What AI Engines Actually Read on Your Website — And Why It Isn’t Your Services Page

Somebody asks ChatGPT how to set up VAT for a webshop selling into Germany. Before it answers, it goes and fetches a handful of real pages from real websites to read. One of those pages might be yours.

That fetch leaves a line in your server logs. We read 66 days of ours to find out what AI engines actually read on our website — not which pages we hoped they would, or which pages we spent the most time on. The answer changed how we think about what a business website is for.

In This Article

• What AI engines actually fetched
• Your services pages are nearly invisible
• What European shop owners are actually asking
• Why the specific page wins every time
• What to do with your own site
• What this data doesn’t tell you
• Frequently asked questions

What AI Engines Actually Fetched From Our Site

There is an important distinction buried in AI traffic, and it is the whole basis of this article. Most bots that visit your site are collecting material in bulk to train a model. Those visits tell you very little — a bot hoovering up the whole internet is not a signal that anyone cares about your business.

A much smaller group behaves differently. When a person asks ChatGPT, Claude or Perplexity a question and the assistant decides it needs to go and read something, it makes a live request right then. Those requests are identifiable in your logs, and they are the interesting ones, because behind each is a human waiting for an answer. We have previously broken down which AI bots crawl a WordPress site and what each one is actually for.

Over 66 days we counted 1,807 of those live fetches. Roughly a third — 624 — were the homepage, which is what you would expect from someone asking who a company is. The remaining 1,183 landed across 150 different pages, and that spread is where the useful information lives.

For the developers: how these requests were isolated

The counts come from raw Nginx access logs, including rotated and gzipped archives, filtered to four user agents that represent live, user-triggered fetches rather than bulk crawling: ChatGPT-User, Claude-User, PerplexityBot and Perplexity-User, and OAI-SearchBot.

Training crawlers such as GPTBot, ClaudeBot, Applebot and Amazonbot are deliberately excluded — their volume is far larger and reflects indexing appetite, not user intent. Requests were further limited to HTTP 200 responses for HTML pages, dropping static assets, robots.txt, sitemap files and REST API calls, so that each remaining line represents one page actually delivered to an assistant mid-answer.

User agents can be forged, so treat these as directional rather than forensic. The pattern held consistently across all nine weeks.

Your Services Pages Are Nearly Invisible

This is the finding that should change something about how you spend your time.

Page type	Live AI fetches	Share
Pages answering one specific problem	1,106	94%
Services pages	63	5%
Theme and product pages	14	1%

Our best-performing services page — managed WordPress hosting — was fetched 23 times in nine weeks. A single article about a firewall blocking a page builder was fetched 168 times. That is more than every services page on the site put together, by a factor of nearly three.

This is not because the services pages are badly written. It is because nobody asks an AI assistant a question that a services page answers. People do not type “who offers WordPress development.” They type “why is my checkout rejecting Belgian customers” — and the page that answers that is the one that gets pulled into the conversation.

What European Shop Owners Are Actually Asking

Group the fetched pages by subject and a clear shape appears. The single largest theme is European commerce regulation — the unglamorous, high-stakes questions a shop owner cannot afford to get wrong.

Page	Live AI fetches
WooCommerce and the European Accessibility Act	81
WooCommerce shipping zones for the EU	58
WooCommerce VAT setup for the EU	52
Shopify to WooCommerce migration (Belgium)	40
WooCommerce payment methods in Europe	38
Shopify to WooCommerce migration (Netherlands)	30
What a WooCommerce store costs	20
Shopify to WooCommerce migration (Germany)	15
Why Belgian customers can’t pay	12

Taken together, the European compliance and commerce pages — on the European Accessibility Act, EU VAT setup, shipping zones and European payment methods — drew 254 live fetches. The cluster about leaving Shopify drew 94, split across separate pages for Belgium, the Netherlands, Germany and a general guide. Pricing questions drew 36.

