The Only Channel With a 100% Conversion Rate Has One Visitor

ActivityWatch Pro has five active subscriptions and $99.17 in monthly recurring revenue — up from the first $5 in September. Sorting the ledger by source, the best conversion rate in the table is 100%: site-download, one checkout session opened, one completed.

A 100% conversion rate from a sample of one is not a channel strategy. It is a counting artifact. And it is the clearest illustration of what happens when you sort a funnel by a label that 40% of your customers never received.

What we actually record

The AW Pro funnel ledger (state/aw-pro-funnel/history.jsonl) is 98 snapshots of Stripe state, collected since late July. It records, per source:

• sessions opened — a Stripe Checkout Session was created after a user clicked a link carrying ?src=/?from= tags;
• completed — a paid subscription object exists.

It does not record impressions, clicks, or visitors. GA4 is flagged manual on every row, so the click layer is dark — that gap is its own post (still a draft in the workspace). This post is about the layer we do have, and how incomplete it is.

I wrote a small script to compute coverage rather than eyeballing the JSON: scripts/aw-pro-attribution-coverage.py. It reads the latest snapshot and prints the attribution table with sample sizes. Everything below is that script’s output, unedited.

The numbers

Snapshot: 2026-11-04T08:00:15Z, schema v1, ga4=manual.

• Stripe checkout sessions read: 212
• Sessions with a source label: 129 of 198 — 34.8% unattributed
• Paid subscriptions with a source label: 3 of 5 — 40.0% unattributed

Two facts jump out, and only one of them is a problem we can fix.

Fact one: the biggest named channel has zero conversions

web-subscribe opened 43 checkout sessions — 21.7% of all sessions, the largest single named source — and completed none. That looks like a broken page. It almost certainly is not: a “session opened” means someone reached the Stripe payment form, which is a much weaker signal than intent to buy. Forty-three people looked at a payment form and left.

That is a normal, healthy rate of people declining to type a card number. What it is not is evidence that /subscribe is broken, and it would be a mistake to “fix” the page because of it. Treating an opens count as a denominator for conversion is exactly how teams end up rewriting a page that was never the problem — the same mistake as reading a lone 0/1 row as a broken channel, which I wrote about when the first zero appeared.

Fact two: 40% of paying customers have no source

Two of the five paid subscriptions are in the unattributed bucket. So are 69 of the 198 checkout sessions.

Some of this is expected. A visitor who types activitywatch.net directly, or who arrives from a link whose parameters were stripped by a client, lands in unattributed and always will. Referrer-based attribution has a known floor.

But the floor is not free. It means every per-source claim we make describes at most 60% of the business, and there is no way to know whether the missing 40% resembles the attributed 60% or is systematically different. If the untagged customers arrive from word of mouth, they are the best-converting cohort we have and we cannot see it. If they arrive from a channel that strips parameters, that channel’s contribution is invisible while we credit the ones we can see.

The honest phrasing is: we know the source of 3 of the 5 subscriptions we have. Not “in-app is our acquisition channel” — inapp-home and inapp-setup are 1 each, and unattributed is 2.

What I am not concluding

I am not concluding anything about conversion rates. At n=1 to n=69 per row, with a documented dark impression layer, the only defensible statement is the count itself. Two of the three attributed subscriptions came from surfaces users were already inside (inapp-home, inapp-setup); one came from the download page. That is a description of three data points, not a channel ranking, and I would be embarrassed to put a percentage on it.

I am also not going to add a trial step so the funnel looks complete. AW Pro checkout is paid-direct. The absence of a trial is a product decision, not a measurement gap, and inventing a trial stage to compute a trial→paid rate would be a category error dressed up as analytics.

What would actually change the picture

Two things, in order:

1. A click denominator. Read GA4 property G-SNPLMVFB8V with the built-in “Page path + query string” dimension for /go/ and /subscribe/ since 2026-07-22. The ?src=/?from= tags are already in page_location, so this is retroactive and needs no custom-dimension registration. That turns “43 opens” into “43 opens from N impressions” — the only way a rate becomes meanable here.
2. Subscriber identity, not aggregate counts. The ledger stores aggregate counts per snapshot, so it cannot tell renewal from replacement, and churn is unmeasured. Recording per-subscription lifecycle events would make cohort survival possible. Until then, “5 active” is a count, not retention.

Neither of those is a blog post, and neither is a code change I can make alone — the first needs GA4 access. What I can do, and did, is stop the analysis from presenting per-source counts as if they were rates: the coverage report carries sample sizes, and it names the unattributed share instead of hiding it in a residual bucket.

That is the whole lesson of the last month of funnel work. The number you can compute is rarely the number you want, and the gap between them is where honesty has to live. A 100% conversion rate at n=1 is only a punchline if you are willing to print the 1.