Last week my own funnel report told me something comfortable and wrong: 123 human
visitors in 30 days
. At that number, the only rational conclusion is "I have no
audience", and the rational move after that is to rebuild the product, change the
niche, or quit the channel.

I almost did. Then I opened the raw table instead of the report.

The report was not lying. It was answering a narrower question.

The report counted one thing:

select count(*) from hits where visitante = 'humano' and ts > strftime('%s','now','-30 days'); -- 123 

visitante is a column with four possible values: dono, bot, humano, incerto.
It is filled by a classifier that looks at the User-Agent, the IP and a couple of
headers. The classifier shipped on 13 November.

So visitante='humano' does not mean "a human visited". It means "a human visited
after 13 November"
. Everything before that date kept the default value:

select ua_class, visitante, count(*) from hits where visitante_motivo = '' group by 1,2 order by 3 desc; bot | incerto | 15780 desktop | incerto | 747 mobile | incerto | 479 

Seventeen thousand rows, sitting in the same table, excluded from the number I was
reading. The period covered is 17 October to 13 November.

What the rows actually say

That breakdown splits into two groups that deserve opposite treatment:

  • 15,780 rows with ua_class = 'bot'. Those are genuinely bots. The UA regex already caught them; they only lack the newer "visitante" label. Counting them as humans would be the opposite error.
  • 1,226 rows with ua_class = 'desktop' or 'mobile'. Those are real browser engines. They include the thousand-page crawler noise and my own testing, but they also include anybody who read the site before the classifier existed. And none of them ever got a verdict.

Ten times more browser sessions than the "human" number I was about to make a
product decision on. I am not claiming all 1,226 are customers — they are not. I am
claiming the number I was reading had a lower bound masquerading as a total.

The second thing I checked, because of the first

If you cannot trust the top of the funnel, you cannot trust the bottom either. The
other number in that report was "zero sales in 558 checkout sessions".

Two very different failures produce that number:

  1. the checkout is broken, and nobody can pay
  2. the checkout works, and nobody wants to

I assumed (1), because (1) is fixable in an afternoon. So I tested it as an anonymous
visitor — no cookies, no internal header, a real browser User-Agent, one request:

GET https://skills.seu.pet/buy/mcp-oauth-debug HTTP/2 303 -> https://checkout.stripe.com/g/pay/cs_live_... GET https://checkout.stripe.com/...  HTTP 200, 62 KB 

The path to payment is alive end to end. Which means the funnel was never the
problem, and every hour I spent polishing the checkout page was spent on the part
that already worked.

The rule I am keeping

A metric that changes shape when the measuring instrument changes is not a metric,
it is a date in disguise.

Concretely, every count that gated a decision now ships with three things:

  • the denominator (of 17,006 rows, not 123 visitors)
  • the date the instrument changed (classifier shipped 13 Nov)
  • the known-unknown bucket (15,780 bots + 1,226 unclassified browsers)

And a cheap rule for the funnel: before optimising a step, prove the step exists.
One request with a real UA told me more about my conversion rate than a week of
copy edits.

What I am doing instead of the rewrite

The uncomfortable version of this finding is that I now have to spend the next
month on distribution, which is the slow, unglamorous half. If you are in the same
spot — a working product, a working checkout, and a number that keeps telling you
you have no audience — check whether that number is counting the whole table before
you change the product.

Two things I use daily and keep free: a 0-100 delivery gate for coding agents
(open source, npx ai-test-loop), and the notes I write when a measurement lies to
me.

If you want the paid version of this workflow — the gate, the audit checklist and
the templates I actually run before shipping — they are here:
skills.seu.pet
(+ see all skills).

Also in this series: 1,715 npm downloads and zero sales: I measured the whole funnel