ZDNET’s key takeaways

• Amazon research failed, while sentiment analysis worked well.
• File organization succeeded, despite an unsettling permission request.
• Google’s agent still cannot control Chrome or match rivals.

---

Over the past few months, we’ve done some pretty deep dives into what you can accomplish using the agentic capabilities of ChatGPT Work and Claude, which now bundles the former Cowork agentic mode into the main Claude experience. But we haven’t spent much time with Gemini Spark, Google’s entry into the field.

Also: I gave Claude Cowork 7 non-coding jobs, and it earned a spot in my toolbox

That changes today. Here’s the top line: Gemini Spark is beta, and it feels like that. Some things work, some things fail miserably. And not the things you might expect.

One other note before we get started. Gemini Spark is the agentic AI tool for Gemini. Muse Spark is the language model that Meta (Facebook) Muse uses.

Test 1: Amazon product research

I have found that using ChatGPT Work to do product searches on Amazon can be a huge time-saver. When I’m searching for a product with a bunch of specific requirements, I can feed that assignment into ChatGPT Work, and it does the research. I used this approach for my nine vertical mouse shootout.

Also: How to use ChatGPT Work – and my top 10 tips for getting started with agentic AI

ChatGPT found the Amazon products I needed and even opened each in a browser tab, which I find helpful for comparison and elimination. When I ran the search, ChatGPT identified 11 products, of which I chose nine to review.

To test Gemini Spark, I fed the same prompt into Spark. Note the small Spark tab on the screen. To use the agentic AI, click Spark before giving your assignment.

Spark immediately divides the Gemini window into three panes. I’m not thrilled with this interface, but it is what it is.

One thing that surprised me was how quickly Spark presented results. ChatGPT took 10 minutes and 21 seconds. Spark took 30 seconds at most.

Also: How to get free Google AI Pro for an entire year – and save $240

As soon as the output was returned, I immediately saw Spark’s first limitation compared to ChatGPT Work. Spark won’t open browser tabs.

Instead of finding 11 candidate products on Amazon, Spark found only four. Well, that’s not quite right. Spark didn’t find any candidate products on Amazon. Instead, it presented one product from OzBargain, an article in Popular Science, a product from Pacifiko, and another from Ugreen.

And just in case you thought it couldn’t get any worse, check out this sentence before the product links I just showed you: “Here are top wireless vertical mouse candidates available on Amazon with Prime delivery.”

So the AI clearly knew I wanted Amazon listings, but it gave me web search results, probably from Google’s index. Then the AI misrepresented the results as available from Amazon with Prime delivery. I’m fairly sure it didn’t do any actual research because the results came back so fast.

Also: 5 ways I actually use ChatGPT Work in my small business to save time and money

We end this test with a starting score of Pass: 0, Fail: 1. Oops.

Test 2: Sentiment analysis

Recently, when Claude updated from Opus 5 to Opus 5.5, I thought I saw a fairly measurable performance improvement in Claude Code. I was curious whether other developers felt the same, so I launched a task to analyze social media to see what people are saying.

I gave Gemini Spark this prompt.

Here, Spark did an okay-ish job. It identified several Opus 5.5 issues and concerns across the web and consolidated them into main themes. Many users confirmed performance improvements, but they also discussed some issues.

The AI sourced a lot of information from Reddit (which is fine for this task). It pulled posts from Substack. But it completely ignored other social feeds like X and Facebook, which were in my prompt. Perhaps Google doesn’t have access to them, but I would have liked to see a mention.

In any case, I did derive value from this agentic assignment, so I’m calling it a win. We end this test even, with a score of Pass: 1, Fail: 1. Better.

Test 3: Organize my files

Next up is what’s become my canonical computer control test: organizing my ever-growing pile of PDFs. When I ran the test previously, ChatGPT Work and Claude (then Cowork) succeeded, but ChatGPT Work failed to ask permission to work with my files at appropriate times.

Also: I let ChatGPT Work and Claude Cowork loose on my files – only one made me nervous

To set up this project, I copied my big folder of PDFs to a temporary directory. I did this so I wouldn’t have to worry about what damage the AI might do to my files. Then I told Gemini which folder to work on. Slightly confusingly, the folder selection is in the left-hand pane. Most other AIs let you select the folder using the plus icon. Not Spark.

Gemini Spark asked permission, but it was one really broad permission request that made me very nervous. Notice what it said (bold is my freaked-out emphasis): “By allowing, Gemini will be able to work on tasks with this folder, including sharing the content of this folder with third parties.”

