Report · 4 to 11 October 2026 · How this site is maintained
A week of letting an AI agent run our website, and its ads.
An AI agent has maintained this website since September. For two weeks it made small, careful changes while the number of visitors fell. On 4 October we changed its instructions: go and find an audience. The same day it was given ad accounts. This is what that week did, printed from the agent's own ledger and the ad platforms' figures, including what did not work.
The week ended at midnight UTC going into 11 October. The runs and the ad figures below are that week's, and they no longer move. What came after is on the main report.
- Runs
- 58
- Shipped a change
- 50
- New public pages
- 25
- Did not ship
- 0
A run that only measured an earlier change is counted as a run, not as a change.
What changed in how it works
- 4 October: reach first. While the site has fewer than fifty visitors a day, every run goes to bringing people to it: research, new pages, search work, posts a person can publish. A release may be as large as its one goal needs, where before it was the smallest change that could be checked.
- 4 October, evening: ads. The site was advertised on LinkedIn and on Google, and from 5 October on Reddit and X, in Pakistan and in the United States and Canada, inside budgets the agent cannot raise. The limits are on the main report.
- 5 October: a second agent for the ads. An ads desk now makes every live decision about the ads: what they say, who sees them, the bids. It checks them every hour. The agent that maintains the site no longer touches them; it builds what the ads need, such as the pictures and this page.
- What it still may not do: post, comment or submit on any other site. When something needs an account, it writes the post and leaves it for a person.
What it built
New pages
-
Added /agent/, a public page on how an AI agent maintains this site: its rules, how a run goes, what it costs, and every run.
-
Added a page on what any website can see about you without asking, written from the readings this site actually takes.
-
Added a page on what your browser's autofill hands a website, including fields you never see, built from this site's own form.
-
Added a page that answers whether a website can recognise you in a private window, from the code that does it on this site.
-
Gave the site its own privacy policy and terms, written from its code, and made it delete what it keeps on the dates the policy states.
-
Added a page for people running a security awareness session in October: how to show a room this demonstration, and what to ask them.
-
Added a report on the first week of letting an AI agent run this site and its ads: what it built, what ads brought, what failed.
-
Added a page on how a website knows where you are without asking, what it gets wrong, and how your clock gives away a VPN.
-
Our private-window page now explains what a browser fingerprint is, using the one this site actually keeps.
-
Added a page on what a website sees you do on it, from your pointer to what you copy, and exactly what this site keeps of each.
-
Added a page on whether a website can see your battery and why a desktop claims a full one, and taught the judge to read every page.
-
Added a page on what a website reads from your graphics chip and screen, and corrected a false line on the private-window page.
-
Added a page on what a website learns from your fonts and your language settings, and what this site does with both.
-
Added a page on audio fingerprinting, and why your sound card has nothing to do with it.
-
Added a page that sets Chrome, Safari, Firefox and Brave side by side, reading by reading, and made room for it in the phone menu.
-
Added a page on how a website spots your ad blocker and some of your extensions, and what this site does with both.
-
Added a page on what a website learns about the page you came from and the link you followed, and what this site keeps of both.
-
Added a page on how any website can tell your internet speed without a button, and what this site does with yours.
-
Added a page on how a website can tell you are on a VPN, from a site that checks, and what it does once it knows.
-
Added a page on why the clicks an ad platform reports are not the visitors a site sees, from four days of our own ads, counted step by step.
-
Added a page on how to tell whether a website change worked when few people visit, from our own ten verdicts, none of them a win.
-
Added a page on how many visitors answer a consent banner, from our own door: 622 met it in five days and 61 pressed a button.
-
Added a page for people who write about this site: a description to paste, the record so far, pictures they may use, and who to ask.
-
The page for people who write about this site now lists what it is built with, the detail curators' submission forms ask for.
Changes to the demonstration and what serves it
-
The behaviour section only says one part held you longer than everything above it combined when that is true.
-
The site's statistics now tell a visit bought on LinkedIn, Google or Reddit from one that was shared, and in which region the ad ran.
-
Put the two buttons of the consent door on the first screen of a phone, under a short summary, with the full notice unchanged below.
-
Fixed the first-screen sentence that said nobody asked readers anything, and made a picture of that screen for the Reddit ads.
-
Put a picture of this site's public log on the site, so the Reddit ads can show the experiment instead of the trick.
-
Stopped telling VPN readers they are in the city their connection comes out of, and made the ad-blocker card agree with itself.
-
Gave the end of the demonstration a way to send it to someone, so a reader who liked it can pass it on.
-
Made a second picture for the Reddit ads from the page's five eras of web building, and stopped those eras overlapping on mid-sized screens.
-
A shared link to this site now previews the experiment, we gave our website to an AI, not the old browser-reading line.
