Browser automation is a fundamental web scraping technique for building your own dataset.

It is essential for investigating personalization, working with rendered elements, and waiting for scripts and code to execute on a web page.

However, browser automation can be resource intensive and slow compared to other data collection approaches.

👉Click here to jump to the Selenium tutorial.

Intro¶

If you’ve tried to buy concert tickets to a popular act lately, you’ve probably watched in horror as the blue “available” seats evaporate before your eyes the instant tickets are released. Part of that may be pure ✨star power✨, but more than likely, bots were programmed to buy tickets to be resold at a premium.

These bots are programmed to act like an eager fan: waiting in the queue, selecting a seat, and paying for the show. These tasks can all be executed using browser automation.

Browser automation is used to programmatically interact with web applications.

The most frequent use case for browser automation is to run tests on websites by simulating user behavior (mouse clicks, scrolling, and filling out forms). This is routine and invisible work that you wouldn’t remember, unlike seeing your dream of crowd surfing with your favorite musician disappear thanks to ticket-buying bots.

But browser automation has another use, one which may make your dreams come true: web scraping.

Browser automation isn’t always the best solution for building a dataset, but it is necessary when you need to:

1. Analyze rendered HTML: see what's on a website as a user would.
2. Simulate user behavior: experiment with personalization and experience a website as a user would.
3. Trigger event execution: retrieve responses to JavaScript or network requests following an action.

These reasons are often interrelated. We will walk through case studies (below) that highlight at least one of these strengths, as well as why browser automation was a necessary choice.

Some popular browser automation tools are Puppeteer, Playwright, and Selenium.

Headless Browsing¶

Browser automation can be executed in a "headless" state by some tools.

This doesn't mean that the browser is a ghost or anything like that, it just means that the user interface is not visible.

One benefit of headless browsing is that it is less resource intensive, however there is no visibility into what the browser is doing, making headless scrapers difficult to debug.

Luckily, some browser automation tools (such as Selenium) allow you to toggle headless browsing on and off. Other tools, such as Puppeteer only allow you to use headless browsing.

If you’re new to browser automation, we suggest not using headless browsing off the bat. Instead try Selenium (or Playwright), which is exactly what we’ll do in the tutorial below.

Case Studies¶

Case Study 1: Google Search¶

In the investigation “Google the Giant,” The Markup wanted to measure how much of a Google Search page is “Google.” Aside from the daunting task of classifying what is "Google," and what is "not Google," the team of two investigative journalists-- Adrianne Jeffries and Leon Yin (a co-author of this section) needed to measure real estate on a web page.

The team developed a targeted staining technique inspired by the life sciences, originally used to highlight the presence of chemicals, compounds, or cancers.

The reporters wrote over 68 web parsers to identify elements on trending Google Search results as "Google," or three other categories. Once an element was identified, they could find the coordinates of each element along with its corresponding bounding box. Using the categorization and bounding box, The Markup were able to measure how many pixels were allocated to Google properties, as well as where they were placed on a down the page for a mobile phone.

Browser automation tools' ability to collect and analyze rendered HTML pages can be essential. This is especially the case for search results, since most search results contain modules, carousels, and other non-standardized rows and columns that are more complex than lists.

Rendered HTML can be used to analyze the allocation of real estate on a website, which can be a useful metric to gauge self-preferencing and anti-competitive business practices relevant to antitrust. Take for example this case study, which was placed above the others because one of this section's co-authors happened to work on it.

Case Study 2: Deanonymizing Google's Ad Network¶

Google ad sellers offer space on websites like virtual billboards, and are compensated by Google after an ad is shown. However, unlike physical ad sellers, almost all of the ~1.3 million ad sellers on Google are anonymous. To limit transparency further, multiple websites and apps can be monetized by the same seller, and it’s not clear which websites are part of Google’s ad network in the first place.

As a result, advertisers and the public do not know who is making money from Google ads. Fortunately, watchdog groups, industry analysts, and reporters have developed methods to hold Google accountable for this oversight.

The methods boil down to triggering a JavaScript function that sends a request to Google to show an ad on a loaded web page. Importantly, the request reveals the seller ID used to monetize the website displaying the ad, and in doing so, links the seller ID to the website.

In 2022, reporters from ProPublica used Playwrite to automated this process to visit 7 million websites and deanonymize over 900,000 Google ad sellers. Their investigation found some websites were able to monetize advertisements, despite breaking Google’s policies.

