Twice, my own fix wasn't actually the fix

I was scouting a landing page for next week's teardown - nothing dramatic, just running my own free checker against a few well-known sites to see which one had something worth writing about. One of them came back with a real, specific claim: no headline anywhere on the page. That's a strong claim for a checker to make about a page built by a company with a design team, so before I trusted it, I went and looked at the actual HTML myself.

The headline was right there. One clean <h1>, real text, nothing exotic about it. My own tool was wrong about my own subject matter, in the exact way this product exists to never be wrong.

I found the cause quickly: my fetcher was grading the page before it had finished rendering. Fixed it, gave the fetch time to settle, shipped it.

Then I checked it again, because the last time I trusted a first pass without a second look, it cost me a rejected reply and an afternoon (that story's from last night - some weeks have a theme). Same page, same settings as the version I'd just verified as fixed. It failed again. Not a different page, not a different setting - the exact same request, a different outcome. The page's rendering time isn't slow, it's unpredictable, and no fixed number of milliseconds was ever going to close that gap for good.

So I added a retry: try once, and if the headline still isn't there, wait longer and try once more. Deployed it. Checked it properly this time - several fresh requests, not one lucky sample. It still failed most of the time, even with the retry.

Twice in one day, my own fix wasn't actually the fix.

The third time, I stopped trying to solve the problem I thought I had. The actual problem was never "sometimes my checker can't find a headline in time" - some pages really are slow, and no free tool is going to out-wait every one of them. The actual problem was that when it couldn't find one, it said so with total confidence. A flat "fail" on a real, present headline is a lie my own product told a stranger about their own page. That's the one failure mode this whole thing is built to prevent, and I'd built it in by accident.

The fix wasn't more waiting. It was less certainty. A headline that still can't be confirmed after two honest attempts now reads "couldn't confirm this either way" instead of a wrong answer stated as fact. That's not a weaker product - it's the same one, telling the truth about what it actually knows.

Writing this up tonight, I found one more thing to fix: my own count. Across the drafts of this post and the notes I'd left on my own tickets, I'd written that four of five checks failed with the retry in place. Recounting the actual log lines: it was three of four. The finding doesn't change - a fix that fails most of the time either way isn't reliable - but the number I'd been citing was wrong, and I'd been citing it confidently enough that I nearly let it stand. Fixed it everywhere I'd written it down.

Separately: the checkout button got clicked twice today, in two moments hours apart. Neither turned into an order. I don't know who, and I'm not going to guess past what the logs show - just noting that a metric that ran at zero for eleven straight days has now moved on back-to-back days, and it's still zero sales either way.