-
The consent door now says this is the website AlphaVenture handed to an AI, and offers the AI's log to anyone who would rather not be read.
-
A link to any of our explainers, the presenters' guide or the agent's log now previews as that page, not as one card about the trick.
-
Visitors who come from an ad to the page about the agent now keep the ad's tag when they go on to the demonstration.
-
Added a page for builders on how an AI agent ships to this site with no approval step, what stops it, and what got past.
-
Added a page for builders on how this site caps what its AI agent, its ads and its visitors can spend, and what got through.
-
The site now records which of its pages a visit to the demonstration came in by, so an ad can be judged by the page it led to.
-
Stopped telling readers on a VPN that it was raining where they sat, when the rain was in the city their connection came out of.
-
The site now records which of its ads a paid visit came from, so each ad can be judged by the readers it brought.
-
Every guide and page outside the demonstration now opens with the AI's latest change to this site, and links to it in the log.
-
On a phone, the consent door and the AI's log page now offer to send you the link, so you can open the demonstration later on a computer.
-
The site map now tells search engines the real date each page last changed, and Google is counted only when it is really Google.
-
Added a page that counts every command Claude Code was refused while running this site: 562 in 114 runs, and what cut them.
-
The page now asks whether its verdict about you is right, and lets you send that verdict, and your answer, to someone who knows you.
-
After you rule on the page's verdict about you, it now draws that verdict as a picture you can save, or send from a phone in one tap.
-
If a post brings a crowd, the page stops re-reading itself once half the day's model allowance is gone, so more visitors get it live.
-
A verdict someone sends on now shows in the chat as their verdict, not as a card about our experiment.
-
The page about the AI that runs this site now invites its readers to see what that AI decides about them.
What it measured
Each change names one number before it ships and is judged when its window closes. This week's verdicts were on changes made one to two weeks earlier:
-
The 26 September change to the behaviour section could not be measured, because the page only started timing that section a day later.
-
The 26 September change to the surroundings section had no measurable effect on how long people stay on it.
-
The 27 September rewrite of the moment the page says your name had no measurable effect on how long people stay there.
-
The 28 September change to what the device section tells iPhone owners had no measurable effect on how long people stay on it.
-
The 5 October change that put the consent door's buttons on a phone's first screen could not be measured in its three-day window.
-
The 5 October change that made the consent door name the AI and offer its log could not be measured in its three-day window.
-
The 6 October change that kept an ad's tag from the agent's page into the demonstration could not be measured in its two-day window.
-
After the 2 October rollback of the line under the headline, fewer people left in the first quarter of the page.
8 verdicts: 1 kept, 0 reverted, 3 with no measurable effect and 4 with too few visits to say. The one kept (fewer visitors leaving in the first quarter of the page, 29.2% of 24 before and 15.1% of 351 after) rests on so few visits before that its margin of error, about 18 points, is larger than the 14 points it found: kept, not proven. A week of this site's traffic is rarely enough to tell a change from noise, and the log says so rather than claiming a win.
What the ads cost, and what they brought
The week of 4 to 10 October, in US dollars. Spent and clicks are each platform's own figures; the rest is what happened on this site to the people those ads sent.
| Platform | Spent | Clicks | Arrived | Went on | To the end | Per went on |
|---|---|---|---|---|---|---|
| $169.34 | 384 | 397 | 77 | 18 | $2.20 | |
| $6.11 | 111 | 133 | 3 | 0 | $2.04 | |
| $17.12 | 196 | 265 | 18 | 1 | $0.95 | |
| X | $31.48 | 95 | 251 | 8 | 3 | $3.94 |
| All | $224.05 | 786 | 1,046 | 106 | 22 | $2.11 |
Summed on from the ads desk's daily records; the ads were shown 220,127 times in all. Arrived is the server's access log: every visit an ad's link brought, phones and misclicks included. Went on answered the consent door or went past it. To the end reached the last section of the demonstration; two of X's three were our own check of an ad's link on 7 October (below).
So the week bought 1,046 arrivals for $224.05, about $0.21 each. 885 of them reached the consent door, and 106 chose to go on, at $2.11 each. LinkedIn cost the most and brought 18 of the 22 who watched to the end. LinkedIn counted 44 reactions to the ads and no shares.
The site's own statistics, which count a visit only once the demonstration has loaded and the visitor has done something on it, told a visit from an ad apart from any other from on 5 October. To the end of the week they counted 44 visits from an ad of that kind, 22 of which reached its end. A page of its own follows a click from the ad to the end, step by step.
What did not work
Written down as they were found, each with its date. Times are UTC.
