Web Navigation

When you get something done on the internet, you do it through a chain of small actions. You enter an address, read the page, click a link, type something into a field, scroll down. Web navigation means that a computer program carries out exactly these actions itself. It receives a goal in plain language, for example: Find the cheapest train from Hamburg to Munich on Friday. It then works its way step by step through real websites until the goal is reached or it gives up. So the difference from an ordinary chat program is not knowledge, but action.

From answer-giver to task-doer

Until now, AI programs mainly delivered text. They would explain to you how to book a train ticket, but you did the booking yourself. Web navigation shifts this boundary. The program no longer provides the instructions but the result. That is precisely why it is considered one of the most important developments in the industry in recent years.

Economically, this is interesting because an enormous amount of office work consists of clicks. Entering orders into a system, comparing prices from twenty providers, copying data from a portal into a spreadsheet. Such activities are mindless but take hours. A program that reliably takes them over saves companies real money. That is why major providers like OpenAI, Google, and Anthropic are investing heavily in this direction.

But there is also a downside. A program that clicks on its own can also buy the wrong thing, delete an account, or fall for a scam attempt. With a plain text error, you get annoyed. With a click error, money may actually be gone. That is why the central open question is not whether it works, but how reliably.

See, decide, click

The process is a loop that keeps repeating. First, the program perceives the page. This happens either via a screenshot, i.e., an image of the screen, or via the page’s source code, i.e., the blueprint it is made of. From this perception, a description of what is currently visible is generated.

In the second step, the language model decides on the next action. It chooses from a fixed set of actions: click this element, type this text, scroll, go back, wait. This action is passed on to the browser and executed. Then the loop starts over, because the page now looks different. A booking can easily require thirty or forty such rounds.

The biggest problem here is error accumulation. If each individual step is 95 percent correct, a chain of twenty steps is already well below 50 percent reliable. A human quickly notices they have gone astray and turns back. A program often does not notice this and stubbornly keeps going. Current systems therefore only solve everyday test tasks correctly in roughly half to two-thirds of cases.

Where such agents are already at work

The most visible are the features from major providers. OpenAI offers a mode in which the program completes tasks in its own browser window. Google and Anthropic have comparable tools that can operate a computer. In news reports, they usually appear under the keyword agents, meaning programs that carry out action steps on their own.

Less visible but more widespread is the use in companies. There, web navigation is increasingly replacing older automation software that followed rigidly fixed click paths. Its weakness was: as soon as a page changed its appearance, the process broke down. An AI-powered agent often finds the changed button anyway, because it understands the page’s content.

For you personally, this has a practical side. Many websites defend themselves against automated access, for example with puzzles meant to distinguish humans from machines. A tug-of-war is currently developing between site operators and agent providers over this. A common misconception, by the way, is confusing web navigation with a web search. Searching means finding information; navigating means actually doing something on the page you found.