ProPublica's investigation used browser automation tools to trigger event execution to successfully load ads. Often, this required waiting a page to fully render, scrolling down to potential ad space, and browsing multiple pages. The reporters used a combination of network requests, rendered HTML, and cross-referencing screenshots to confirm that each website monetized ads from Google’s ad network.

Browser automation can help you trawl for clues, especially when it comes to looking for specific network requests sent to a central player by many different websites.

Case Study 3: TikTok Personalization¶

An investigation conducted by the Wall Street Journal, "Inside TikTok's Algorithm" found that even when a user does not like, share, or follow any creators, TikTok still personalizes their "For You" page based on how long they watch the recommended videos.

In particular, the WSJ investigation found that users who watch content related to depression and skip other content are soon presented with mental health content and little else. Importantly, this effect happened even when the users did not explicitly like or share any videos, nor did they follow any creators.

You can watch the WSJ's video showing how they mimic user behavior to study the effects of personalization:

This investigation was possible only after simulating user behavior and triggering personalization from TikTok's "For You" recommendations.

Tutorial¶

In the hands-on tutorial we will attempt to study personalization on TikTok with a mock experiment.

We’re going to teach you the basics of browser automation in Selenium, but the techniques we'll discuss could be used to study any other website using any other automation tool.

We will try to replicate elements of the WSJ investigation and see if we can trigger a personalized "For You" page. Although the WSJ ran their investigation using an Android on a Raspberry Pi, we will try our luck with something you can run locally on a personal computer using browser automation.

In this tutorial we'll use Selenium to watch TikTok videos where the description mentions keywords of our choosing, while skipping all others. In doing so, you will learn practical skills such as:

• Setting up the automated browser in Python
• Hiding signs that are easy tells of an automated browser
• Finding particular elements on the screen, extracting their content, and interacting with them
• Scrolling
• Taking screenshots

Importantly, we’ll be watching videos with lighter topics than depression (the example chosen in the WSJ investigation.).

::: {.callout-tip}

Pro tip: Minimizing harms¶

When developing an audit or investigation, start with low-stakes themes: both to minimize your exposure to harmful content and to avoid boosting their popularity, unnecessarily. :::

Step 1: Setting up the browser¶

Our setup will consist of a real browser and an interface that will allow us to control that browser using Python. We chose Google Chrome because it's the most popular browser and easy enough (famous last words) to set up.

1.1 Installing Google Chrome¶

Please download the most recent version here.

If you already have Google Chrome installed, make sure it's a latest version by opening Chrome and pasting this address in the address bar: chrome://settings/help. Now verify that there are no pending updates.

1.2 Installing the webdriver¶

The webdriver is our interface between Python and the browser. It is specific to the browser (there are different webdrivers for Firefox [called Gecko], Safari, etc) and even to the particular version of the browser. It's easier to ensure we are working with the correct version by installing a webdriver that automatically detects the current version of Chrome.

Run the code in the cell below to download the Python package chromedriver-binary-auto. Adding an exclamation mark before code in Jupyter notebook allows you to run commands as if you were in your computer terminal's command line

Let's see if the installation worked correctly! Run the cell below to import the correct webdriver and open a new Chrome window.

The chrome-driver-auto package should have installed a driver that's suitable for your current Chrome version running the line of code above should have opened a new Chrome window.

This step is notoriously hard, and you might get a version mismatch error:

SessionNotCreatedException: Message: session not created: This version of ChromeDriver only supports Chrome version 112
Current browser version is 113 with binary path /Applications/Google Chrome.app/Contents/MacOS/Google Chrome

It means that you probably updated your Chrome in the meantime. To fix it, reinstall the Python package:

If everything works fine and you have the window open, our setup is complete and you can now close the Chrome window:

Step 2: Hiding typical tells of an automated browser¶

When you open Chrome with Selenium you'll notice that the window displays a warning about being an "automated session". Even though the warning is only displayed to you, the webdriver leaves behind other red flags that inform website administrators that you are using browser automation.

The website admins will use these red flags to refuse service to your browser.

Let's remove those.

This should open a new window without those warnings and navigate to tiktok.com:

Step 3: Finding elements on page and interacting with them¶

We will perform our mock experiment without logging in (but we will also learn how to create multiple accounts and how to log in later).

