41 AI Agents Read Our Site. Analytics Showed Zero.

41 AI Agents Read Our Site in 30 Days. Our Analytics Showed Zero.

The number, and where it came from

We publish a free readiness grade and we count AI readers for customers, so it would be strange not to run both on ourselves. Here is the whole 30-day reading for loudpixel.ai, by operator:

Sixty reads in total. Forty-one of them were AI agents — a reader acting on behalf of an answering system, which is the only category that represents demand from the channel this business is about. Thirteen were classic search crawlers: real traffic, different meaning. Four were SEO tools, which are never evidence of AI demand — somebody's rank tracker was looking at us. Two named themselves and have not been looked up yet, one of which was DeepSeek's crawler.

Ten more requests were recorded and deliberately not counted: bare HTTP libraries claiming no identity at all. Most of those are our own scripts, and counting them would flatter the number.

Why your analytics shows none of this

Every mainstream analytics product learns about a visit in one of three ways: a JavaScript snippet runs, a tracking image loads, or the page fires an event. An AI agent does none of the three. It requests the document, parses the text, and closes the connection.

So the visit happened. The answer that mentioned you — or did not — was written partly from what that agent read. And your dashboard, checked the next morning, is honest and empty.

This is not a gap you can close by configuring analytics differently. The reader never executes the code that would report it. It needs a count taken on the server, where the request actually arrives.

The finding that changed what we did next

The operator breakdown is interesting. What it does not tell you is what those agents were reading. That table is the uncomfortable one:

Forty-two of sixty reads went to the sitemap. A sitemap is a list of addresses. It contains no argument, no offer, no reason to name us. Seven reads in total touched anything that describes the business.

We had spent months making the pages good and had told the agents where the pages were without giving them a reason to open one. On the same date, a separate reading of our AI visibility showed us named by 3 of 14 assistants — and the two facts are one fact. The agents were arriving, taking the index, and going.

Three things changed that week as a direct result: the front page now leads with what only we measure rather than the category's boilerplate; the machine-readable description of the business says the same thing instead of a sentence every competitor publishes; and every page now advertises its own plain-text twin in a response header, because eighteen of those twins existed and nothing anywhere pointed at one. That last one is why exactly one markdown read appears in the table above.

The part where we were wrong

Here is the column we did not want to publish:

Zero confirmed. On a site whose own pitch includes knowing which AI assistants are really reading you.

The cause was not the checking — it was that 29 reads arrived with no visitor address attached, so there was nothing to check. And the reason that went unfixed for a month is worth more than the bug: "no address" was a single verdict covering five different situations, each with a different remedy. A proxy not forwarding the address, a setting too high, a setting too low, checking switched off deliberately. Nobody could tell which, so nobody could fix it.

Both halves are now fixed. The address is read in the one case where it was being discarded for no good reason, and every remaining miss records which of the five it was, so the next look at this table is actionable instead of a shrug.

We are publishing the zero because a measurement product that only shows you its good numbers is not a measurement product. The next reading will be better, and it will be the same counter that produced this one.

What this means if you run a site

Four things follow, in order of how cheaply you can act on them:

1. Assume the agents are already there. Ours were, at 41 reads a month, on a small site. If you sell anything, yours are too, and nothing you currently look at shows them.
2. Check what they are reading, not just that they came. If almost all of it is your sitemap, you have a discovery signal and no argument. That is a content problem wearing a technical costume.
3. Give the cheap document a way to be found. Agents pay tokens for your markup. If you publish a plain-text version of a page, say so in the response — ours existed for months and were pointed at from nowhere.
4. Do not trust a reader's name on its own. Anything can call itself GPTBot. Where an operator publishes a way to verify, the check should run; where it does not, the honest label is "a reader identifying itself as X".

Frequently asked questions

Why can't Google Analytics see AI agents?

Because it learns about a visit from JavaScript running in a browser, and an AI agent runs no JavaScript. The same applies to any analytics product that depends on a snippet, a tracking pixel image, or a client-side event. The request reaches your server; nothing reaches your dashboard.

Are these the same as search engine crawlers?

No, and mixing them up is the common mistake. In our 30 days, 13 reads were classic search crawlers feeding an index, and 41 were agents reading on behalf of an answering system. Both are real traffic. Only the second is demand from AI answers, and the report keeps them apart.

How many AI readers can be identified?

66 assistants and crawlers are recognised by name, across 28 operators. A reader we do not recognise is recorded and labelled as unrecognised rather than dropped or guessed at.

Can a reader fake being ChatGPT or Googlebot?

It can claim to be. Several operators — Google, Microsoft, Apple, OpenAI, Anthropic, Perplexity among them — publish a way to check a claim, and where they do, we run it. Where they publish nothing, the claim cannot be checked by anyone and we say so rather than implying confirmation.

How do I count AI agents on my own site?

It needs a server-side count, because the reader never runs the code a client-side tool depends on. Our free report tells you what an agent can and cannot do on your site today; counting the readers over time is what the pixel does, and we will tell you plainly whether it is worth it for your traffic — send us your site.

The honest summary. Our site was being read by AI assistants 41 times a month, our analytics showed zero, almost all of the reading was the index rather than the argument, and we could not confirm a single reader was who it claimed. Three of those four are fixed. Get your own readiness level free, and if you want to know what the grade means before you see yours, the five levels are explained here.

Written with AI assistance and reviewed by LoudPixel for accuracy.