DEV Community

Ted
Ted

Posted on Originally published at tedagentic.com

Nothing Failed. Things Just Stopped Being Looked At.

I run a travel directory that has grown from one region to six. Each region has its own pages, its own listings, its own prices, and its own list of open questions waiting on someone: a host who hasn't replied, a price that needs re-checking, a page that should be built once enough listings are confirmed.

The site has a lot of monitoring. Uptime, broken affiliate links, schema validity, a weekly ranking tracker, a weekly re-read of every listing's rules. On a normal Monday the alerts are quiet, and quiet felt like healthy.

Then I spent one day actually walking through every region, and found this:

  • One region's search clicks were down 25% week on week, after its rankings had been slipping for two weeks. The ranking tracker had seen the drop, flagged it once as "waiting for confirmation", and moved on.
  • Three pages were quoting nightly prices "for Tuesday 10 November". It was 16 November. The prices had been accurate when they were checked. They had simply expired.
  • A page that had been retired months earlier still carried three placeholder listings in its code, with made-up ratings. A redirect hid them. Remove the redirect and they'd be live.
  • Four regions had been researched, scored and planned in October. Nobody had touched them since. They existed only in a notes file.
  • Three listings had been quietly removed from the platform they were booked on. Their pages still recommended them.

Not one of those threw an error.

Neglect doesn't fail. It just stops being looked at.

Every monitor I had was asking "is anything broken?" None of these things were broken. A stale price is a correct number with a date attached. A forgotten plan is a perfectly valid file. A slow slide in clicks is a series of normal-looking days. The system was doing exactly what it was told. What had gone missing was attention, and attention doesn't have a status code.

The failure was memory, not work

With one region, I didn't need a system. I knew what was open because there was only one thing to think about. By six regions, "what's the next task?" had quietly turned into "what do I happen to remember?", and the answer always favoured whatever I'd touched most recently.

So the fix wasn't another monitor. It was a register: one file, one record per region, where everything that can go stale carries a date.

{
  "id": "region-a",
  "status": "live",
  "price_pages": [
    { "page": "/region-a/cabins", "checked": "2026-11-16", "for_date": "2027-01-12" }
  ],
  "last_full_check": "2026-11-16",
  "next_actions": [
    { "what": "Judge the retitle on the weekly tracker", "due": "2026-11-30" },
    { "what": "Follow up the three silent hosts", "due": "2026-11-18" }
  ]
}
Enter fullscreen mode Exit fullscreen mode

Three fields did most of the work.

for_date, not just checked. I had always recorded when a price was checked. That isn't what makes it wrong. A price checked yesterday for a stay last week is already useless. The register stores the date the price is for, and flags any page where that date has passed or is two weeks away.

due on every open action. "Follow up the hosts" isn't a task until it has a day attached. Without one, it sits in a notes file and becomes a fact about the past.

last_full_check plus a rotation. Once a week the oldest region is due a full pass: re-read every listing, re-check prices, look at its rankings. Six regions, about one a week, so nothing goes more than a month without someone actually looking at it.

A message that only lists what's due

Every Monday a plain script (no AI, about 200 lines) reads the register and posts one message. It joins in live data: listing counts from the database, search clicks week on week, the ranking tracker's last two runs. Then it lists only what needs attention:

  • clicks down 20% or more week on week
  • a ranking drop of three places or more
  • prices with no check date, or for a date that has passed or is close
  • actions due within three days, or overdue
  • regions still being built, against how many confirmed listings they need
  • this week's rotation

Regions with nothing due go on one line at the bottom: "All clear: …". That line turned out to matter. Without it, a silent region and a forgotten region look identical. With it, every region is accounted for every week, either in a list of problems or in the all-clear.

The first run flagged two price pages with no check date at all, one of them still showing two prices that had been re-checked that morning on a different page. Same listings, same week, two prices. Nothing about that was broken either.

Making the register the source of truth

The register only works if it's the first place I look and the place every change gets written down. So two rules came with it:

  1. "What's the next task?" gets answered from the register, not from memory.
  2. Any work that changes a region's state updates the register in the same sitting: a price re-check, a full pass, a new region, a finished action.

A second script reads the same file every morning and puts the top three due items across all my sites in one short briefing. That covers this blog too, which, it pointed out, hadn't had a post in fifteen days.

What I'd tell myself at region two

Monitoring answers "is it broken?" Running more than a couple of things also needs an answer to "when did someone last look at it?", and that answer has to live in data, not in your head.

The cheapest version is a file with a due date on everything that can go stale, and something that reads it every week and says out loud what's overdue. The overdue items had been there all along. The difference is that now something tells me.

Top comments (0)