Google Photos: finding a photo by scrolling

Help people find a photo they remember but can't describe. The brief's questions assumed people search. Most never get that far.

I built a discovery engine that pulled 5,640 public posts from the Play Store, App Store, Reddit, Google Photos Help, YouTube, Hacker News and Stack Exchange. An AI model read every post: 740 were about finding a photo, and 287 were real "can't find it" failures. The top-scoring problem was search showing nothing or the wrong photos.

But the people who post are the people who search. And only 7% of failure posts mentioned any workaround: most people just got stuck.

How strong is this? I checked the engine against 100 AI-made reference labels, and about 3 in 4 matched. Good enough to find patterns, not to quote as exact counts.

I asked 25 people about the last time they looked for an old photo, not their opinions, and watched 2 of them search live.

0 of 15 Google Photos users searched first. 11 scrolled and 4 browsed albums. 7 of 15 took over 5 minutes, against 1 of 10 on other photo apps. What they remembered most was who was in the photo (16 of 25), ahead of where (14), what was happening (12) and roughly when (11).

The sharpest group: people with 5,000+ photos who scroll. 5 of 6 took over 5 minutes, against 2 of 9 for everyone else.

Photos already knows who is in your photos, where they were taken and what was happening. It uses that in search and on People pages, but the timeline, where my segment spends all its time, only shows dates.

The moment I designed for: finding a photo of specific people, more than a month old, in a big library.

A timeline that knows people. Flick fast and it shows whole moments instead of single photos. Tap a face and marks light up on the scroll bar wherever that person appears. Open a photo and you can jump to the same people in other years.

Where AI does the work: recognising the same person across years, and grouping photos into moments. In the prototype the face signals are simulated; the grouping runs live.

• Better date jumpingMonth markers and hiding screenshots are cheap to ship, but Google and Apple already have them, and testers still opened 8 wrong photos.
• Photos that notices you huntingIt could solve more, but it has to learn how people hunt first, which means data collection and a privacy review. Months away, not this year.
• Improving searchIt's the loudest complaint online, but 0 of 15 users I asked started there.

Everyday photos: wrong photos opened fell from 8 to 0, and median time to the right moment fell from 31.6 to 23.5 seconds.

Similar events, like two Diwalis: worse. 9 of 10 found it on the plain timeline, 6 of 10 on mine, and it took longer. Still unsolved.

The "same people, other times" strip: tapped 2 times in 100. It needs a redesign.

The catch: my testers weren't my segment. They were 10 friends, 7 had seen an earlier version, and the test library had 2,517 photos, not 5,000+. Next: time real hunts with 5 to 6 people who have 5,000+ photos, then a 1% in-app test.

• People hate a changed timeline. My engine found 39 posts about updates that moved browsing features. Fix: opt-in first, and slow scrolling stays exactly as it is.
• Moments pop up while people just browse. Fix: a higher speed trigger and an easy off switch.
• Old photos sit under the wrong date. 5 of 23 people wanted childhood photos, likely scanned years later. Fix: spot scanned prints and let people fix a date in one tap.
• Faces send people to the wrong place. Face groups were 35 of 161 Google failures in my engine. Fix: "Not this person", already in v2.