DEV Community

Cody Wang
Cody Wang

Posted on Edited on

My hit log counted 36 humans. One row came from Google.

I ship one sales page a week. T1 Harbor Counsel went up on 9 November, T2 Northmark on the 12th, T3 Summit Legal on the 16th. Before writing a fourth, I did the thing I'd been avoiding: I instrumented the site and looked at who actually arrives.

The logger is about thirty lines in a Next.js route handler. It appends one JSONL row per pageview: timestamp, path, referrer, viewport width, language, a user agent cut at 120 characters, and a boolean from a regex. No cookie, no IP, no third-party script. That last part is a decision with a cost, and the cost is the rest of this post.

Eleven days of rows: 45. Nine are flagged as bots. Thirty-six are not. That number felt good for about ten minutes.

Then I sorted by timestamp. Twenty-one rows arrive inside nineteen seconds, all from one user agent, all at a 1919px window: /, /contact-us, /about-us, /team, /imprint, /impressum, /legal, /privacy, /terms, /privacy-policy, then seven template pages in under four seconds. Nobody opens impressum and privacy-policy 200ms apart. It is a link checker, and it got counted as a human twice over, because I had quietly treated "not flagged" as "a person".

Why did it slip through? My filter is /bot|crawler|spider|headless/i, and the user agent was a plain Chrome string on X11. A cheap filter only catches bots that confess in their name. The ones that cost you anything don't.

The mirror case is worse. Nine rows are flagged and, when I open them, the stored user agent ends at (compati, cut at the 120-character limit with the bot token past the cut. I have rows I know are bots and cannot name. I traded storage bytes for diagnosis and lost the diagnosis. Storing the matched token instead of the raw string would have cost fewer bytes and named the crawler.

Then there is the block of eight rows on 17 November: 15:49:01 to 15:50:43, all to /, one user agent, Chrome 114 on a Mac, every one at 1280px. My own capture scripts render at 1440. Either someone on an old Chrome reloaded the homepage eight times in 102 seconds or a monitor did, and no field in my log can tell me which. Without a cookie or an IP, a first-party log cannot subtract itself. That is not a bug I can patch; it is the price of the design.

Two of the seven remaining rows are self-referrals from my own other domain: internal navigation, not acquisition. One arrives from google.com on 17 November, en-GB, landing on /templates/rooks-salon, a template detail page, not the homepage. Two more, on the 18th, are a Mac at 1470px with a Chinese locale going homepage, then /contact, 34 seconds later. That sequence is the only one in eleven days that looks like a person reading.

So the honest tally for eleven days of work is three rows that plausibly belong to strangers, one of them from search, and it landed on a page untouched by any of the three sales pages I was proud of.

Which changes what I do next. I was going to spend this week on a fourth sales page in the same shape, a fourth variation of stat band, three lanes, process, intake. The rows say the bottleneck is not the sales page, because nothing points at one. The only entry point search bothered to use was a detail page. So the next block of hours goes into the twenty-five template pages already in the sitemap of Slice Lab — real titles, and one honest paragraph each about what the layout argues — instead of a fourth landing page nobody enters through.

I know forty-five rows prove almost nothing. That is the other half of the tradeoff: a log small enough to read in one sitting is too small to test anything. I am trusting it for direction only, and the direction is to stop polishing the door nobody walks through.

The same measurement problem is what I chase on client sites, with header and render checks instead of logs — that work lives at prismicaudit.com.

Top comments (0)