Buried: how people store and lose information in a chat interface
Semester 2 Design Research coursework · M.Des Interaction Design, NID Bengaluru · 3 weeks, February 2026
Intro
People do not lose information in WhatsApp. They lose the path back to it.
A three week design research study into how Indian users store and re-find information inside a chat interface that was never built to be an archive. It started from a complaint everyone recognises, important messages getting buried in group chats, and ended somewhere specific enough to design against: WhatsApp has quietly become the default Personal Information Management system for its users, while offering none of the structure a PIM system requires. Users have closed that gap themselves, by building private archives out of features meant for something else.
The course was about writing a hypothesis you can actually test, testing it hard enough to find out whether it holds, and documenting the findings so someone else can build on them. Mine did not hold, twice. No interface was shipped at the end of it, deliberately: what a product team needs before it draws a screen is a problem definition that has survived being tested, and the behavioural model underneath it.
Buried in the Stream: Personal Information Management and Re-finding in High-Volume WhatsApp Groups. An unpublished student paper written for the course in ACM SIGCHI format, with the full method, sample and references.
Key findings
Overload here isn’t about volume.
Retrieval fails in seven separable ways, not one.
The non-text archive is effectively invisible.
People build parallel archives that recreate the original problem.
When the system fails, people ask a person.
Scrolling isn’t a strategy. It’s the fallback nobody admits to.
Impact
What the research produced
Four outcomes, written as what the research produced rather than as metrics.
11
contextual enquiries, ages 23–31
59
survey respondents
7 + 4
failure modes, and repairs people built
3
weeks, start to finish
A reframed research question, twice
The study began as a chronology complaint and ended as a Personal Information Management question. Narrowing was the finding, not a detour.
A coded vocabulary of seven failure modes
Including Chronological Burial, Keyword Rigidity, Media Non-Indexability, False-Storage Assumption, Contextual Decay and Non-Linear Nav Deficit. Named failure modes are what let a team argue about a fix without arguing about the problem.
Three behaviourally derived personas
Built from a behavioural variable mapping across eleven participants, not from demographic buckets: The Systematic Archivist, The Nostalgic Searcher, The Selective Minimalist.
One counterintuitive finding that changes what you would build
When retrieval fails, users do not escalate to a better tool. They escalate to a person. Human memory is the actual search engine, and any feature that ignores that is competing with a friend who answers in thirty seconds.
Discovery
Three research questions, and why the first two were wrong
I brought three candidate questions to the first review. The one I was most attached to came back with two notes, and both were correct.
How do viral unverified allegations affect the wellbeing of the people they target?
Cut: too complex for the available scope.
It needed sociological fieldwork over months, not three weeks and a classroom.
How do viral unverified allegations spread and affect online communities?
Cut: not strongly within the scope of UX research.
A real question, and an important one, but a sociology question wearing a UX costume.
Struggles and strategies to find information buried in WhatsApp group conversations under information overload.
Kept: mundane, universal and observable.
The least glamorous option, and the only one I could watch happen in front of me in twenty minutes. Then it moved twice more.
The question, in three versions
- Question
Struggles and strategies to find information buried in WhatsApp groups under information overload.
Why it changed
Opening frame. Behaviour led, no mechanism named.
- Question
Exploring the limitations of chronological messaging interfaces in supporting information retrieval in high-volume WhatsApp groups.
Why it changed
Fieldwork pointed at chronology, so I named the mechanism. Too narrow: it presumed the answer.
- Question
How do Indian users employ diverse Personal Information Management strategies to facilitate data storage and retrieval within WhatsApp’s non-hierarchical interface?
Why it changed
Chronology alone could not account for self-groups, social outsourcing or emotional retention. Opened the frame back up around the behaviour, with a literature grounded construct.
Takeaway
From a broad behaviour problema presumed mechanisma literature grounded research frame.
