Nine research questions, answered
- What this shows
- Each answer comes from the evidence below, with its source. Slides cite these as Engine RQ1 to RQ9.
- RQ1
Where in a search does it break?
When the app tries to understand it: 364 of 505 posts. Once a first try fails, only 9 of 82 people who tried again found the photo.
Public postsEvidence → - RQ2
How do people word a search when memory is partial?
As one word about a thing: 98 of 118 real searches are one or two words. Only 6 describe a moment.
Public postsEvidence → - RQ3
What do people remember about a lost photo?
Where it was (16 of 30) and what was happening (16 of 30) top our survey, ahead of who (13) and roughly when (12).
Our surveyEvidence → - RQ4
What do they forget or get wrong?
The year: 7 of 21 found photos showed a different year than remembered. In 1 the memory was off, in 2 the app’s date was wrong, and in 4 people couldn’t tell which.
Our surveyEvidence → - RQ5
Can Google Photos use what people remember?
Mostly not: 101 of 162 things remembered hit a feature that failed or doesn’t exist. “What was happening” has no feature at all.
Public postsEvidence → - RQ6
Why don’t people just search?
17 of 30 didn’t try the search box. The top reason: “didn’t know what to type” (7).
Our surveyEvidence → - RQ7
Which photos are hardest to find?
Photos that weren’t taken as memories: 46 posts about saved or shared images going missing, and quick-note photos (orders, lab results) never found in 6 of 8. Survey groups are still too small to compare by type.
Public postsEvidence → - RQ8
What actually helps people find the photo?
Naming a concrete memory (23 of 136 found vs 16 of 233) and leaving the app to rebuild the date (8 of 26 vs 31 of 343). Both pass the chance test.
Public posts, testedEvidence → - RQ9
Where should Google Photos act?
On unreliable dates and on life events having no feature. General search quality is ruled out by the brief. Pattern finder: 6 testable ideas.
All sourcesEvidence →
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: saying what you remember, the app understanding it, judging the results, and recovering after a miss.
- How to read it
- Bigger bar, more posts break at that stage. The orange line under each count says how often those searches never found the photo.
- 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 people word their searches
- What this shows
- Every search people quoted in their posts, copied word for word, and what each search was built on: a thing, a date, a person, a full sentence.
- How to read it
- Bars show how many searches rest on each kind of detail, split into ones that worked and ones that didn’t. Real examples sit underneath.
searches quoted word for word in 59 posts. Each one is checked to really appear in its post.
are one or two words long, like “dog” or “selfie”. The typical search is 1 word.
describe a moment or ask a full question, which is how people actually remember photos.
What each search was built on
A search can rest on more than one kind of detail.
And in our survey
What the 13 people who searched actually typed, and whether they found the photo.
Examples from public posts:
People remember plenty. Google Photos can’t use most of it.
- What this shows
- Everything people said they remembered about a photo they couldn’t find, and whether Google Photos has a matching feature that worked. 101 of 162 things remembered ran into a feature that failed or no feature at all.
- How to read it
- Left: what people remembered, with the feature it needs written under it. Right: what happened. Orange means the feature broke, dark brown means the app has nothing to match it against. Hover a row to trace it, click a band to read the posts.
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.
What people remember, and what they forget
- What this shows
- Public posts can only show what people still recall. So in our own survey, 30 people so far wrote down what they remembered about a hard-to-find photo before searching, then checked the year the photo really shows. Early numbers, shown as counts; they update as responses come in.
- How to read it
- Left: how many people remembered each kind of detail. Right: how often the year they wrote down was wrong, split by how sure they said they were.
What they remembered, and what they didn’t
Ticked before opening the app. Out of 30 people.
Did the year they remembered match?
For photos they found, the year they wrote down vs the year the photo shows. 21 photos checked.
- Said “sure”2/8year didn’t match
- Said “fairly sure”3/9year didn’t match
- Said “guessing”2/4year didn’t match
Overall 7 of 21 didn’t match (likely between 17% and 55%). 2 of 30 had no idea of the year at all. When the year was wrong, it was off by 1, 1, 1, 1, 2, 2, 4 years. Of the mismatches, 1 were memory errors, 2 were the app’s date being wrong, and in 4 people couldn’t tell which.