- The door stopped every paid visitor (5 October). Early that morning the server's access log held 52 visits from ads. 45 of them reached the consent door and not one answered it. 36 of the 52 were on a phone, where the door's two buttons sat about 1,460 pixels down a page about 1,750 tall. At 01:46 the buttons were moved onto the first screen, under a two-sentence summary, with the full notice unchanged below them. By 03:35 about 29 more visitors had come from ads, and none of them had answered it either. Whether the change helps is read on 8 October.
- Google's picture ads in Pakistan brought the wrong visits (5 October). About 45 visits in their first day, 40 of them on a phone, none past the door. The ads desk paused them.
- Half the first search keywords could never be shown (4 to 5 October). Of the thirty searches the first Google campaigns were built on, sixteen were too rare for Google to show at all, among them both that describe an agent running a website. In their first two days the search campaigns were not shown once.
- Google turned down one headline as clickbait (4 October): "It's a magic trick. On you." The ads carrying it ran on their other headlines.
- LinkedIn spent past the day's budget (4 October). $18.00 against $15 in Pakistan on the first day: the platform runs up to about a fifth over a campaign's total before it stops. The extra came off a later day.
- The page put VPN users in the wrong city (found 5 October). A synthetic visitor on a VPN-style connection, with a clock set to Karachi, was told they were in Nuremberg, Germany, where the connection came out. Since 09:14 that row says where the connection comes out, not where the reader is.
- Our own checks counted as visitors (4 October). The statistics counted 83 visits that day, 80 of them returning visits on a desktop with no campaign tag, spread across the releases shipped that day: most likely the people running the site looking at their own work. The statistics keep nothing that could say whose they were, so this report gives no count of unpaid visitors.
- Nobody has posted the link anywhere (as of 5 October). The posts are written and waiting for a person, because the agent may not post on other sites itself. Search engines had sent no one to the new pages yet.
Found later in the week, added when it closed (11 October):
- LinkedIn was switched off for the wrong reason (7 October). The ads desk paused every LinkedIn ad because hardly anyone was getting through the door. A person then pointed out that the ads' own text carried no link and no ask: on the post's page there was nothing to click. The link and the ask went in that morning, the ads went back on, and LinkedIn became the channel that brought most of the week's visitors who watched to the end.
- Our own check counted as two of X's full views (7 October, found 8 October). Two of the three visits from X that reached the end came from one machine that keeps opening the site, through the ad's own link. Read that day, X looked several times cheaper per full view than it was. Opening an ad's link counts as a visit from the ad, so an ad is checked by opening its address without the tag.
- The door fix of 5 October could not be judged (read 9 October). Too few visits on either side of the change, so no verdict. Over the whole week 885 visitors from ads reached the door and 106 went on. On 10 October the door's first screen was rewritten again, by a person this time; its effect is not this week's.
- Google's search ads stopped being shown (6 to 10 October). After their first two days they were shown ten times in five days and spent nothing; two campaigns were marked as limited by Google's ad policies, and the desk could not see why through the tools it has. Google gets no money from 11 October until there is a new idea for it.
- Reddit hardly showed the North American ads (5 to 10 October). They sat in Reddit's review for most of the week and were shown 123 times without a click, so nearly every Reddit figure above is Pakistan's.
- About twenty guides, and nobody from search (6 to 10 October). The agent wrote a public page or two a run, each answering a question about what websites can see. In the seven days to 10 October search engines sent none of them a visitor, and Google had read only a few, once each. The people who did read them came from this site's own links. From 10 October a new page needs a named way of reaching its first readers before it is built.
- It was refused 562 commands and called every run a success (counted 10 October). Over 114 unattended runs since September its permissions refused 562 commands; each of the 102 runs with a refusal still ended as a success. What was refused, and what fixed it.
- Nobody shared it. 22 visitors from ads watched to the end this week and none passed it on that we could see. Since 10 October the end of the demonstration hands the reader a verdict about them to send.
Runs this week that were undone, or did not ship:
-
Fixed the first-screen sentence that said nobody asked readers anything, and made a picture of that screen for the Reddit ads.
-
Stopped telling VPN readers they are in the city their connection comes out of, and made the ad-blocker card agree with itself.
-
The site now records which of its pages a visit to the demonstration came in by, so an ad can be judged by the page it led to.
-
Stopped telling readers on a VPN that it was raining where they sat, when the rain was in the city their connection came out of.
-
The site now records which of its ads a paid visit came from, so each ad can be judged by the readers it brought.
Why we publish this
Snitch is built by AlphaVenture, which builds AI systems for other companies. Clients ask whether an agent can be trusted with a live product. A week in which it built 25 pages, saw its own door turn every paid visitor away and had it rebuilt that night, and wrote down the rest, is part of the answer. So are the two weeks before, in which it was careful and nothing grew. The main report has its rules, its costs and every run since September, and a page for builders sets out what lets it release with no approval step.