This is Google asking permission to share my most private files before I’ve even given it a task. This should have been a far more granular request, and multiple access requests should have come after my prompt to work on those files. This issue, alone, will keep me from using Gemini Spark for local computer assignments.

Also: The best and worst AI for your privacy, ranked – and how each handles your data

But, for you, Dear Reader, I soldiered on. First, I asked Gemini what it could tell me about the folder. It divided the contents into categories and presented them, but while both Claude and ChatGPT found concerns (such as duplicate files, files with non-specific names, and more), Gemini reported nothing.

On the other hand, Gemini does something that no other AI does: it keeps a backup of the files it works on. I’m not sure whether that’s how it handles sharing, whether it uses up Google Drive space, or what. But the idea that if the AI completely nukes your files, there’s a backup, is nice.

Also: How to keep your conversations with ChatGPT, Gemini, Copilot, or Claude private

That said, you should have (and I do have) your own backup infrastructure. Don’t rely on an agent to cover your back from errors.

I then told the AI to rename the generic files.

The AI accomplished this task quite successfully.

After it presented its analysis, I told the AI to rename the files, which it did.

Next, I had the AI scan all the files to create a more comprehensive set of categories.

Finally, I told it to go ahead.

The process worked. You can see how it cleaned up the folder with nice categories and organization.

I would call this assignment a success, but because of the sharing warning and the lack of granular permissions, it also made me hesitant to use it. We end this test with a score of Pass: 2, Fail: 1, Overshare: 1. Ugh.

Test 4: Chrome browser use

One hallmark of agentic AI is that agents can use browsers (usually Chrome) to access web sites and get work done.

While some sites provide programmatic interfaces, AIs use browsers to get the work done, even when not set up for algorithmic data interchange. After all, if you want your AI to be your personal assistant, it has to do everything you need online.

Also: How to switch from ChatGPT to Gemini – without starting from scratc

That leads us to Google Gemini Spark and Google Chrome. Sure, Claude and ChatGPT have very nice browser interfaces, but Chrome is Google’s product. You’d expect Chrome to be incredibly capable, efficient, and powerful when Spark uses it.

Enter surprising plot twist, stage left. Oh, Google. You seemed so promising. The fruit was so low-hanging. The runway, it was clear. The skies, they were blue. All you had to do was land the plane. And yet.

“I cannot directly access or interact with your open Google Chrome window.”

Wait. What?!?

So, yeah. We end this with a score of Pass: 2, Fail: 2, Overshare: 1, Missed Opportunity: 1.

How much Spark do you get?

One concern across all agentic implementations is just how much usage you get before the AI companies cut you off for a time. In this regard, Spark did fairly well. I ran four fairly substantial tests on my $20-per-month Google AI Pro plan, using just 20% of the five-hour allocation.

Don’t underestimate Google

While I recommended Gemini for entry-level buyers because of the goodies that come with the $20-per-month fee, I don’t recommend using Gemini Spark over Claude or ChatGPT Work.

Also: ChatGPT Plus vs. Google AI Pro: Same $20 price, very different perks

But that’s not the whole story. Not by a long shot.

When Google first entered the generative AI business with Bard (later known as Gemini), its approach was very much subpar. Yet Gemini is now quite good. Nano Banana, Google’s image generation tool, is exceptional.

Google has a long track record of fielding subpar products and then cancelling them or wildly improving them. Android is a good example of something that started rough and became a world-dominating technology.

I fully expect Google to fix these issues. I have no doubt that Spark will eventually have excellent Chrome control. I do worry that Google won’t fix its permissions issue, or, as everything is Google Drive-related, will ask for overly broad sharing options. But beyond those issues, don’t discount Google.

The tech giant is certainly playing catch up, but it’ll get there. The company is relentless. It’s constantly improving and adding capability. So, when we come back here in a year, expect Spark to have captured lightning in a bottle and given Google’s agentic AI ambitions one heck of a power surge.

---

You can follow my day-to-day project updates on social media. Be sure to subscribe to my weekly update newsletter, and follow me on Twitter/X at @DavidGewirtz, on Facebook at Facebook.com/DavidGewirtz, on Instagram at Instagram.com/DavidGewirtz, on Bluesky at @DavidGewirtz.com, and on YouTube at YouTube.com/DavidGewirtzTV.