New problems the AI found that we didn’t look for
- What this shows
- The theme finder read all 505 posts and proposed problems outside our own categories. It then checked every post again and only counted a match when it could quote the exact words from the post.
- How to read it
- The big number is how many posts show that problem. The quotes underneath are the exact words that matched.
Saved and shared images go missing
Photos that arrived from somewhere other than the camera, WhatsApp/Messages, screenshots, downloads, saved images, other-device uploads, never appear, appear only sometimes, or land in a folder the user didn't know existed.
Why it matters: A large share of the photos people urgently need are received or saved rather than taken, and the app's camera-first model makes exactly these the hardest to retrieve.
Never found in 33 of the 37 posts that say how it ended.
“it is really uncomfortable way of hard to find photos from different sources then Camera” · Play Store
“Google Photo doesn't find (or automatically backup) printscreens, why?” · Reddit
Works differently on different devices
The same library behaves differently depending on where you look, phone app vs web vs desktop vs another app, so a photo, its date, its location or a whole feature is present on one surface and absent on another.
Why it matters: Users waste time doubting their memory or re-searching on a second device, and it hides real data from them without any error message.
Never found in 28 of the 36 posts that say how it ended.
“They kept telling me to search for some Gemini function in Preferences but the setting just isn't there on my phone!” · Reddit
“File-based categories like Videos and Creations work, but AI/content categories like Selfies, People, Places, and Things show no results at all” · Google support forum
New photos not yet searchable
About photos just taken, outside the brief’s focus on old photosSearch only works for photos that are backed up, synced or already indexed, so very recent photos or non-backed-up photos are unfindable until some invisible processing finishes.
Why it matters: People most often hunt for the photo they took minutes ago, and a silent indexing precondition makes search look simply broken.
Never found in 16 of the 17 posts that say how it ended.
“Most features are not available unless you backup your photo as everything is happening on Google's servers not on your phone including searching” · Play Store
“two weeks ago I added a description to one photo, but still this photo was not found by this description” · Google support forum
Can't search by file details
The user tries to retrieve a photo using technical or file-level handles they actually remember, filename, IMG number, file type such as dng/raw, camera or phone model, or whether it was edited, and search ignores them.
Why it matters: When content words fail, file attributes are often the only exact, verifiable handle a user has left, and the app treats them as meaningless text.
Never found in 18 of the 18 posts that say how it ended.
“Want to find a photo or folder with an exact name...doesn't work.” · Play Store
“What in the world is happening???? I can't search pictures by name????” · Play Store
Flashback photos you can't find again
The app itself showed the photo first, in a Memories card, 'X years ago today' notification, or cinematic highlight, and the user then cannot get back to that photo afterwards.
Why it matters: A user who has just seen the photo has maximum motivation and maximum certainty it exists, so failing to re-find it is the most trust-destroying failure of all.
Never found in 7 of the 9 posts that say how it ended.
“Can't find the pictures that were in a notification from Google Photos” · Reddit
“I can scroll all the way down and it only goes to December 2024 but my memory is clearly show stuff from 2022 and 2023” · Play Store
Photos taken as quick notes
The photo was taken as a functional note, an order, receipt, licence plate, lab result, ID, repair documentation, a screenshot to post later, and the user searches by why they took it or the job it belongs to, not by what is visually in it.
Why it matters: These retrievals are time-critical and task-blocking, and the purpose of the shot is exactly the thing no index captures.
Never found in 6 of the 8 posts that say how it ended.
“Just spent over 1 hr trying to find a photo I had taken of an order that the seller asked for me to upload a pic” · Play Store
“For work, I'd look up "maps" to pull up maps of locations I service, but the AI assumes I want Google Maps” · Play Store
Hidden patterns, and how we’ll test them
- What this shows
- The pattern finder ran 35 of its own queries over the 501 posts collected before the App Store reviews were added, comparing groups of posts to find ones that behave unusually. Every number below was counted by code, and the next section tests whether each difference is bigger than chance.
- How to read it
- Each card is one pattern, the numbers behind it, what we think it means, and how we’ll check it with real users. The label shows how much to trust it given the sample size.