Version two was my favourite and it was the weakest. It named a culprit before the data had finished talking.
Fieldwork
Contextual enquiry, in the room where it happens
Eleven participants, M.Des students aged 23–31, interviewed in their own classroom environments during live academic schedules rather than in a booked room. The setting matters: their groups were active while we talked, so message dumping happened during the session instead of being recalled afterward.
Each session ran 20–25 minutes and collected four data types at once.
01
Verbal response
To a semi-structured guide covering usage and context, storage behaviour, retrieval practice, emotional and motivational drivers, group dynamics, deletion and storage management, and attention.
02
Live retrieval tasks
I asked each participant to find a specific artefact in front of me: a document, a graduation photo, a voice note, an ID card. This is the part that produced the real data.
03
Interaction observation
Which affordance they reached for first, how long before they abandoned it, what they did next.
04
Emotional reaction during failure
Logged separately, because the affect was doing analytical work.
The instrument
Research interview guide
Lost in the Chat: exploring the UI limitations of chronological messaging interfaces in supporting information retrieval and management within high-volume WhatsApp groups
To identify and analyse the UI limitations of chronological messaging interfaces in high-volume WhatsApp groups that hinder effective information retrieval and management, and to generate design insights for improving information accessibility and organisation.
- Date
- 14 February 2026
- Researcher
- Akhil Cleatus
- Participants criteria
- NID students and faculty
- Purpose
- Exploratory: routines, meanings and problems
- Type
- Semi-structured
- Setting
- Their own classroom, during live academic schedules
How a session was planned
- Introduction5 minWelcome, purpose, consent, recording
- Warm-up5 minIcebreakers, background
- Core questions35 minMain research questions
- Wrap-up5 minFinal thoughts, participant questions
The script
Introduction
- Welcome, introduce, purpose, consent, confidentiality, session outline
Warm-up and background
- Tell me about yourself.
- How frequently do you use WhatsApp?
- Which categories are those active groups?
- What are your main goals when using WhatsApp groups?
Main interview questions
- How often do you use WhatsApp in a typical day?
- What types of groups are you part of?
- In very active groups, how do you deal with too many unread messages?
- How do you feel about messages being shown only in time order?
- If you need to find something important in a group, what do you usually do first?
- What makes it difficult to find information in group chats?
- Have you ever failed to find something important? What happened?
- How do you usually save important messages or files?
- Do you regularly delete chats or media? Why or why not?
- Do you behave differently in family, friends and academic groups?
- If you could improve WhatsApp to make finding information easier, what would you change?
Wrap-up
- Anything else to share?
- Anything I didn’t ask?
- Questions for me?
- Thank you.
Planned at fifty minutes. The sessions as run came in at twenty to twenty-five.
Self-report said search is mostly fine and the live task said otherwise, and not one participant described the difference as a failure of WhatsApp rather than as something they are bad at.
What the task showed
Content type, through the method they tried, to how it ended
All 22 episodes. Press any card to follow it through.
Every episode that ended inside a minute was a keyword search, and all four were in a self-chat. Not one retrieval from a group conversation was fast, whatever the group's size or the age of the thing being looked for.
Research evidence
Retrieval attempts
How people find, and fail to find, information in WhatsApp groups.
22 attempts · 11 participants · 26 coded variables
- 11Participants
- 22Retrieval attempts
- 26Variables per attempt
- 11 of 22Took over five minutes, or never ended
Retrieval outcome
Across 22 attempts
- Found quickly418%
- Found eventually732%
- Found after a struggle418%
- Not found29%
- Search abandoned523%
Retrieval method used
Across 22 attempts
- Scroll/Browse11
- Search (keyword)8
- Scrolled media thumbnails3
Chat volume
Across 22 attempts
- Medium (50-200/day)13
- Low (<50 msg/day)7
- High (>200/day)2
Individual attempts
22 attempts from 11 participants, 26 coded variables each. The name column stays put; scroll sideways for the rest of the sheet.
