Logging into the AdMob console after midnight just to check whether yesterday's eCPM dipped — if you run apps as an indie developer, you probably know how much time these "just looking" tasks quietly eat. I have been shipping mobile apps on my own since 2014, and even now that my catalog has crossed 50 million cumulative downloads, I still review mediation lists and adjust bidding source priorities by hand.
Browser-agent tools have finally moved into a usable range for me in 2026. I spent a week running Gemini Computer Use and Claude in Chrome in parallel, pointed at the same AdMob management tasks. This is the journal. It is less about which one "wins" and more about where the differences actually showed up.
The three tasks I gave both agents
To keep the comparison honest, I gave both agents the same workload.
The first task was setting up ad units for a new app: banner, interstitial, and rewarded, all following my naming convention. The second was exporting the priority order of bidding sources inside a mediation group (Meta Audience Network, AppLovin, Unity Ads, and so on) to a CSV. The third was reviewing the past seven days of eCPM per ad unit and flagging any unit that dropped by 30% or more compared to the prior week.
All three need a mix of clicks and drag operations and cannot be done purely through APIs. AdMob exposes a Reporting API, but mediation configuration is largely outside its surface, so browser-side work is unavoidable. That is where an agent earns its keep.
How Gemini Computer Use behaved
Gemini Computer Use, on Gemini 3.x Pro models through Vertex AI, declares a computer_use tool. The model sees screenshots and proposes coordinate-based actions, which the runtime (a Playwright or Chrome DevTools Protocol wrapper) executes.
What stayed with me was how stable its screen interpretation felt. Sidebar transitions like "Apps" → "Ad units" stayed on track even when label positions shifted. On the ad unit creation form, it filled in "Format," "Ad unit name," and "Frequency capping" correctly whether I prompted in Japanese or English, which was a pleasant surprise.
The mediation priority export, on the other hand, hit one failure in twenty rows. The mediation table uses virtual scrolling, so rows that scroll out of view drop from the DOM. Computer Use makes decisions from what it visually sees, so this is a structural limitation. I worked around it by telling it "only transcribe rows currently visible," which then ran cleanly.
For the eCPM anomaly task, I used the report screen in "Compare" mode (past 7 days vs prior week), had it screenshot, and read the diff visually. Reading numbers off of charts was accurate enough to trust: one week produced two false positives and zero misses.
How Claude in Chrome behaved
Claude in Chrome runs as a Chrome extension. Because it uses extension permissions, it has direct DOM access — a structural difference from the coordinate-based Computer Use approach. DOM access means table elements can be read directly as elements, which neatly sidesteps the virtual scrolling problem above.
On the mediation priority task, Claude in Chrome walked the <tr> rows and returned the full list as CSV in roughly half the time Computer Use took, with no scrolling glitches. When the DOM is available, this approach is just plainly faster.
The trade-off appeared on some AdMob screens where React was still hydrating. Claude in Chrome occasionally clicked before event handlers were attached, and a few clicks went into the void. This is a web-app behavior issue rather than the agent's fault, but in practice I had to retry once or twice. Computer Use, because it advances visually, tends to wait until the page looks loaded — that paid off here in stability.
For eCPM anomaly detection, Claude in Chrome read the numbers straight from the DOM rather than going through a chart image, so the values were exact. False positives over the week: zero.
How I ended up dividing the work
After a week, I settled on "Claude in Chrome where the DOM is reachable, Gemini Computer Use everywhere else." Looking at AdMob alone, Claude in Chrome edges ahead on speed and accuracy. But once you start crossing AdMob, Firebase Console, and App Store Connect in a single workflow, Computer Use's general-purpose nature starts to matter.
This reminded me of my grandfather's toolbox — both of my grandfathers were temple carpenters (miyadaiku). A plane and a saw both shape wood, but the situations they fit are different, and that is not a question of which is "better." Agents are the same: pick by the task and the web app in front of you.
When I started writing HTML on my own back in 1997 as a sixteen-year-old, browsers were things humans operated. Almost thirty years later, AI moves around inside them on my behalf, but deciding what to delegate and what to watch myself is still a human judgment. Agents are not magic wands; used carefully, they hand time back to you, reliably. That is what one week with both reminded me.
My next experiment will be the AppLovin dashboard and the SKAdNetwork report screen, with the same comparison. I will post a follow-up once it is settled. Thank you for reading.