People who describe a concrete memory of the photo they want (a person, a rough date, a place, what was happening) end up succeeding about three times as often as people who just complain that search is broken.
23 of the 155 posts that mention at least one concrete memory ended in the photo being found, versus 15 of the remaining 346 posts; overall only 38 of 501 posts end in success. Posts with a concrete memory are also far less ambiguous: 21 of 155 unclear outcomes versus 115 of 346.- What it suggests
- Successful searching depends on the person being able to convert a fuzzy memory into a usable handle, and today the app gives no help doing that conversion; those who can't articulate a handle fall straight into generalised frustration.
- How we’ll test it
- In sessions, ask participants to say out loud everything they remember about a target photo before touching the phone, then compare hit rate and time for people who produce two or more concrete details versus those who produce none; in a prototype, offer memory prompts ('roughly when?', 'who was there?', 'where?') and measure whether the low-detail group improves.
- Why this confidence
- Based on a large group (155 posts with concrete memories against 346 without) and a consistent gap in both success and clarity of outcome.
“Just spent over 1 hr trying to find a photo I had taken of an order that the seller asked for me to upload a pic.”
The users who did the most organising work themselves, file names, captions, descriptions, their own keywords, album titles, are the group that almost never finds the photo, because the app ignores the labels they created.
Only 1 of 43 posts about the app ignoring the user's own labels ended in success, against 38 of 501 overall; of the 7 of those posts where the person had also built albums, 0 of 7 found the photo.- What it suggests
- The product rewards passive users with automatic recognition and punishes the diligent ones, so the effort people invest in naming and filing becomes invisible, a strong betrayal signal and a likely driver of people leaving for file-based tools.
- How we’ll test it
- Recruit people who name files or write captions, have them search for a photo using their own label, and record whether it surfaces; in a prototype, make user-written names, captions and album titles first-class search matches and measure change in success rate for this group.
- Why this confidence
- 43 posts is a reasonable group and the pattern is extreme (1 success), though the sub-group with albums is only 7 posts.
“search hasn't even worked for years. Want to find a photo or folder with an exact name...doesn't work.”
Remembering the words printed in a photo, a receipt, a meme, a screenshot, a sign, is the least reliable memory of all, and this is reported overwhelmingly by Reddit users rather than by app reviewers or support users.
Only 1 of 24 posts where the person remembered words in the image ended in success, compared with 9 of 49 where they remembered roughly when it was taken; of the 32 posts reporting that word-in-photo search failed, 0 ended in success and 20 of 32 came from Reddit versus 4 of 32 from the support forum.- What it suggests
- Reading text inside pictures was a power-user superpower that appears to have degraded; the people who relied on it are a vocal, technical minority who notice the regression precisely because they used it daily.
- How we’ll test it
- Give participants a set of their own screenshots and documents and ask them to retrieve one by an exact word they can see in it, measuring hit rate and whether partial or pasted phrases work; instrument the prototype to log word-in-photo queries that return nothing despite a matching image existing.
- Why this confidence
- The clue group is only 24 posts, although the related failure group of 32 posts points the same way.
“Can't find text in images on Google Fotos anymore, super annoying”
Complaints that search returns an incomplete or bloated set of results, 'only 6 dog photos when I have hundreds', are voiced almost entirely in app reviews and on Reddit and essentially never taken to the support forum.
Of 60 posts about results being incomplete or swamped, 45 came from app reviews and 14 from Reddit but only 1 from the support forum, even though the support forum supplies 126 of all 501 posts; only 2 of those 60 ended in success.- What it suggests
- People treat 'the results list is lying to me' as a quality judgement to broadcast rather than a problem to report, so this failure is invisible in support channels while quietly destroying trust in every future search.
- How we’ll test it
- Ask participants to search for a common subject they know they photograph often, then show them the full ground-truth set and capture their reaction and whether they would ever report it; in the prototype, test a visible 'showing X of Y' or 'see all matches' affordance and measure repeat-search behaviour.
- Why this confidence
- 60 posts, and the source imbalance (1 from a channel that normally supplies a quarter of posts) is far too large to be noise.
“when I search for "dog", only 6 pictures come up even though I am SURE I have more dog photos”
Complaints about the conversational AI assistant are mostly about losing the old search rather than about a real photo hunt, these posters rarely describe any actual photo they were trying to find.
