Dev.to AI 🤖 Ai 👁 0 📖 7 min read

Grok Bot #1: I Set Up an Agent to Watch My Gumroad Store for 7 Days. 6 Alerts, 6 Were Noise

I sell a few n8n workflow packs on Gumroad. I don't check the dashboard every day, and when I do, I mostly see zeros. So I did the founder thing: instead of fixing the store, I built something to watch it. For 7 days (1

I sell a few n8n workflow packs on Gumroad. I don't check the dashboard every day, and when I do, I mostly see zeros. So I did the founder thing: instead of fixing the store, I built something to watch it.

For 7 days (10 to 16 November) I had a watcher agent, set up in Grok Bot, check my store every evening and tell me about sales and drop-offs. My plan was a neat post: "it caught X real things, Y were noise, here's how I tuned it."

The honest result: 6 alerts, 6 were noise, 0 were real. The store made 0 sales. And the one thing that actually went wrong that week never raised an alert at all.

This is the write-up of what the watcher saw, why nearly every alert was noise, and the rules I'd use now.

The setup

Nothing fancy. Every day at 9 PM IST the watcher:

  1. Opens my Gumroad sales dashboard (the analytics page returned a 404 on day one, so it uses the sales page).
  2. Reads views, sales, revenue and conversion for "today" for each of my 3 products: a free sample, a 2-workflow pack, and a 10-workflow pack.
  3. Appends one JSON line per product to a log.
  4. Checks three alert rules and labels each alert real or noise, with a one-line reason I can check.

The rules, written before the week started:

  • Any sale. Always tell me.
  • Views drop of 50% or more against the trailing average.
  • 0 sales with 20 or more views. That's the "people are looking but not buying" signal.

A logged day looks like this (product name shortened):

{"date": "2026-11-12", "product": "Founder AI Agent Pack", "views": 0, "sales": 0, "revenue": 0,
 "conversion": null, "views_7d": 5,
 "note": "calendar-day filter (dashboard TZ not shown), not rolling 24h; conversion not shown at 0 views",
 "alerts": [{"rule": "views drop >=50% vs trailing avg", "call": "noise",
             "why": "2 views on 11 Nov to 0; counts too tiny to mean anything"}]}

I didn't change the rules during the week. I wanted the count to be clean, even if it was embarrassing.

The week in numbers

Result
Days covered 7 (10–16 Nov)
Successful reads 6
Failed reads 1 (14 Nov)
Product rows logged 18
Sales 0
Revenue $0
Views (7-day total, best product) 7
Views (other 2 products) 0 every day
Alerts 6
Noise 6
Real 0

All 6 alerts came from the one product that got any views at all:

Date Rule that fired What actually happened
10 Nov views drop ≥50% 1 view the day before, 0 today
11 Nov views drop ≥50% 0 views vs a trailing 3 views in 7 days
12 Nov views drop ≥50% 2 views the day before, 0 today
13 Nov views drop ≥50% 2 views the day before, 0 today
15 Nov views drop ≥50% 1 view the day before, 0 today
16 Nov "views drop / no sale" 1 view, no sale

That's a "100% drop" alert five days out of six. Technically true. Completely useless.

Why almost every alert was noise

1. Percentages lie at single-digit traffic

My best product's trailing 7-day views went 2, 3, 5, 7, 7, 7 over the week. When you're getting one or two views a day, any quiet day is a 100% drop. The rule was written for a store with traffic. Mine doesn't have any yet.

2. I was comparing half a day to a full day

This one took me a while to spot, and it's the most useful thing I learned.

The evening snapshot showed 0 views on 5 of the 6 days it read. Across all 18 product rows, the snapshot views add up to 1. But the 7-day total kept climbing from 2 to 7, and the next day's alert would say "2 views yesterday".

So the views were real. They just landed after my 9 PM read, or the dashboard's "today" isn't my time zone. The log notes say the dashboard doesn't show which time zone it uses. Either way, every evening the rule compared a partial today against complete earlier days, and of course today looked like a drop.

If you're building any daily "compare today to the average" check, compare complete days only: yesterday vs the trailing week, not today-so-far vs anything.

3. One rule fired when it shouldn't have

The 16 Nov alert was labelled "views drop / no sale" for 1 view and 0 sales. My no-sale rule needs 20 or more views. So that's not just tiny numbers. The rule fired below its own threshold. The watcher marked it as noise, which was correct, but it shouldn't have fired at all. A labelled false alarm is still a false alarm.

