I can read in a spreadsheet while the path they already walked goes untagged.

That's the work:

1. name the path I believe they walk
2. fire a tag at each stop: time, count, success or fail
3. read the sequence, not the vanity total
4. change the product where the path stalls

A survey and a tag tell different stories

Ries's line was never "pick one channel and stop." It was: stop sitting in the building with the solution I love, and go see what the customer does. A survey answers a real set of questions. Did this feel worth it? What were they trying to do? What would they miss if it disappeared? I want those answers. They're not the same answers the path gives.

A feedback form tells me what someone remembers, chooses to say, and has words for. A tag tells me what they did: which stop they reached, how long it took, how often they came back, where they dropped off. One is stated preference. The other is behavior. Both belong in the stack. The miss I keep seeing is treating the form as a stand-in for the path, as if a good question replaces watching the walk.

The industry objection writes itself. Product analytics has been around for more than a decade. Mixpanel, Amplitude, Google Analytics. Of course something is being tracked. The miss isn't the tool, and it isn't the survey either. It's that neither got wired to the journey I claimed they walk. Pageviews. A homepage bounce rate. A weekly active number that cannot tell me which stop died. A survey at the end that cannot tell me they never got past setup.

When the path is untagged, the form is the only instrument that got a name. The rest of the walk is empty.

Surveys are good at the why, the felt pain, the feature request in plain language. They're weak at the corridor: the screen they skip, the workaround they run twelve times a week, the stop where half the cohort disappears before anyone answers. That story lives in the path. I still run surveys. I just don't ask them to do a job only behavior can do.

Every stop gets a name, a time, and a count

A tag is not a pageview. A tag is a named stop on the path I claimed they walk, with two numbers attached: how long they spent on this fragment, and how often they come back. Time is duration. Not the date the event fired. The warehouse will stamp a clock because that's how events work. The number I read is milliseconds on this stop, and milliseconds to the next milestone.

Success or fail rides on the same stop, because a stop that fires and then dies is a different story from a stop that never fires.

I name the stop as the outcome, not the click. export_clicked tells me they touched a button. report_exported tells me the product did the job. Time to value is the outcome. A click is theater.

capture('first_value', {
  stop: 'report_exported',
  duration_ms: Date.now() - stopEnteredAt,
})

The fail event carries the same stop, so I can read success and death as one path, not two dashboards:

capture('core_action_failed', {
  stop: 'report_exported',
  duration_ms: Date.now() - stopEnteredAt,
  reason: error.code,
})

duration_ms is how long this fragment took. Time to the next milestone is the same shape, across stops: first_value minus signup. How long they sat here. How fast they moved. A stop they clear in seconds is doing the job. A stop that holds them for forty minutes is the ticket.

Count without duration is a popularity contest. Duration without count is one long session from one curious person. Together they are a pattern: this is where they live, this is where they stall, this is what they hired the product for. I don't need a model for that sentence. I need the events to exist.

Once the feed is thick, the same events are what a clustering job or a propensity model would train on. Users who never hit first_value by day two. A feature that fires a lot and fails a lot. A path that always detours around the screen I designed as the happy path. That is a query on tagged stops. It isn't a science project. I haven't trained the model. I would not start there. I would start with the path.

The warehouse is not the map

The capture layer is a solved problem. I don't care which warehouse, as long as it stores named events I defined, not a pile of pageviews I inherited.

Google Analytics is the one most teams already have. It answers which URL got traffic. It does not answer which job in the product succeeded. GA4 can take events. Most installs I see never fire them. A Mixpanel project with no taxonomy is the same miss in a nicer UI.

Autocapture is a bigger pile, not a map. Heap and PostHog will record every click. That is useful the day I forgot to tag a stop. It is not a substitute for naming the journey before launch. Session replay is the tape. I pull it when an event looks wrong, not to understand ten thousand users. Sentry is the failure tag when the stop is a crash. If the event name in Sentry and the event name in Mixpanel are different words for the same death, I'm back to stitching by hand.

The install is not the work. The work is the taxonomy: every stop I believe they walk has a name, and that name is the same in the product, the warehouse, and the crash report.

Usage is the story the form does not tell

A feature can be well designed, on-brand, and still be used as a corridor to something else. Or never used. Or used once, in a panic, the day before a deadline, which is a different product than the one in the Figma file. A survey can tell me the export felt slow. It cannot tell me they skip the dashboard every morning and live in the export. People cannot narrate a pattern they don't know they have.

That's the whole point of tagging the path. I want to see the corridor, the dead room, the room they never leave. Taste is how I design the stop. Behavior is how I find out what the stop actually is.

The pretty feature that nobody repeats is not a design failure I can see in a screenshot. It's a count that stays at one. The ugly workaround they run twelve times a week is the job they hired the product for. The feed shows me that. The form might confirm it later, in words I can quote in a meeting. It won't surface it first.

Time to value and dead features are already in the feed

If first_value carries how long they spent getting there, time to value is already a number. I don't need a kickoff about onboarding copy to find out they sat in setup for a day. A long duration is the path asking for work. A short one is the stop doing the job. If the event never fires, they never got value.

If core_action is the only stop with a high count, that is where the next work goes, even if the roadmap is in love with a different screen. If a stop I shipped last quarter still has a count of zero, that is a candidate to kill, not a candidate for a nicer empty state. If core_action_failed keeps pace with core_action, I know which feature is valuable and also broken, which is the worst kind of miss: people want it, and it dumps them.

Those are product decisions. Shorten the path to first_value. Stop spending design on a stop that never fires. Fix the stop that fires and fails. None of them require a model. They require the tags to have been there at launch. That's the part that keeps not happening, because by then I'm in love with the solution and the warehouse is "already set up."

People keep saying data is gold. Gold is not the Mixpanel bill. Gold is a path with names on it, read as a story about how they use the product, next to the story they chose to tell in a survey.

Takeaways

• Pageviews are not a journey. I write the path as named stops, then I fire those events.
• Time is how long they spent on that fragment, not the date the tag fired. A long duration is a stall. A short one is a stop that's working.
• A click count is theater. I tag report_exported, not export_clicked.
• Weekly active users cannot locate a stall. I read the sequence, and the stall becomes the ticket.
• A feature at count zero is not an empty-state problem. I cut it, or I follow the corridor they actually use.
• If first_value never fires, they never got value. I shorten that path before I rewrite the copy.
• A crash in Sentry and silence in Mixpanel is two names for the same death. I use one.
• A survey tells me what they say. Tags tell me what they did. I need both, and I stop asking either to do the other's job.

None of that is a tooling failure. Each one is a path that already happened, left untagged or misread, while I sat in the building with the solution I loved.