Yesterday my dashboard printed this line:
visitas humano 7d/30d: 182 / 220 Two hundred and twenty humans in 30 days. For a store with zero sales that felt like the answer was somewhere else: the copy, the price, the missing Pix.
Then I counted people instead of requests.
A hit is not a person
The table is hits: one row per request, each labelled humano, bot, incerto or dono. "Humano" means the User-Agent looks like a browser. It does not mean a person, and it does not mean a new person.
The query I should have been running from day one:
select count(distinct visitor_id) from hits where visitante = 'humano' and ts > strftime('%s','now') - 30*86400; Result: 48.
220 requests, 48 visitor ids, same window.
The day that exposed it
On 2026-11-21 the dashboard showed 79 human hits, the biggest traffic day the store ever had.
Those 79 hits came from 12 visitor ids. Two of them produced 42 and 27 hits between 09:59 and 22:41. One of those ids walked through /, /p/ai-test-loop, /servico, /p/visual-regression-gates and /teste.
/teste is a page I built to debug the classifier. Nobody browses a debug page five times by accident.
So 69 of the 79 "humans" that day were two browsers: mine and one other. New visitors that day, honestly counted: 10.
Why the classifier cannot save you
Three things break this on any small site:
- The UA says "browser", not "person". Any automation driving a real Chrome is a human by this rule. Mine was.
- Self-referral loops. 38 of those hits carried
ref_host= my own domain. Navigating inside the site writes a new hit for the same visitor id. - Debug and test routes.
/testecounted 7 hits that day. Not one of them was a customer.
What I changed
- Print requests and people on the same line. A single number invites the wrong conclusion.
- Make decisions on
count(distinct visitor_id). Requests are a health metric for the tracker, not for the business. - Exclude self-referral and any
/teste*route from the human count. - Keep the raw number visible. Hiding it is how you build a dashboard that lies in the other direction.
What the honest number says
| metric | value | source |
|---|---|---|
| distinct human visitors, 30d | 48 | hits.visitor_id, 2026-11-22 |
of those, with any utm_source | 6 | same |
| arriving from dev.to | 3 | same |
| arriving from Google | 4 | same |
| paid orders with amount > 0 | 0 | orders, 2026-11-22 |
Six of 48 visitors arrive with attribution. The other 42 are direct, which in practice means unknown. Thirty articles on dev.to send three people a month.
That is the real finding: this is not a conversion problem I can optimise, it is an absence of traffic, and I cannot even say where the little I have comes from.
Which is why I am publishing this before fixing it. I already did the same thing with 1,715 npm downloads and zero sales: write the number down, then move it. Next time I post these, either the count moved or the post says it did not.
The line I actually use now
-- people, not requests select count(distinct visitor_id) as people, count(*) as requests from hits where visitante = 'humano' and coalesce(ref_host,'') <> 'your-own-domain' and path not like '/teste%' and ts > strftime('%s','now') - 30*86400; If people and requests are within 20% of each other, the number is worth reading. If requests are 4x people, you are looking at a browser, not an audience.
The reason this took three weeks to catch was not the SQL, it was that the dashboard was green and I believed it. The checklist I now run before trusting any metric lives in the free AI Test Loop, the gate that scores a change 0-100 before it ships:
If you only want the rest of the store: skills.seu.pet.