The failure that didn't alert

On 14 November the read failed. The machine the watcher runs on had a full disk (128 GB of 128 GB), the browser couldn't start, and there was no data for the day.

What did the watcher do? It wrote a note line in the log:

{"date": "2026-11-14", "note": "Gumroad read FAILED: box disk full (/ 128G of 128G), browser could not start; no data"}

No alert. Nothing pushed to me. The next day's read filled the gap (1 view, 0 sales), so I didn't lose much. But think about what that means: a watcher that silently stops looks exactly like a store with no sales. On a store that does $0 most days, I would never have known the difference from the numbers alone.

It's the same lesson I keep relearning with my n8n workflows: the dangerous failures aren't the loud ones, they're the ones that look like a normal quiet day.

What other indie sellers seem to do

Before writing the new rules, I read through a bunch of public posts on X from people selling digital products. I didn't find a single magic setup, but a few patterns came up again and again:

  • Lots of people post daily "N views, 0 sales" logs while building in public. At the start, product page views are the only signal you have, so that's what everyone tracks.
  • 0 sales gets split into two different problems. Traffic (nobody is seeing it) vs conversion or purchase friction (people see it and don't buy). The common advice is to work out which one you have before you touch anything.
  • Conversion from cold traffic is low. One seller shared numbers from cold outreach to a paid product link that worked out to just over 1%. At 7 views a week, a 1% conversion rate means I shouldn't expect a sale for months. My zero is just arithmetic, not a broken funnel.
  • Relying on one traffic source is a known risk. A few people noted that most of their visits came from one place, and that a single source is fragile.
  • Most revenue-alert tools are built to catch revenue drops. That assumes there's revenue to drop.

Put together, it made my situation pretty clear: I don't have a conversion problem yet. I have a traffic problem. My "0 sales with 20+ views" rule, the one meant to catch the funnel leaking, never got enough traffic to fire for a real reason.

The tuned rules (replayed on the same week)

I rewrote the rules and replayed them against the exact same 7-day log:

Rule Change Alerts on the same week
Any sale unchanged 0
Views drop complete days only (yesterday vs trailing 7 days), and only when the trailing week has at least 70 views (about 10 a day) 0
Views but no sale 0 sales with 20+ views, enforced strictly 0
Watcher health (new) alert if a read fails or a day has no row 1 (14 Nov)

Old rules: 6 alerts, 0 real. Tuned rules: 1 alert, and it's the one that mattered.

The replay check is basically this:

for r in product_rows:
    if r["sales"] > 0:
        alert(r["date"], "sale")
    # drop rule: only with real traffic, only on complete days
    if r["views_7d"] >= 70 and r["views"] <= 0.5 * r["views_7d"] / 7:
        alert(r["date"], "drop")
    if r["sales"] == 0 and r["views"] >= 20:
        alert(r["date"], "views, no sale")

for r in failed_reads:
    alert(r["date"], "watcher health: read failed")

Two other changes I'm making:

  • A weekly digest instead of daily pings while traffic is under about 10 views a day: total views, sales, and where the views came from. Daily numbers at this scale are just noise with a timestamp.
  • Read yesterday, not today. The evening snapshot of "today" was the root of most of the noise. Reading the previous complete day removes the time-zone guessing.

What I'd tell myself a week ago

  1. Write alert rules for the traffic you have, not the traffic you want. Percentage rules need a minimum volume, or they fire on every quiet day.
  2. Compare complete periods. Today-so-far vs yesterday is a drop alarm that goes off every evening.
  3. Monitor the monitor. The most important alert of the week was "I couldn't read the data", and my rules didn't have it.
  4. A labelled false alarm is still a false alarm. It was nice that the watcher explained why each alert was noise. It's nicer to not get them.
  5. The watcher can't fix a traffic problem. It just told me, very accurately and six times, that nobody was visiting. The fix for that is distribution, not monitoring.

I'll rerun the watcher with the tuned rules and report back. The watcher itself is not in my GitHub repo. The free n8n workflows I write about here are: https://github.com/arkintof/n8n-ai-agent-workflows

If you sell anything small: how do you track your store? Do you have alerts at all, or do you just open the dashboard when you feel brave? And has your monitoring ever quietly stopped without telling you?

Written with AI assistance; workflow built and tested by me.

📰 Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.