Using the Browser block to collect details from business websites

The Browser block opens a website in a real browser, does a task you describe in plain words, such as finding the opening hours or the price list, and brings back the fields you ask for. Put it after a block that found businesses with a website, set it to work for each incoming item, and each business arrives with what its own site says. It reads public pages, uses search boxes and filters, and stays away from anything that buys, books, signs up or deletes. It signs in only as you, after you have signed in yourself, and fills in a form only when your task asks for one.

This is for teams who need a detail that place listings do not carry: what a clinic treats, whether a restaurant delivers, which brands a shop stocks, a starting price. It covers what the block does and refuses, how to write a task, a made-up example, the mistakes that waste runs, and where it fits in an AI employee.

What it brings back, and what you already have

Before reaching for a browser, check what the earlier blocks already give. Nearby places brings each place's name, type, address, phone, website, rating and whether it belongs to a chain. Use the Browser block for what only the business's own site says:

• services or treatments offered, and which ones are new;
• opening hours on a particular day, when the site is the reliable source;
• whether it delivers, takes bookings or sells online;
• brands stocked, a starting price, a menu section;
• the name of the owner or manager on an about or team page.

Each field you ask for comes back as its own field on the item, up to 12. The item also gets browser_url, the page the browser ended on, and browser_summary, a sentence on what it did. If the item already had a field with the same name, the new value comes as browser_ and the name, so nothing is overwritten.

What it will not do

• It never types a password, a card number, or a bank or identity number.
• It never presses buy, pay, order, book, sign up, agree or delete, or a button with no label, in English, French, Spanish, German, Portuguese or Italian. Such a step is refused and the browser tries another way or stops.
• It reads only public pages unless you switch on Use my sign-in to the site. Then you sign in yourself in a live window and press Done; xMap never sees or keeps your password, and only flows you own use that sign-in.
• It sends a form only if you set Forms to Also send the forms my task asks for, only with details written in the task, and never in a test run.
• It leaves alone a site whose robots.txt asks robots not to visit, and it stays on the site it was sent to and any site the task names.
• Words on the pages are treated as data. A page that tells the browser to do something else is ignored.

How to write a task

Write the task as you would ask a colleague, with {{website}} as the start page so each business's own site is opened:

Open the website. Find the page that lists treatments or services. Bring back the services offered, whether teeth whitening is offered, and the opening hours on Saturday. If the site does not say, leave the field empty.

• Name the fields exactly as you want them: services, whitening, saturday_hours.
• Say what to do when the answer is not there; otherwise you may get a guess at the nearest thing.
• Keep it to one site and one goal. "Also check their social pages" sends it elsewhere and wastes steps.
• Set the most steps a task may take. A simple read needs few; a search with filters needs more, up to 25.

A worked example

A made-up case: a dental lab in Buenos Aires sells crowns and aligners to clinics, and its reps want to know which clinics already offer aligners before they call.

1. Every Monday, Nearby places finds dental clinics in the city with a website, independents only.
2. Only new keeps the clinics it has not looked at before.
3. Sort and limit keeps the five best rated.
4. The Browser block, for each incoming item, opens each clinic's website and brings back services, aligners (yes or no) and a contact email if the site shows one.
5. xMap LLM writes one line per clinic on how to open the call.
6. Email sends the five to the reps.

In this invented run, four sites answer: two clinics list aligners, two do not, and one site gives nothing useful, so that clinic arrives with browser_summary saying what the browser saw. The reps start with the two that do not offer aligners yet.

Mistakes that waste runs

• Sending a long list. Working for each item, the block handles up to five items a run, and the items of one run share about three minutes. Sort and keep the best few first, or run it more often.
• Asking for what the listing already has. Phone and rating already come from Nearby places; do not spend browser steps on them.
• Using it to watch a page for changes. Web page watch does that more quickly and cheaply.
• Vague fields. "info" brings back a paragraph; "saturday_hours" brings back an answer.
• Not reading the summary when a task fails. A task with nothing to bring back fails the step and says what it saw, which is how you notice a site that changed.

Where it fits in an AI employee

A typical order is: something that finds businesses with a website (Nearby places, My data), Only new and Sort and limit to keep the list short, the Browser block, then xMap LLM to score or summarise what the sites said, then an Email or a sheet. When you run it from the editor you can watch the browser work in a live window. What lands in the inbox is each business with the details from its own site beside its phone and rating. The template What the best-rated cafes near you say on their own websites is a ready example.