Instead of logging in, our first interaction will be dismissing this login window. Doing this programmatically has two steps:

1. We need to identify that [X] button in the page source
2. And then click it

Let's inspect the button element:

In my case, the particular element that the Developer Tools navigated to is just the graphic on the button, not the button itself, but you can still find the actual button by hovering your mouse over different elements in the source and seeing what elements on page are highlighted:

Our close button is a <div> element, whose data-e2e attribute is "modal-close-inner-button".

There are many ways to fish for the exact element you want, and many of those methods are built into Selenium. One way to find it would be using a CSS_SELECTOR, like so:

If Selenium successfully finds an element, you'll get a WebElement object of the first match. However, if Selenium does not find the element-- for example because the element hasn't loaded yet, you will get an empty object in return. This will crash your script if you try to interact with the empty element.

One thing you can do is to tell Selenium to wait up to X_seconds for that particular element before trying to click on it, like this:

We seem to have found something, let's click it! WebElements come equipped with special functions you can use to interact with them:

Did you notice a change on the page? Congratulations! You just automated the browser to click something.

Step 4: Scrolling¶

We now have a browser instance open and displaying the For You page. Let's scroll through the videos.

If you are a real person who (for whatever reason) visits TikTok on their computer, you could press the down key the keyboard to see new videos. We will do that programmatically instead:

When you run the cell above you will see that your browser scrolls down to the next video. You just automated scrolling!

Step 5: Finding TikTok videos on the page¶

Now that the site loaded and you can browse it, let's find all the TikTok videos that are displayed and extract the information (called metadata) from each of them.

1. Right click on the white space around a TikTok video and choose "Inspect".

1. Hover your mouse over the surrounding <div> elements and observe the highlighted elements on the page to see which ones correspond to each TikTok video.

1. You will see that each video is in a separate <div> container but each of these containers has the same data-e2e attribute with the value of recommend-list-item-container.
2. Similarly to how we found the close button, we can now use this to find all videos on page:

When we searched for the "dismiss" button we used the driver.find_element() function because we were only interested in the first element that matched our CSS selector.

Now we're trying to find all videos on page, so we use the driver.find_elements() function instead - it returns the complete list of elements that match the selector.

Step 6: Parsing TikTok metadata¶

Now that we found all the TikTok videos on the page, let's extract the description from each - this is how we will decide whether to watch the video, or to skip it. The process of extracting a specific field from a webpage is "parsing".

1. Pick any description, right click, "Inspect".
2. Let's locate the <div> that contains the whole description (including any hashtags) and make note of its data-e2s attribute.
3. Now let's write the code that, extracts the description from a single video (note that you can get the text content of any element by calling element.text)

::: {.callout-note} Note: We previously searched for elements using driver.find_element() and driver.find_elements(). That allowed us to search the whole page. Notice that here, instead of driver, we're using a particular element which we called video: this way we can search for elements within an element, rather than on the whole page. :::

Step 7: Finding the TikTok video that's currently playing¶

We know how to scroll to the next video, and we know how to find all videos that are loaded. At this point we could either:

1. Assume that at the beginning, the 0th video is playing, and then every time we press arrow down, the next video is being displayed
   
2. Assume that the arrow down does not always work and each time verify which video is actually playing

The problem with the first approach is that even if scrolling fails just once, our experiment will be compromised (after it happens we will be watching and skipping different videos that our script tells us). This is why we will go with the second approach and verify which video is actually playing. Back to our favorite tool- inspect element!

When you right click on the playing video, you will see that instead of our familiar UI we get a custom TikTok menu, so that won't work. Try right-clicking on the description of the video instead, then hovering over different elements in the inspector and expanding the one that highlights the video in the browser. Dig deep until you get to the div that only contains the video.

Still in the inspector try looking at the video below. You will see that the div that contains the video is missing and there is no element with the tag name video. That's how we can find if the video is currently playing - its div will contain the video element that we can find by TAG_NAME:

Step 8: Taking screenshots and saving page source¶

The presentation of your results might be more compelling, when its accompanied by screenshots, rather than just data. Selenium allows you to take screenshots of the whole screen, or just a particular element (though the latter is a bit cumbersome):

In the spirit of bringing receipts, you can also save the entire webpage to parse it later.

