At half past eleven at night, I opened paulkuo.tw and glanced at the traffic section.
Visitors: 0.
That could not be right. People had read articles and used the translation tool that day. Cloudflare Dashboard showed Page views 193 and Visits 130.
Another analytics page then showed 1,100 Unique Visitors. Three very different numbers, on the same day, for the same website. I rushed to find an explanation that fitted them together.
My answer then was that sampling caused the zero, 130 meant humans, and most of the remainder were bots. Revisiting this article, I need to withdraw that judgment. The numbers were observations from that time; the explanation did not have equally complete evidence.
Visits are not a headcount
I naturally read Visits 130 as “130 people came today.” I had already changed the unit.
Under Cloudflare’s current metric definitions, a Web Analytics Visit is a page view originating from another website or a direct link, with the referrer’s hostname involved in the determination. A visit can include multiple page views. This does not establish each person’s identity and count their heads.
Likewise, IP-deduplicated Unique Visitors cannot simply be treated as individual people. An office can share a public IP; one person switching networks can appear under another IP.
I was therefore dealing with at least two units: visits and visitors deduplicated by a particular method. Labeling both “people” muddled their meaning before comparison began.
The values here are those recorded at the time. I have not reconstructed the historical queries. Current official documentation helps correct my interpretation; it cannot replace the original request records.
Browser reports and network requests cover different ground
Cloudflare Web Analytics reports data through a JavaScript snippet on the page, often called a beacon. HTTP traffic analytics instead observes requests passing through the network service. Those different vantage points can cover different activity. Web Analytics documentation
My earlier assertion that “bots do not run JavaScript, so beacons count only humans” was too strong. Executing JavaScript does not prove that a reader is sitting behind a browser; automated browsers can also load pages. Conversely, a human reader’s report can fail to arrive because of browser settings or network problems.
I switched from network analytics to browser reports to obtain a metric closer to readers’ use. That goal remains reasonable. But successful reporting must be defined, along with its blind spots; choosing a tool does not guarantee accuracy.
Sampling is an investigative lead, not a complete explanation for zero
During troubleshooting, I noticed several zero API values over a period and many multiples of 100. I connected that pattern to adaptive sampling and even wrote that “130 was rounded down to zero.”
That statement does not work. Ordinary rounding to the nearest hundred would not turn 130 into zero. More importantly, the shape of the values cannot establish what sampling was used, at what rate, or whether another issue was involved.
Cloudflare’s sampling documentation explains that some Dashboard and GraphQL figures estimate totals from samples. We therefore cannot assume that Dashboard is always exact while only the API samples. Current documentation also describes confidence-interval queries for some Adaptive datasets; applicability still requires checking the dataset and fields.
For this historical incident, the supportable conclusion is that the API and display differed substantially and my original acceptance checks were inadequate. Establishing the cause requires query records aligned on timezone, time range, hostname filters, dataset, and metric, followed by checks on sampling and update delays. Without those records, the root cause remains unresolved. A plausible technical term cannot fill the gap.
Three wrong turns
First, I checked only the fields. I confirmed the API returned Visits without establishing how it was calculated or comparing a few days with the display. A number being available made me think it was ready for decisions.
Second, I rushed to change data sources. Seeing zero, I assumed another API would do better, pushed a revision, then reverted it. The back-and-forth cost time without supplying what mattered most: how the two queries differed.
Third, I conflated human detection with numerical quality. Even if a system filters automated traffic, completeness, deduplication, and sampling still require separate checks.
I often remind myself to investigate before acting when collaborating with AI. A zero on the homepage still made me want a quick fix. The useful lesson is to stop that impulse and preserve the problematic query and result first.
Counting for yourself still requires a definition
I later added a custom visit beacon: page loads report to a Worker, with daily deduplication based on IP plus User-Agent. User-Agent is software-identification information sent by a browser. Here that combination is only a counting key, not proof of a person’s identity.
A custom counter lets me define events, retain troubleshooting records, and understand how a figure is produced. It still sees only reports that arrive. Deduplication can merge different people or split one person. Hashing alone also cannot guarantee full anonymity; retention and use matter.
I went one step further in the original article: subtracting beacon counts deduplicated by IP plus User-Agent from Zone Analytics IP-deduplicated counts and calling the difference AI/bot visitors. That calculation is invalid. The units, coverage, and deduplication differ. There is no basis for assuming one group is exactly the human subset of the other.
Thus, 1,100 minus 130 equals 970 is only arithmetic. It cannot prove there were 970 bots, much less establish that machines accounted for 88% of the website’s visitors.
To investigate automated access, inspect request records and the basis for identification separately, explaining what can and cannot be recognized. A crawler fetching an article also does not establish that a model cited it or that the content had an impact. Those are different questions.
What to preserve the next time you see zero
Reviewing this experience, I would first write a definition for each metric before deciding whether it belongs on the homepage.
Browser-report trends, network-request changes, and article-generated inquiries each have a use. They can sit alongside one another. They should not be forced into one human-versus-machine pie chart just to make a neat picture.
If Dashboard and API results differ greatly on your site, start small: select one time period and record the timezone, filters, metric name, and full response. Confirm that both sides ask the same question before investigating update delays, collection gaps, and sampling. Building your own tool is an option, not something every small website must do.
That night’s zero made me rush to prove somebody had visited. Looking back, I would rather first understand what the word “visitors” on the homepage promises readers.


💬 Comments
Loading...