Your website now has two readers

For twenty years the deal behind a company website was simple. Someone typed a question into Google, saw ten links and clicked one. The website caught the visit and tried to turn it into a call, an order or a booking.

That deal is coming apart at both ends. More questions are now answered before anyone clicks. And more of the browsing that remains is done by software, on a person’s behalf. The website is still needed. What it is needed for has changed.

Fewer clicks after the answer

Google says its AI Overviews now reach more than 2.5 billion people a month. AI Mode, the conversational version of its search, passed one billion monthly users within a year of launch. Sundar Pichai gave both numbers at Google I/O in May.

What that does to clicks has been measured from several sides. Pew Research Center followed the browsing of 900 American adults through March 2025. When a Google results page showed an AI summary, people clicked a traditional result on 8 percent of visits. Without a summary, they did so on 15 percent. The links inside the summary itself were clicked on 1 percent of visits.

Ahrefs repeated its own study this year, across 300'000 keywords. In December 2025 an AI Overview on the page went with a click rate for the top result about 58 percent lower. And SparkToro, working with Similarweb clickstream data, found that 68 percent of Google searches in the United States ended without any click in the first four months of 2026, up from about 60 percent in 2024. None of these studies is a proof on its own. All of them point the same way.

Browsers that act

The second change is in the browser. In October 2025 OpenAI launched ChatGPT Atlas, a browser with ChatGPT at its core. In January Google gave Gemini in Chrome a feature called auto browse, first for paying subscribers in the United States. It compares prices, fills in forms and puts products into a cart, and it is designed to pause and ask before it buys something. In August it reached paying subscribers on Android in the United States.

In July OpenAI said it would shut Atlas down and move its agent features into the ChatGPT desktop app, a browser in its cloud and an extension for Chrome. TechCrunch read the decision as OpenAI concluding that “the browser is a feature, not the destination.” At the end of September OpenAI showed where the feature went: Dots, agents that work in the background, each with its own cloud computer and browser, navigating the web for their user, who approves what they do.

Whichever product wins, the consequence is the same. Some of the visitors to its website will not be people. They will be agents sent by people, reading the page, comparing it with others and trying to finish a task. A page an agent cannot read, or a form it cannot complete, is a sale that goes to someone else.

Agents at the checkout

Buying is where agents are pushed hardest. In September 2025 OpenAI launched Instant Checkout in ChatGPT, built on the Agentic Commerce Protocol, an open standard developed by Stripe and OpenAI. Stripe issues a payment token limited to one merchant and one basket value, so the assistant can start a payment without exposing the buyer’s payment details.

Google is building the same layer. Its Universal Commerce Protocol, developed with Shopify, Etsy, Wayfair and Target among others, is meant to let agents act for a user, and Chrome will support it. At the end of August Google began letting people in the United States book hotels inside AI Mode and pay with Google Pay, with the hotel or the booking platform as merchant of record. Flights planned in AI Mode are still bought on the airline’s or a booking platform’s site.

The pattern so far is plain. The research moves into the assistant. Where the money changes hands is still being decided. Either way the assistant chooses which companies get the chance, and the merchant’s own site has to hold up when the agent or the person arrives.

What machines read

What do the machines want from a page? Google has answered for its own search, in a guide for site owners updated in July, and the answer is less exotic than the industry hoped. Its AI features build on its ordinary ranking systems. No special schema markup is needed, and no AI text files either. Google says its search does not use llms.txt or similar files, so they neither help nor harm a site there.

The advice that remains is old advice made sharper. Let the crawler in. Put the facts in the page as text. Write what Google calls non commodity content: useful, reliable, and made by you rather than repeated from others. Structured data stays, because it still earns rich results in ordinary search.

Many sites fail the first step. Adobe found this spring that around a third of product pages could not be properly accessed by AI. We still include an llms.txt in Cited’s fixes, for the assistants that read one. It is a short index and cheap to make. Nobody should sell it as a lever on its own.

What people still come for

When fewer people click, the ones who do arrive later in their decision. They have read a summary, compared a few names and asked their follow up questions somewhere else. Adobe’s data shows what that does to a visit: by March 2026, shoppers who came to American retail sites from an AI assistant converted 42 percent better than other visitors, where a year earlier they had converted worse.

What they want from the website is what a summary cannot give them. The price, or how a price is built. The work itself, shown properly. The people behind it. A way to reach a person. A next step that works on a phone late at night. This is also the part no machine writes for you: the taste in how the work is shown, the judgement about what to promise, a person who picks up the phone. That is where trust is still made.

So the company website now has two jobs. It is a source for machines, plain, consistent and easy to quote. And it is the place where a person decides, which is a design problem, a writing problem and a measurement problem at once.

Where our products sit

Our products split along the same two jobs. Cited works on the first. Every week it runs the questions buyers really ask across ChatGPT, Claude, Gemini, Perplexity, Copilot and Google AI Overviews, records whether a company is named and who is named instead, and proposes the fixes: structured data, clear facts, pages the models can quote and the sources they trust. A person approves every change.

Studio works on the second. It puts a CMS on top of a built site, with no plugins on the public site, publishing that is tested first and rolled back if anything fails, and every publish committed to the company’s own repository. The assistant drafts, and a person approves every change. Proof tests two versions of a page, a shop or an email on real visitors and keeps the one that sells. North puts the numbers in one place, Search Console included, and marks a missing figure as missing. Studio and Proof are in early access.

A website will be read more often than it is visited. It is worth writing for both readers.