The values, the column titles and the colour of every chip are the sheet's own, read back off the research board rather than redrawn. The log below is this table, read as sentences.
The live retrieval task, coded
I asked each participant to find one specific thing, there and then. This is every attempt.
15 of 22 attempts ended with the thing in hand.
- Found quickly4
- Found eventually7
- Found after a struggle4
- Not found2
- Search abandoned5
Filter to their own self-chat and six of seven attempts end with the file in hand; the one failure is a file that had corrupted rather than a search that failed. Filter to a sticker and neither was ever found. Every failure in the log happened in a chat someone else fills, and every one of them was for something other than text. In five of the twenty-two the next move was to ask a person, which is the study’s least comfortable finding: when the interface runs out, people do not reach for a better tool.
The same attempts, counted
Where retrieval fails, and what it fails on
One dot per attempt, all twenty-two of them. Counts rather than percentages: at a sample this size a percentage would claim a precision eleven participants cannot give.
By where the thing was being looked for
- Own self-chat
- 6 of 7 found
- Small group
- 9 of 13 found
- Large group
- 0 of 2 found
By what was being looked for
- Text and documents
- 6 of 8 found
- Photos, voice, stickers
- 9 of 14 found
By how they went looking
- Keyword search
- 6 of 8 found
- Scrolled the chat
- 6 of 11 found
- Scanned thumbnails
- 3 of 3 found
- Found, under a minute
- Found, 1 to 5 minutes
- Found, over 5 minutes
- Abandoned or never found
The taxonomy, tested against the attempts
Each kind of thing fails in its own way
What was being looked for, against what stopped the search. If retrieval failure were one undifferentiated problem, this grid would be crowded down a single column. The claim the taxonomy makes is that it is not.
One dot is one attempt. 9 of the 48 cells carry any, which is the point: if this were one problem rather than several, every dot would be in one column.
Asked about search, people defend it. Asked to use it, people lose fifteen minutes and then apologise for themselves.
User insights
The eleven, in their own words
11 participants
- The Storage Illusion3 participants
“If I’m not tagged, I usually don’t read it.”
Ashish Aravind Ansari, 27
“Keyword search helps only for text; images remain hard to find.”
Parag Gawade, 31, Mumbai
“Most important files must be somewhere in my chats.”
Zain, 26, Kerala
- Effort Outsourcing2 participants
“Global search shows too many unrelated results.”
Nila, 23, Chennai
“Data sorting should be easier, especially for sensitive info.”
Srilekshmi Gireesh, 24, Kerala
- Social Entrapment2 participants
“Deleting WhatsApp could mean losing everything.”
Vidushi Sharma, 26
“Group pinning helps but isn’t widely used; AI summarisation may strip emotion.”
Samreen, 23, Delhi
- Attention Fragmentation1 participant
“Date-wise filtering like Google Photos would help.”
Abhijith Joy, 25
- Search Pessimism3 participants
“Search is generally sufficient, but similar keywords slow retrieval.”
Achute K, 23
“Audio and photo search is frustrating.”
Niveditha MR, 24, Kerala
“Voice notes are hard to track; starred messages are forgotten.”
Bharath Krishna, 26, Kerala
Secondary research
The construct that organised everything
I anchored the study in The Reappropriation of Instant Messaging: Texting Ourselves, Message Dumping, and Revisiting Conversations (Proceedings of the ACM on Human-Computer Interaction, Volume 6, Issue MHCI, September 2022). It gave the field’s existing vocabulary for behaviour I was watching raw, and let me run a structured comparison rather than a summary.
Personal Information Management, in the literature and in the field
The useful word was reappropriation. Users are not misusing WhatsApp. They are extending it, and the extensions are legible, repeatable and shared across people who have never discussed them.
