I wanted to see if I could get my browser to do my chores.
Not just search for things, but actually do them. Chrome recently introduced Gemini Auto Browse—a feature designed to let an AI complete multi-step tasks for you on the web.
So, I picked two tasks on opposite ends of the spectrum: buying groceries on Safeway.com and automating achievement points on the Google Developer Forums.
One worked surprisingly well. The other was a hilarious failure.
Task 1: The Safeway.com Grocery Run
Groceries are a perfect target for browser automation. It’s a repetitive, tedious search-and-add-to-cart loop that eats up 20 minutes of my weekend. It also gets my goat as an ADHD shopper when I get distracted from my shopping list and just open a new tab to do something else. I get just as distracted inside the grocery store as I do online.
I fired up Gemini Auto Browse and gave it my list.
But there’s a catch. If you give the AI a raw, messy copy-paste job, it chokes. To make this work, you need to sanitize and normalize your data.
Here is the difference between a list that causes the browser automation to spin its wheels, and the one that actually works.
The Messy, “Human” Copy-Paste (Chokes the AI)
- “Hey, can you grab a pack of that ground beef? The 80% lean one is fine, whatever size.”
- “Some of that shredded cheese. The Mexican style blend.”
- “Yogurt – Fage brand, 2%, get the big 32 oz tub.”
- “A loaf of Signature Select sourdough bread (sliced, 24oz).”
- “That tube of plain original polenta.”
- “A 3-pack of sweet Fuji apples.”
- “A fresh bunch of green cilantro.”
- “Unsweetened almond milk (carton).”
- “That oats and more cereal (Honey Bunches style).”
- “A bag of white rice.”
- “Flour tortillas, burrito size (8 count).”
- “Fiber supplement gummies.”
If you paste that conversational list directly into Auto Browse, Gemini has to spend cognitive cycles parsing pronouns, brand names, and parenthetical fluff. It will get stuck on a query like “Yogurt – Fage brand, 2%, get the big 32 oz tub” because the Safeway search bar expects catalog keywords, not shopping chat.
Plus, a lack of specificity like “cheese” will result in a completely non-deterministic quantity of cheese—it could add a single string cheese, or a 3lb wheel of Gouda to your cart! You have to be specific.
The Sanitized & Normalized List (Flawless Execution)
This is what clean data looks like. No verbs, no conversational fluff, just exact product characteristics:
- Ground beef 80% lean
- Shredded cheese Mexican style
- Fage yogurt 2% 32 oz
- Bread signature select sourdough sliced 24oz
- Polenta original plain
- Apples fuji 3ct
- Cilantro fresh 1 bunch
- Almond milk unsweetened
- Cereal oats and more
- Rice white bagged
- Tortillas burrito size 8ct
- Fiber gummies
Because this list is clean, Gemini can execute sequential queries in the Safeway search box, pick the closest product matches, click “Add to Cart”, and move to the next item without hesitation.
Sanitization and Normalization 101
Here’s the difference:
- Sanitization: Stripping away the fluff. The catalog search doesn’t care about “artisanal, hand-massaged, organic, locally sourced.” It cares about “butter.”
- Normalization: Standardizing quantities and names so the search engine matches them accurately.
When you feed Gemini a clean, normalized list, Auto Browse navigates Safeway like a pro, clicking search bars and adding items to the cart while you watch.
Task 2: Hunting Achievements on Developer Forums (The Disappointment)
Next, I wanted to see if I could automate my way to some shiny badges on the Google Developer Forums.
The goal: have Gemini browse threads, like helpful replies, and maybe post a few introductory comments to trigger those early forum achievements.
It didn’t work. At all.
As soon as Gemini landed on the forum interface, it refused to behave like a human. It wouldn’t click the “Like” buttons or write simulated replies.
At first, I was cranky. I wanted my badges.
But honestly? This is probably for the best. If it were that easy to automate human-like interactions, developer forums would be overrun by AI bots farming karma and trust scores. The fact that the guardrails held is a win for the future of trust on the web.
The Takeaway
We are entering the era of the agentic web.
AI can navigate custom web layouts, but it requires us to be better communicators. If you want a browser agent to do your shopping, you have to feed it clean data.
And if you want to rank up on developer forums? You’ll still have to do it the old-fashioned way—by actually being human.
