Why can’t people find a photo they know they have?
An AI engine that read 25,065 public posts from Google Photos users. Ask it anything: it counts the matching posts and shows you the real ones.
Reads public complaints, sorts them, finds where photo search breaks
- Collectcode
- Sort every postAI agent
- Double-check2nd AI · 82% agree
- Read key storiesperson · 20 of 49 kept
- Find new problemsAI agent · 6 themes
- Find patternsAI agent · tested vs chance
- Answer questionsAI picks, code counts
Most searches break when the app tries to understand them
- What this shows
- Every failure placed on the four stages of a search named in the brief.
- Stage 1
Saying what you remember
48 posts break hereNever found: 31 of 34 that say how it ended
The person can’t put the memory into a search: no way to say a date range, two people together, or the words just aren’t there.
- Stage 2
The app understanding it
364 posts break hereNever found: 274 of 292 that say how it ended
The app misreads a sensible search or holds wrong details: wrong dates, wrong faces, unread text, no place.
- Stage 3
Judging the results
61 posts break hereNever found: 43 of 45 that say how it ended
Too many results to look through, or only a handful shown, so the right photo can’t be picked out.
- Stage 4
Recovering after a miss
111 posts break hereNever found: 79 of 89 that say how it ended
After the first try fails, the person switches approach or gives up.
Once the first try fails, recovery rarely works: of the 82 people who tried a second approach and said how it ended, 9 found the photo. A post can break at more than one stage. Public posts under-count stage 1, because people seldom write “I didn’t know what to type”; our survey measures that directly.
How this is calculated
- What is counted
- One post about finding a photo (505 posts). Each post was tagged by AI with what went wrong.
- Formula
- A stage counts a post if the post has at least one failure that belongs to it. Stage 2, the app understanding it, covers wrong or no results, Ask Photos misreading, a wrong date, wrong faces, no place, unread text and ignored captions: 364 posts. Stage 4 counts posts that tried two or more approaches or gave up. “Never found” = posts in the stage that ended not found ÷ posts in the stage that say how it ended, for stage 2 274 ÷ 292.
- Keep in mind
- A post can count in more than one stage, so the stages don’t add up to the total. Public posts under-count stage 1 because people rarely write “I didn’t know what to type”; the survey covers that.
People remember plenty. Google Photos can’t use most of it.
- What this shows
- 101 of 162 things people remembered ran into a feature that failed, or no feature at all.
162 things remembered, from the posts where people say what they remembered. Not shown: where the photo came from (33 mentions), because there is no way to tell from a post whether that worked.
How this is calculated
- What is counted
- One thing remembered in one post: a post that remembers who and when counts twice (162).
- Formula
- Each kind of memory needs one feature: who → face recognition, when → the photo’s date, text → reading text in photos, where → location, what it looked like → AI image recognition. What was happening and why they took it have no feature. “Feature failed” = the same post also reports that feature failing. Unusable = failed + no feature = 85 + 16 = 101 of 162 (62%).
- Keep in mind
- “Failed” means the same post reports both, which is likely but not proven to be the cause. “Where it came from” (33 mentions) is left out because no failure type can show it breaking.
People remember moments, but search with one word
- What this shows
- Every search people quoted in their posts, copied word for word, plus what our survey respondents typed.
searches in public posts are one or two words, like “dog” or “selfie”
describe a moment or ask a full question
searches in our survey failed; each named an event and a year, like “Party 2018”
How this is calculated
- What is counted
- One search quoted in a post (118 searches in 59 posts), plus the 13 searches typed in our survey.
- Formula
- An AI copies each search exactly as the person wrote it; code keeps it only if those exact words appear in the post. “One or two words” = searches with 2 words or fewer: 98 ÷ 118. “Describes a moment” = built on an event or written as a full sentence: 6.
- Keep in mind
- Only posts where people quote what they typed are counted, so this describes searches people chose to share.
When the year doesn’t match, people can’t tell who is wrong
- What this shows
- From our survey: people wrote down the year before searching, then checked the year the photo shows.
How this is calculated
- What is counted
- One survey respondent who found the photo and wrote down a year (21 people).
- Formula
- Mismatch = year written before searching ≠ year the photo shows: 7 of 21. Each mismatch is split by the answer to “Is that when it was really taken?”: yes → memory was off (1), no → the app’s date was wrong (2), not sure → can’t tell (4).
- Keep in mind
- 21 people is a small group: the true mismatch rate is likely between 17% and 55% (95% Wilson interval). Updated as responses arrive.
What helps people find the photo, tested against chance
People who named a concrete memory (who, where, when, what) found the photo in 23 of 136 posts, against 16 of 233 who named nothing: very unlikely to be chance.
How this is calculated
- What is counted
- One post that says how the search ended.
- Formula
- Fisher’s exact test asks: if naming a memory made no difference, how often would a gap this big appear by chance? Here p = 0.004, about 4 in 1,000.
- Keep in mind
- This shows the two go together, not that one causes the other.