My Blog Report Said 358 Post Views. Some of Them Weren’t Posts.
I asked a simple question on November 6: were people finding my technical posts? My access-log report answered with 358 “post views” over the previous seven days. The list included my contact page and a privacy page.
The requests were real log entries. The label was wrong. I’d classified a URL shape as a content type, then used the result to judge editorial demand.
I caught the mismatch before choosing my next batch of topics. Otherwise I’d have let pages that weren’t articles influence what I wrote next.
What the first counter actually counted
My report filters known bot user agents, common scanner paths, static assets, and WordPress administrative endpoints. It then treats a long, hyphen-compatible path ending in a slash as a post.
That shortcut doesn’t consult WordPress’s post type or publication status. A contact page fits. A privacy page fits too. The code wasn’t secretly measuring sessions; I was giving its output a stronger interpretation than it deserved.
For the second pass, I fetched the actual published-post paths. There were 32. I merged the available archive and container logs, removed identical log lines, and matched successful GET requests against those paths.
| November 6 analysis | Previous seven-day window | Latest seven-day window |
|---|---|---|
| Requests matching published articles after known bot/scanner filtering | 287 | 294 |
| Requests with a recognized external search source | 3 | 6 |
| DuckDuckGo source | 2 | 6 |
| Bing source | 1 | 0 |
| Google source | 0 | 0 |
These are the windows analyzed that morning, not today’s traffic totals. They’re rolling windows rather than calendar weeks. The article-only query also used the merged, deduplicated records, so I don’t present 358 minus 294 as an independently audited count of excluded fixed pages.
Six requests gave me a direction, not a growth curve
The latest six search-source requests went to two articles. Four targeted my CPU expert-offload tuning post. Two targeted my Qwen3-Coder setup on an RX 9070 XT.
That’s a useful editorial clue. I have some evidence that those specific problems attract search referrals. I don’t have enough evidence to say all local-LLM topics are popular, or that traffic has entered sustained growth.
I also won’t call three becoming six a proven doubling of my audience. The sample is tiny. A request isn’t a unique person, and a referrer header isn’t a verified human identity.
The other 288 article requests in that latest window didn’t all come from new readers. Of the 294 total, 262 had no attributable external source and 26 referred from within the site. I can’t turn the first group into loyal direct visitors by changing its label.
The visits my server can’t see
My site sits behind Cloudflare. When a cached response is served without reaching the origin, that visit doesn’t add a corresponding page request to my WordPress access log. Cloudflare’s cache documentation explains that cached copies reduce origin traffic.
This means my origin count can miss real page deliveries. At the same time, it can include automated requests that my heuristics missed. Those two limitations don’t cancel each other into a trustworthy audience number.
I’d also inspected pages myself during publication and layout checks. A normal browser used for operational checks can look like a reader. I didn’t have a complete, reliable exclusion list for all of that activity.
I can’t recover session duration, bounce behavior, or returning-reader counts from these aggregates. I haven’t queried Search Console’s search terms for this analysis either. Search-source requests don’t reveal the actual query someone typed.
Don’t turn a crawler into a recommendation
My script has a separate section for AI crawler names. Those requests aren’t evidence that an assistant recommended my article to someone. User-agent strings can also be imitated.
Even a request carrying an AI-related referrer or campaign parameter needs careful wording. It records a claimed source; it doesn’t prove a particular answer contained a citation or that a human clicked it.
I haven’t rewritten the reporting script as part of this analysis. Its broad “post views” label remains a known limitation. The figures above came from the separate article-path comparison, not a claim that the daily report is already corrected.
For now, I use the result to prioritize concrete GPU troubleshooting subjects. I keep the sample size beside that decision. My CPU offload article and AMD setup record earned that attention through observed referrals, not an invented keyword-volume estimate.