Blog · English
One Week of eStat: Who Really Reads f97.be?
380,000 page views in seven days, and at most 3,400 of them from humans. Why I built my own statistics tool without cookies and what it reveals about the machines reading my website.
One Week of eStat: Who Really Reads f97.be?
Nov 21, 2026 · @Erik
Of the roughly 380,000 page views my web server delivered in the seven days from November 14 to 20, at most 3,400 came from humans. That is less than one percent. The rest are automated: two thirds say so openly, one third pretends to be a browser.
I know this because eStat has been counting since November 12, a statistics program I built myself. This text explains why eStat exists and what the first week has shown.
Why eStat Came About
I no longer knew who reads f97.be. Since consent became mandatory, Google Analytics hardly shows any visitors, because it only counts those who click "Accept" first. Anyone who declines or dismisses the notice does not exist for the statistics.
I did not want another off-the-shelf tool. Almost all of them set cookies, store IP addresses or send my readers' data to a third-party service. But the web server is at my home, and what happens there should stay there.
There was already a first step: a small CGI script kept a text file per page with the running total. That answers the question "how often?", but not "who, from where, when and via which search term?". That is exactly what I wanted to know, especially since the energy glossary is currently being revised, translated into four languages and moved into new language folders.
So eStat was born. The first version ran on November 12 at 4:14 p.m., six more followed by the evening of the same day, and by the next morning version 1.5.3 was in place.
How eStat Counts
The web server counts as it delivers, the analysis runs on a Mac. Between them lies only an ssh connection.
- On the server, a script writes one line for every HTML page delivered: time, page, referrer, browser language, device, country and network operator. It looks up country and network in a local database.
- The IP address is not stored. From it, the user agent string and a daily key, a truncated hash value is created. The key is re-rolled every night and the old one overwritten. The same person is a different value tomorrow.
- Anyone who identifies as a bot ends up by name in a separate log and does not count as a visitor. My own networks do not count at all.
- On the Mac, the app fetches the daily files, analyzes them and additionally stores everything in a SQLite database. Added to this are the search terms from the Google Search Console.
No cookies, no consent banner, no third-party service. The figures in this text come from that database.
The Week in Numbers
In the seven full days from November 14 to 20, the server delivered 379,614 HTML pages. This is how they break down:
| Who requests | Requests | Share |
|---|---|---|
| Bots that identify as bots (104 different ones) | 257,558 | 67.8% |
| "Browsers" that enter with an internal referrer and fetch exactly one page | 55,362 | 14.6% |
| "Browsers" with bulk requests (100 or more pages a day) | 33,848 | 8.9% |
| "Browsers" from data centers | 24,936 | 6.6% |
| "Browsers" without a browser language or with one that does not exist on the website | 4,488 | 1.2% |
| Remainder: possibly humans | 3,422 | 0.9% |
The first row was sorted out by the server, the middle four are only detected by the analysis. They look like Chrome on Windows or Safari on the Mac and behave like no human does.
The "remainder" is an upper limit. How much of it is really readers is covered further below.
Finding 1: The AI Is Reading Along, and Thoroughly
The largest group of readers of f97.be are AI crawlers: 111,139 requests in seven days. Around 8,000 different pages were fetched. Arithmetically, the training and index collectors fetched each of them fourteen times in one week.
| Type | Requests |
|---|---|
| AI crawlers (Meta, Anthropic, Amazon, ByteDance, OpenAI, Perplexity …) | 111,139 |
| Search engines | 60,012 |
| SEO services | 54,836 |
| Other bots | 29,241 |
| Social media previews | 2,208 |
| AI assistants acting on behalf of a human | 92 |
Three observations on this:
- Googlebot is a supporting actor. With 3,598 requests it ranks 16th. Applebot comes to 41,565, Meta's crawler to 30,872, ClaudeBot to 24,959.
- SEO services read more than Bing and Google combined. AwarioBot, MJ12bot, SemrushBot and DotBot survey the website for customers I do not know.
- A lot is learned, little is looked up. For every 1,200 requests for collecting there is one where a human has just asked ChatGPT, Claude or DuckDuckGo something and the assistant reads one of my pages for it. ChatGPT came 53 times, almost always only as far as the home page.
Finding 2: A Third of the Automatons Wear a Mask
122,056 requests came from senders posing as a normal browser. After subtracting the automatons, 3,422 of them remain. The disguised ones appear in three forms.
The swarm. 55,235 "visitors" from 6,058 networks in 194 countries each fetched exactly one page and claimed they came from another page of my website. A human cannot do that: anyone coming from one of my pages has visited it beforehand. The connections belong to ordinary access providers, above all in Brazil (5,519), Argentina, Spain and Canada. Almost only the glossary is read: 40,245 requests, 12,779 of them on the German overview alone, plus each of the 1,074 German articles an average of 15 times. This is what it looks like when someone siphons off a website via rented residential connections, one page per address, so that no block kicks in.
The data centers. 18,002 "visitors" came from the cloud, 10,505 of them from Singapore. Alibaba and Tencent together account for 12,000 of them, again with one page per address and again almost only on the glossary overviews. Humans browse via Telekom, Proximus or Vodafone, not via a cloud server in Singapore.