Two things stand out. The first is that country matters. We did not write one migration page — we wrote several, one per market, because the tax treatment, the payment habits and the consumer law genuinely differ. Each of those country pages gets pulled independently. A single generic “how to migrate from Shopify” page would have collected a fraction of that, because it would not have been the best answer to any of the specific questions being asked.

The second is that the regulation questions are early-stage buying signals. Nobody researches the European Accessibility Act for entertainment. They read about it because a deadline is approaching and they suspect their store does not comply. That is a person with a budget and a problem, arriving at your site through a machine that just recommended you.

Why the Specific Page Wins Every Time

An AI assistant answering a question is doing something closer to citing a source than browsing a catalog. It needs a page that directly, unambiguously addresses the thing being asked. A page that covers a topic broadly is worse for that purpose than a page that covers one narrow slice completely — even though the broad page might look more impressive to a human visitor.

Our most-fetched page is a good illustration. It documents one problem — a web application firewall blocking a page builder from saving — and how to fix it. It is not a strategic overview. It is not a sales page. It is one problem, named precisely, solved completely. That is exactly the shape of thing an assistant reaches for.

The practical translation for a business is uncomfortable but simple. The page you would most like people to read is probably not the page that earns you the visit.

What To Do With Your Own Site

• Write down the last twenty questions a customer actually asked you. Not topics — questions, in their words. That list is your content plan, and it will look nothing like your services menu.
• Give each question its own page. One question, answered completely, beats a long page covering six things partially. If two questions differ by country or by rule, they are two pages.
• Answer in the first paragraph. Assistants lift the part that directly answers the question. Burying it under three paragraphs of preamble costs you the citation.
• Cover the boring, high-stakes topics. Tax, compliance deadlines, shipping rules and payment failures are unglamorous and they are exactly what people ask machines about, because the answer matters and getting it wrong is expensive.
• Keep your services pages sharp anyway. They are where the visit converts once an AI answer has pointed someone at you. They just are not what earns the visit.

What This Data Doesn’t Tell You

Being straight about the limits, because plenty of AI traffic commentary is not.

This is one website, in one niche, with a readership that skews technical. Our single most-fetched page is a developer troubleshooting article, which tells you as much about our audience as about AI behavior. Your mix will differ.

A fetch is also not a citation. We can prove an assistant collected the page while answering someone. We cannot prove it used the page, credited it, or sent anyone onward. Those are separate questions and the logs are silent on all three.

And the volume is not climbing in a straight line. July produced 1,040 live fetches; August produced 665. That may be European summer, or a change in how one of the assistants decides to browse. Either way, anyone selling you a story about relentless exponential growth in AI referrals is going beyond what this kind of data supports.

Frequently Asked Questions

• How do I know if AI engines are reading my website? Your server access logs record every request, including those from AI assistants. Live, user-triggered fetches appear under user agents such as ChatGPT-User, Claude-User and PerplexityBot, and each one represents an assistant collecting your page while answering somebody’s question.
• Which pages do AI assistants actually fetch? In our logs, pages answering one specific problem took 94% of live AI fetches outside the homepage, while services pages took 5%. Assistants reach for the page that directly answers the question asked, not the page describing what a company sells.
• Why do my services pages get so little AI traffic? Because people do not ask assistants questions that a services page answers. They ask about a specific problem, such as a failing payment method or a compliance deadline, so the page solving that problem is the one that gets pulled into the answer.
• What are European shop owners asking AI about? In our data the largest cluster was European commerce regulation — the Accessibility Act, VAT setup, shipping zones and payment methods drew 254 live fetches. A second cluster, about migrating away from Shopify, drew 94 and split cleanly by country.
• Should I write one broad page or several specific ones? Several specific ones. Our country-specific migration pages for Belgium, the Netherlands and Germany were each fetched independently, because each was the best answer to a different question. One generic page would have matched none of them as precisely.
• Does an AI fetch mean my page was cited? No. A fetch proves an assistant collected your page while answering someone. It does not prove the answer used your page, credited it, or sent a visitor to you. Server logs cannot show any of those things.