Of 49 posts mentioning the AI assistant, 42 flag an assistant problem and 7 complain a feature was taken away, but only 7 of 49 contain any concrete memory of a photo and only 4 of 49 describe hunting for one specific picture, against 155 of 501 posts overall that contain a concrete memory; just 3 of 49 ended in success, and only 3 of 49 came from the support forum versus 26 from Reddit and 20 from app reviews.- What it suggests
- The assistant backlash is as much about removed control and forced change as about retrieval quality, so improving answer accuracy alone will not calm it, restoring a plain, predictable keyword path may matter more.
- How we’ll test it
- Run a side-by-side session where participants complete the same retrieval task with the assistant and with a plain keyword list, and ask which they would choose and why even when the assistant succeeds; A/B a visible 'classic search' toggle and watch adoption and sentiment.
- Why this confidence
- 49 posts is a usable group, and the near-absence of concrete photo hunts (7 of 49) is a clear contrast with the 155 of 501 base rate.
“I don't have a conversation with AI I want to see all the pictures that come up with that word.”
Switching surface, going to the computer or leaving the app for a file browser, is one of the few things that actually works, and it is reported mainly by support-forum users rather than app reviewers.
5 of 17 posts where someone moved to a computer ended in success, and 2 of 13 where they searched outside the app, versus 38 of 501 overall; 11 of the 17 computer posts and 9 of the 13 outside-the-app posts came from the support forum, and the computer group is dominated by wrong dates (7 of 17).- What it suggests
- The phone app quietly hides or reorders things the web version shows, so experienced users have learned a 'go to the big screen' rescue ritual, a workaround that the app itself should be surfacing instead of leaving people to discover.
- How we’ll test it
- Watch what people do after two failed searches on the phone and count how many spontaneously reach for a laptop; then take ten real failed phone searches and re-run the identical query on the web to confirm whether results genuinely differ.
- Why this confidence
- Only 17 computer posts and 13 outside-the-app posts, so the raised success rate is suggestive rather than solid.
“I often find it filtering things out and I have to go to the desktop version to locate the missing photos”
Which patterns hold up, and which could be chance?
- What this shows
- Each key comparison tested with Fisher’s exact test, the standard check for small groups. Ranges show where the true rate most likely sits (95% Wilson interval).
- How to read it
- Blue means the difference is bigger than chance would produce. Grey means it could be chance, so we don’t build on it.
| Comparison | Group A | Group B | Real or chance? |
|---|---|---|---|
| Naming a concrete memory found the photo: posts naming who, where, when or what, vs posts naming nothing | 23 of 136 17%, likely between 12–24% | 16 of 233 7%, likely between 4–11% | very unlikely to be chance p = 0.004 |
| Leaving the app found the photo: people who went to a computer, another app or another person, vs everyone else | 8 of 26 31%, likely between 17–50% | 31 of 343 9%, likely between 6–13% | very unlikely to be chance p = 0.003 |
| Labelling photos themselves found the photo: people whose own captions or names were ignored, vs everyone else | 1 of 37 3%, likely between 1–14% | 38 of 332 11%, likely between 9–15% | could be chance p = 0.154 |
| Trying a second approach found the photo: people who tried two or more ways, vs people who tried one | 9 of 82 11%, likely between 6–20% | 30 of 287 10%, likely between 7–14% | could be chance p = 0.841 |
| Wrong dates by source mention the app's date being wrong: Google support forum vs Play Store | 28 of 126 22%, likely between 16–30% | 21 of 255 8%, likely between 5–12% | almost certainly not chance p = 2.7e-4 |
| Too many results by source complain of too many results: Play Store vs Google support forum | 41 of 255 16%, likely between 12–21% | 0 of 126 0%, likely between 0–3% | almost certainly not chance p = 2.6e-8 |
Each source tells a different part of the story
- What this shows
- How often each problem comes up in each source. Reviewers react to app changes, Reddit users hit the AI search, and people asking Google for help have wrong dates and places underneath.
- How to read it
- Darker orange means the problem is more common in that source. Percentages are out of that source’s posts about finding photos.
YouTube and the App Store are left out: only 2 and 4 of their posts were about finding a photo.