While scrolling through LinkedIn the other day, I spotted an executive wearing a striking dive watch. The image was clear enough to pique my curiosity, yet just grainy enough to frustrate my watch-collector instincts. Determined to uncover the model, I turned to three headline-grabbing AI products—Grok 3, Gemini 2.5 Pro, and ChatGPT o3—to see whether machines could outperform the human eye.

For my original analysis, I used the executive’s photo, but to stay on the right side of image-rights law, I’ve recreated the test with a shot of my own wrist and watch. The reference image appears below and serves as the baseline for every AI run that follows.

I framed the photo to be deliberately ambiguous, recreating the very challenge that stumped me with the executive’s wristshot. (For the curious: the watch hiding in plain sight is a MING 37.07.)

Prompt used in every AI run: “What watch is this gentleman wearing?”

Grok: Creative First Impressions, Room to Impress Further

1. Baseline run (Think and DeepSearch off) – Grok politely admitted it couldn’t name the model, but offered to pull “visually similar” references. A candid miss, yet a helpful next-step prompt.

2. Think-enabled run – After 11 seconds of deeper analysis, Grok broke the image into case geometry, dial texture, strap material, and color palette. It still stopped short of a precise ID, but narrowed the field to aesthetic peers—Nomos, Junghans, Seiko, and Citizen.

3. DeepSearch run – Five minutes of cloud-scale searching later, Grok surfaced two concrete candidates: Seiko SSA231K1 and Movado Face 3640106. Both were thoughtful picks—neither was correct—but Grok explained its rationale and ultimately favored Seiko.

Watching Grok reason through each layer felt like peering into a living engineering diagram. Even when its guesses missed the mark, the transparent step-by-step logic—and the elegant interface that exposes it—made the exercise an interactive horology primer rather than a mere Q&A exchange.

Gemini 2.5 Pro: Keen-Eyed Feature Spotter, Still Chasing the Mark

Gemini pronounced the watch an Orient Bambino within seconds, backing the call with a tidy breakdown of case profile, dial layout, strap grain, and crown-to-bezel ratio. The analysis was methodical; the conclusion was dead wrong.

Curious, I toggled Deep Research. Gemini crawled 359 sites before returning—again, and with unwavering conviction—to the Orient Bambino. Determination admirable; accuracy unchanged. Still, the granular rationale offered a fascinating glimpse into how rigorously an LLM can audit visual cues even when its final leap falls short.

ChatGPT o3: Immersive Analysis, Nearly on Target

ChatGPT homed in on the watch region, digitally zooming and sampling color tones to tease out subtle design clues. For 2 minutes 14 seconds, it annotated dial markings, bezel hue shifts, lug geometry, and even the faint crystal reflection. It felt less like a static Q&A and more like a live masterclass in computer-augmented sleuthing.

The verdict: Movado Museum Classic. Given the minimalist dial and polished case, the guess made aesthetic sense, even if it wasn’t the bull’s-eye. What stood out most was ChatGPT’s closing advice: practical tips for shooting a sharper follow-up photo (diffuse lighting, macro focus, alternate angles) to improve future recognition.

Wrong watch? Yes. Worth the two-minute show? Absolutely—and I’d rerun it just for the spectacle.

Google Image Search — The “Un-AI” Baseline

Before conceding defeat, I opened the Google app and ran a straightforward image search, manually cropping to the watch. Voilà!—In near real-time, Google matched the reference image and correctly flagged the MING 37.07, even offering purchase links. Old-school computer vision still rules when the integration is tight.

Integration First: Why AI Assistants Must Tap Proven, Specialized Tools

Modern LLMs excel at language and broad reasoning, but real-world accuracy often hinges on narrow tools that already solve discrete problems brilliantly—Google Lens for visual lookup, Shazam for audio matching, FlightAware for live flight data. The fastest route from good demo to indispensable product is seamless orchestration of these domain-specific engines.

Why This Matters Most to Google

No company owns more world-class vertical services than Google: Lens, Search, Maps, Photos, YouTube, Gmail, Calendar, Flights, and more. Yet in my watch-ID experiment, Gemini never called on Google Lens, leaving ChatGPT as the more memorable experience. Had I not manually tried Google Lens, I would have defaulted to ChatGPT for future queries, despite its wrong answer, simply because the overall flow felt more helpful and integrated.

Google’s moat isn’t just model size or benchmark scores; it’s an unmatched portfolio of mature, best-in-class services. Rapidly stitching those assets into Gemini—so a single prompt can summon Lens for images, Maps for places, or Flights for itineraries—would turn every interaction into a demo of Google’s ecosystem strength before rivals can bolt on comparable capabilities. The assistant race won’t be won by the LLM with the largest context window; it will be won by the smartest switchboard, routing each user request to the specialized tool that already nails the job.

The lesson for AI builders—especially Google—is clear: embrace your best-in-class specialist tools and integrate them instantly, or risk losing users to platforms that feel more cohesive, even when they’re occasionally wrong.