DEV Community

Jerrod Kim
Jerrod Kim

Posted on

Scraping Social Media with Gemini for sentiment analysis

Scraping Social Media with Gemini for Sentiment Analysis

Hackathon project: using Gemini + a user's browser to analyze real sentiment hidden inside social media comment sections.

Inspiration

If you want to understand how people actually feel about a product, brand, or topic, the best data source is often comment sections.

The problem: scraping social platforms is getting harder every year.

Platforms now deploy:

  • expensive APIs and paywalls (e.g. Reddit's API changes)
  • aggressive bot detection
  • rate limits and scraping prevention

Running scrapers from servers or headless browsers usually gets blocked almost immediately.

So we asked:

What if the scraper wasn't a bot at all — but a real user's browser?

Modern browser automation combined with Gemini's computer-use capabilities makes that possible.

A real browser session comes with some powerful advantages:

  • ✅ already authenticated to social platforms
  • ✅ trusted by anti-bot systems
  • ✅ capable of normal browsing behavior

In other words: the ultimate scraping environment already exists — the user's browser.


What It Does

Textpot leverages a user's browser to explore social media and collect comments for sentiment analysis.

The architecture separates browser control from AI decision making.

The system runs a loop (up to 3 turns) to navigate a page and analyze comments:

  1. Extension captures screenshot
  2. Screenshot POSTed to Cloud Run
  3. Gemini analyzes the screen and returns the next action
  4. Extension executes the action via CDP

The result is a feedback loop where:

  • the browser acts
  • Gemini decides what to do next

This allows Textpot to automatically explore comment sections and extract sentiment insights.


Architecture

The system is split into two parts.

1. Chrome Extension (User Machine)

The extension owns the browser.

Responsibilities:

  • opens the extension popup
  • attaches to the page via Chrome DevTools Protocol (CDP)
  • performs actions (click, scroll, keypress)
  • captures screenshots of the page

Everything runs directly inside the user's local Chrome session.


2. Cloud Run Backend

The backend owns the AI logic.

Responsibilities:

  • receives screenshots from the extension
  • sends them to Gemini
  • stores conversation history across turns
  • returns the next action to perform

Importantly:

The backend never directly touches the browser.

It only tells the extension what action to perform next.


Why This Architecture Works

Splitting responsibilities between browser and AI backend solves a major scraping problem.

The browser:

  • has real authentication
  • has real cookies
  • behaves like a normal user

Cloud Run simply tells it:

"Click here."

"Scroll down."

"Open this comment thread."

This approach bypasses many of the traditional scraping roadblocks.


Challenges We Ran Into

The first version of Textpot looked very different.

Initially we built it as a web app running a headless browser on Cloud Run.

That approach quickly failed.

Problems included:

  • bot detection blocking the browser
  • authentication failures
  • restricted access to social media pages

The fix was simple but important:

Move the browser to the user.

Once we pivoted to a Chrome extension, the system could use:

  • real user sessions
  • real cookies
  • normal browsing behavior

That solved most of the blocking issues immediately.


What's Next for Textpot

Next steps include:

  • polishing the extension UX
  • improving comment extraction
  • adding deeper sentiment analysis
  • launching on the Chrome Web Store

Because part of the system runs on Google Cloud Run, we'll also need to figure out a sustainable pricing model.


Final Thoughts

AI-powered browser automation opens up a new way to interact with the web.

Instead of fighting platform restrictions with bigger scrapers, we can:

  • use real browsers
  • keep AI in the backend
  • let models like Gemini decide how to navigate

For sentiment analysis and market research, this could unlock data sources that are otherwise extremely difficult to access.


If you're experimenting with Gemini computer use, browser automation, or AI agents, I'd love to hear how you're approaching it.

Top comments (0)