::: {.callout-tip}

Pro tip: Keep these records to sanity check your results¶

Taking a screenshot and saving the page source is a useful practice for checking your work. Use the two to cross-reference what was visible in the browser and whatever data you end up extracting during the parsing step. :::

Let's close the browser for now, and kick this workflow up a notch.

Step 9: Putting it all together¶

At this point, we can read the description of TikTok videos and navigate the "For You" page.

That's most of the setup we need to try our mock experiment:
let's watch all TikTok videos that mention food in the description and skip videos that do not mention food.

After one hundred videos, we will see whether we are served videos from FoodTok more frequently than other topics.

::: {.callout-tip}

Pro tip: Use functions!¶

So far we wrote code to open the browser, close the dialog, and find videos as separate cells in the notebook. We could copy that code over here to use it, but it will be much easier to understand and maintain the code if we write clean, well-documented functions with descriptive names. :::

Ok, with that out of the way, let's set up our first data collection.

First, let's make a directory to save screenshots. We will save screenshots here whenever we find a video related to food.

::: {.callout-tip}

Pro tip: Be careful about keywords¶

For experiments that use keywords, the choices we make will directly shape our results. In the field, you can mitigate your own predisposition and biases by working with domain experts to curate keyword lists. :::

The figure above shows when during our 100-videos-long session we were recommended a video about food (from keywords). The x-axis is chronological, the 1st video displayed is on the left, and the most recent video is on the right. The y-axis is "yes" or "no," depending on if the video was related to food.

Results¶

You can look back to the data/screenshots folder we created to check whether the videos we watched appear to be food-related.

If the feed was indeed increasingly filled with food videos, we would see more lines towards the right of the graph. At least here it does not appear to be the case.

Does it mean that the WSJ investigation was wrong, or that TikTok stopped personalizing content?

The answer is "No," for several reasons:

1. We only scrolled through 100 videos, this is likely too few to observe any effects. Try re-running with a higher number!
   
2. When studying personalization you should use an account per profile and make sure you're logged in, rather than relying on a fresh browser. So, instead of closing the login dialog, try actually logging in! You know how to find and click buttons, and this is how you put text in text fields.
   
3. When you're not logged in, you will be presented with content from all over the world, in all languages. If you filtered keywords in just one language, you will miss plenty of target content in other languages.
   
4. You should always have a baseline to compare to. In this case, you should probably run two accounts at the same time - one that watches food videos and one that doesn't. Then you compare the prevalence of food videos between these two.
   
5. The WSJ investigation was run on the mobile app rather than on a desktop browser. Perhaps TikTok's personalization works differently based on device or operating system.

Advanced Usage¶

Above we highlighted some ideas to make your investigation or study more robust, some are methodological choices, but others are technical.

There are some advanced use-cases and tasks you can perform with browser automation that include

• Authentication using the browser and storing cookies for later use.
  
• Intercept background API calls and combine browser automation with API calls. See selenium-wire as an example.
  
• Signing in with one or more email addresses.
  

We may cover some or all of these topics in subsequent tutorials, but you should feel free to experiment.

Let us know what you're interested in learning more about!

Related Readings¶

More tutorials on the same subject:

• "Using real browsers" - Sam Lavigne

Notable investigations, audits, and tools using browser automation:

• "Blacklight" - a investigative tool by Surya Mattu
  
• "TheirTube" - an art and advocacy project by Tomo Kihara
  
• "Worlds Apart" - a TikTok investigation by Henrik Bøe and Christian Nicolai Bjørke
  
• "WebSearcher" - A Python package by Ronald E. Robertson
  
• "Googling for Abortion" - Yelena Mejova, Tatiana Gracyk, and Ronald E. Robertson
  
• "webXray" - A website forensics tool by Tim Liebert
  
• "OpenWPM" - A privacy-measurement tool

Please reach out with more examples to add.

Citation¶

To cite this chapter, please use the following BibTex entry:

@incollection{inspect2023browser,
  author    = {Sapiezynski, Piotr and Yin, Leon},
  title     = {Browser Automation},
  booktitle = {Inspect Element: A practitioner's guide to auditing algorithms and hypothesis-driven investigations},
  year      = {2023},
  editor    = {Yin, Leon and Sapiezynski, Piotr and Raji, Inioluwa Deborah},
  note      = {\url{https://inspectelement.org}}
}

Acknowledgements¶

Thank you to Ruth Talbot and John West for answering questions about their two respective investigations.