AI and the Future of Websites: Fewer Clicks, More Agents, and What to Do Now

For twenty-five years a website’s job has been simple: show up in search, get the click, persuade the visitor. AI is changing the first two steps. More questions are answered before anyone clicks, and more of the “visitors” are software acting for a person.

This isn’t a prediction about some distant future. It’s already visible in the data, and in the WordPress install behind this site.

Fewer clicks from search

In March 2025, Pew Research Center tracked the real browsing of 900 adults in the US across 68,879 Google searches. About 18% of those searches showed an AI summary at the top.

People were also more likely to stop browsing altogether: they ended their session on 26% of pages with an AI summary, against 16% without. Google disagrees with the study, and real traffic varies widely by topic. But the direction is hard to argue with: for simple factual questions, the answer increasingly arrives without a visit.

The searches that still lead to clicks are the ones a summary can’t settle: “who should build this”, “what will it cost me”, “can I trust this person”. That suits a consultant’s website better than a general-information site.

AI is becoming a visitor

The second change is quieter. AI systems are starting to use websites, not just read them.

WordPress 6.9 added the Abilities API to core: a standard way for a site to describe what it can do (create a post, look up an order, check availability) with typed inputs, outputs and permissions. WordPress’s MCP Adapter exposes those abilities through the Model Context Protocol, so an AI assistant can call them directly, with the same permissions as the user it’s acting for.

I found this on my own install without looking for it. When I counted the REST API routes behind this site, two namespaces stood out:

I never set out to make my site agent-ready. Updates did it for me. That’s how these shifts usually arrive: through the platforms, before most site owners notice.

How sites are responding

To see how established sites are adapting, I checked 50 major Indian and global websites for two signals.

Blocking AI crawlers. 27 of the 43 sites I could read block at least one AI crawler in robots.txt. Global news publishers block almost all of them. Developer platforms block almost none. I’ve covered the details in a separate article.

Publishing llms.txt. This is a proposed standard: a Markdown file at /llms.txt that gives AI tools a short, clean guide to a site’s important content. 11 of the 50 sites have one:

It’s a proposal, not an official standard, and it isn’t clear how much the big AI systems rely on it. But it’s cheap to add, and the early adopters are exactly the sites whose audiences already ask AI for help.

What I expect over the next few years

These are opinions, not forecasts, based on the data above and what I see in client work.

1. Less traffic, better traffic. Quick factual questions will be answered by AI. The visits that remain will come from people further along in their decision. Conversion rate will matter more than visitor count.
2. Original data and real experience will be what stands out. An AI can summarise what’s already widely published. It can’t produce your benchmarks, your client outcomes or your opinion. That’s why the articles on this blog are built on something measured.
3. Websites will get an “agent layer”. Booking, quoting, checking stock and placing orders will increasingly happen through structured interfaces such as APIs, MCP and the Abilities API, as well as through pages. Sites that only work through a visual interface will be harder for agents to use.
4. The fundamentals still decide. Fast pages, clean HTML, structured data and accessible markup are what make a site easy for search engines, screen readers and AI agents alike. Good engineering serves all three.
5. Trust becomes the scarce thing. When anyone can generate plausible content, a real name, a real track record and honest advice carry more weight, not less.

What I’d do now

The websites that do well with AI won’t be the ones that fight it hardest or chase it hardest. They’ll be the ones that are easiest for both people and machines to understand, and that offer something worth finding.