For every visitor Google sends a website, its crawler fetches roughly five pages. For some AI companies, Cloudflare's network data has put the same ratio in the hundreds or thousands of pages per visitor sent back, according to analyses of Cloudflare Radar figures through mid-2026. The exact numbers move every month. The shape does not: machines now read the web far more than they send people to it.
Every website now has two readers. The first is a machine that reads it to answer a question on someone's behalf; the second is the person who arrives afterwards, already informed and ready to act. A website that serves only the second reader is invisible to the first, and one that serves only the first gives away answers and closes nothing.
Reader one: the machine
The machines are not one thing, and treating them as one leads to bad decisions. Some crawlers collect pages to train models. Others fetch pages live, at the moment a person asks an assistant a question, so the assistant can answer and cite a source. The second kind can send visitors; the first rarely does.
Our own policy reflects that split. We allow the search and user-triggered fetchers from OpenAI, Anthropic and Perplexity, because those are the requests that can put us in an answer and send a visitor. We also allow the major model makers' training crawlers, because being part of what a model knows is part of being recommended by it. We block bulk scrapers that collect pages for open datasets and offer nothing back. Reasonable businesses decide differently, and that is fine; the mistake is blocking or allowing everything without knowing which is which.
What the machine reader needs is unglamorous. Content in the HTML the server sends, because several AI fetchers do not execute JavaScript reliably. Facts in text rather than images. Clear headings that say what a section answers. Structured data that names the organization and the page type. None of this is new; AI simply punishes its absence faster than search did.
Reader two: the person who already knows
The person who arrives after an AI answer behaves differently from a search visitor. On an online store we run, visitors from ChatGPT completed a key event in about 24% of sessions over four months, against about 15% for Google organic. On our own agency site, AI-assistant visitors were engaged in 61% of sessions, against 31% from Google.
They also arrive deeper. Only 16% of the store's ChatGPT visitors landed on the homepage; most landed on a specific product or delivery page. For a service business the pattern can flip, with assistants linking the homepage when they recommend the company by name. Either way, the landing page has to carry the whole decision: what it is, what it costs, whether it is available to this person, and what to do next.
Here is the concession. A website built this way can feel less impressive. Long scroll-driven storytelling, hero animations that delay the content, and navigation that withholds prices until a sales call are all optimized for a visitor with time to browse. That visitor is becoming rarer, and the design budget should follow.
Reader three, arriving: the agent
A growing share of visits are not people at all but software acting for them: AI agents that open pages, compare options and fill in forms. They often arrive with a standard browser identity and no referrer, so analytics records them as Direct traffic. On our own site, more than half of recent sessions were Direct, and fewer than one in five of those showed any engagement, a sign that many were automated.
Agents reward the same things as the other two readers: clear structure, standard form fields, honest labels and no unnecessary obstacles. They also create new risks. Forms need protection from automated abuse, and analytics need filtering, a subject we cover in measuring a website when clicks fall.
What does not change
The website is still the only place you fully control. An AI answer is a summary written by someone else, from sources you did not choose, that can change tomorrow. Your site is where the facts live, where the transaction happens and where the visitor decides whether to trust you. If anything, AI makes it more important that the source is correct, because errors now spread through answers you never see.
Where this stops applying
The shift is uneven. Most sites still get the large majority of their visits from search and direct navigation, and AI referrals are a small share almost everywhere. Some businesses, such as a local trade or a store whose customers already know its name, will feel very little of this for years. The point is not to rebuild for AI alone but to stop building only for a visitor type that is shrinking.
Questions to ask of your website
- Does your main content appear in the HTML before any JavaScript runs?
- Do you know which AI crawlers you allow, and why?
- Could someone landing on any of your top ten pages make a decision without visiting another page?
- How much of your design budget goes to visitors who browse, versus visitors who arrive ready to act?
This is how our website development team plans new builds, and how our AI automation work connects a site to the systems that act on what arrives.