The vacuum cleaners. On November 14, eStat counted 45,715 views, a good three times a normal day. Ten addresses were responsible for 33,848 of them. A single one in the Philippines fetched 13,096 pages in 134 minutes between three and six in the morning, almost 100 per minute. One from China managed 10,067 in 138 minutes. Just under half of those were for my genealogy research, almost as many for the glossary.
One side effect surprised me: the referrer lies too. On November 13, 712 views apparently came via Google, almost all from Vietnam, on the 14th another 700 from China. It was two vacuum cleaners that sent "google.com" as the referrer with hundreds of their requests. There were 45 credible views via Google in the entire week.
Finding 3: A Few Dozen Readers a Day Remain
After all filters, 3,269 visitors with 3,422 views remain, and even that figure is too high. 2,012 of them come from the USA, almost all with Chrome on the Mac and via networks such as Web2Objects, B2 Net Solutions, GTT and Oxylabs. Oxylabs rents out proxy addresses. That is the same trick as the swarm, just without the telltale internal referrer.
A more reliable view is of the countries I write for. From Germany, Belgium, the Netherlands, Luxembourg, Austria, Switzerland and France, eStat counted 208 unremarkable visitors in seven days, around 30 a day. Subtracting the hosting networks among them, it is about 10 a day.
Three clues support this order of magnitude:
- Mobile devices are missing. Only 1.4% of all "browser" views came from a phone. For human readers today it would more likely be half.
- The country list does not fit the website. At the top are the Philippines, the USA, China, Singapore and Brazil. Germany is in 9th place, Belgium with 296 views far behind. A German-language energy glossary has no audience in Manila.
- The Search Console reports similar figures. Google reports 18 clicks from search for the week of November 12 to 18.
Humans downloaded 34 files during the week, mainly the chess app Zugwunder. Bots fetched 65.
That is sobering and still the most useful figure of the week. Any statistics based on server logs that filters only by user agent would have reported 81,000 visitors to me.
Finding 4: Google Has Discovered the New Glossary
The good news is in the search data. Until November 5, Google displayed pages from f97.be around 20 times a day in search results. On November 6 there were 193 impressions, on November 18 there were 1,084.
| Period | Impressions | Clicks |
|---|---|---|
| October 29 to November 4 | 143 | 4 |
| November 12 to 18 | 4,692 | 18 |
Almost all of this belongs to the revised German glossary under /de/glossar/: 3,723 impressions on 351 articles. In October, f97.be appeared for 5 to 10 different search terms per week, now it is over 600. Searches come from Germany, Switzerland and Austria, that is, where the readers should be.
The catch is the position. The articles rank on average between 25 and 57, that is, on pages three to six. Hardly anyone clicks there, which is why 4,692 impressions resulted in only 18 clicks. Terms with many impressions are "Solarenergie", "BESS", "Lastflüsse" and "Niederspannung", all beyond position 40.
Close to the first page, on the other hand, are the specialist terms: "Residualnachfrage" in position 5, "Energy Efficiency Directive" in position 6, "Clean Spark Spread" in position 7. The glossary wins where there is little competition.
Google delivers this data with a delay of two to three days. Whether impressions turn into readers will become clear in the coming weeks.
Finding 5: The Error Page Is a To-Do List
7,299 times during the week the server answered "not found", spread across 2,376 different addresses. The list falls into two parts.
One part is intrusion attempts. At least 2,853 requests were for PHP files, WordPress plugins, .env and .git/config. Neither PHP nor WordPress runs on the server, yet the scanners knock on every door. That is background noise and confirms the decision to keep the server simple.
The other part is my own mistakes, and I would not have found them without eStat:
- The home page links to four book lists that do not (yet) exist: Reclam, comics, cookbooks and Suske en Wiske.
favicon.icois missing and is requested in vain around 25 times a day.- The ScriptAIX viewer looks for
config.jsonandscriptaix.json, both are missing. - Some redirects from the old glossary point to articles that are missing in the new one, such as CCS, self-supply and frequency containment reserve. In total this affects 16 addresses.
- Under /books/, bots run into dead ends on 360 addresses. Old pages there are still in search indexes and training data.
What Follows From This
One week of measurement has turned my picture of my own website upside down. I thought I knew too little about my readers. In fact, I knew nothing about the machines that make up 99 percent of the traffic.
For eStat, three tasks result from this:
- Humans as the default. Every figure in the app should first appear without suspicious visitors. The raw figure is the exception, not the other way round.
- Measure reading instead of requests. A small script that reports reading time and scroll depth separates humans from automatons better than any network list. A vacuum cleaner does not scroll.
- Alert on outliers. November 14 would have been a case for a notification while it was happening, not the day after.
What remains open is how I deal with the collectors. The named AI crawlers follow rules and could be locked out. The swarm from residential connections cannot. Whether I even want to shield the glossary, I do not yet know: I write it so that it is read, and these days machines read too.
Until then, I will stick to the figure that counts: a few dozen humans a day, 18 clicks from search in one week and a glossary that Google is only just getting to know. In four weeks I will check again.