Card sort
Letting mentors break the clusters
Key terms were extracted from every transcript onto cards and clustered with senior mentors in the room. The mentors’ contribution was mostly destructive and mostly correct: two of my early clusters were restatements of my own hypothesis rather than groupings of participant language. Removing them is what produced version three of the research question.
- Extract key termsKey terms were taken off every transcript onto cards, in the participants’ own words.
- Cluster with mentorsThe cards were grouped and argued over with senior mentors in the room.
- Throw two clusters outTwo were restatements of my own hypothesis rather than groupings of participant language. Removing them produced version three of the question.
What came off the transcripts
110 coded behaviours, before clustering
- Checking
- Rechecking
- Selective-checking
- Priority-checking
- Ignoring
- Muting
- Skipping
- Opening
- Closing
- Leaving-unread
- Marking-read
- Late-reading
- Night-checking
- Hourly-checking
- Scrolling
- Fast-scrolling
- Slow-scrolling
- Endless-scrolling
- Date-scrolling
- Skimming
- Scanning
- Jumping
- Backtracking
- Thread-following
- Reply-tracing
- Searching
- Keyword-typing
- Guess-searching
- Name-searching
- Year-searching
- Retry-searching
- Media-browsing
- Link-checking
- Group-switching
- Common-group-search
- Audio-playing
- Audio-replaying
- Finding
- Refinding
- Locating
- Failing-to-find
- Giving-up
- Retrying
- Cross-checking
- Confirming
- Verifying
- Saving
- Resaving
- Screenshotting
- Forwarding
- Self-forwarding
- Downloading
- Redownloading
- Backing-up
- Restoring
- Deleting
- Bulk-deleting
- Clearing-chat
- Archiving
- Starring
- Unstarring
- Ignoring-starred
- Forgetting-starred
- Pinning
- Unpinning
- Drafting
- Redrafting
- Testing-stickers
- Storing
- Hoarding
- Accumulating
- Keeping
- Organising
- Not-organising
- Sorting
- Not-sorting
- Labeling
- Not-labeling
- Captioning
- Not-captioning
- Forgetting
- Remembering
- Recalling
- Misremembering
- Assuming
- Expecting
- Depending
- Outsourcing-memory
- Calling
- Asking
- Clarifying
- Delegating
- Requesting-summary
- Explaining-context
- Avoiding
- Withdrawing
- Observing
- Lurking
- Silencing-self
- Hesitating
- Overthinking
- Venting
- Not-sending
- Reacting
- Emoji-reacting
- Sticker-reacting
- Contextless-replying
- Conflict-reading
- Gossip-reading
- Nostalgia-reading
The terms in black are the ones that survived into the seven named failure modes and the four repair strategies beside them. Starring, Ignoring-starred and Forgetting-starred were three separate observations before they became one: Star-Vaulting.
The wall
Seventy-nine observations, grouped under the prompt that produced them
79 notes under 13 prompts
why u use search bar in whats app group
In chronologic interface when using search option and multiple item come a list how do u confirm what u real want , do u use arrow to go next
what make you skip , ignore and close the group without reading content
How are u retrive audio files
what are the main thing u set as a key factor for recalling and getting the message , is it thumpnain , date , file name , key words etc
whats the maximum time u invest in retrieving information from whatsapp
what are the assumentions u made about the data u had, thinking that it was ther but not able to find in some urgent situation —
What are the method u use to same whats app information to your phone storage — Screenshots — download chat — image saving etc — —
what are the major times your whats app checking increases — morning — night — events — personal preferred message
why do u use Mark as Read feature on whats app.
why do u use persons common group search for
what are the problems u face when retrieving self-forward message in whatsapp
Which u dont prefer doing while searching for informations
Media-browsingRetry-searchingSearchingClosingIgnoring-starredSkippingLeaving-unreadLate-readingAudio-replayingVerifyingConfirmingKeyword-typingOpeningFindingSkimmingAvoidingMutingIgnoringNot-sendingForgettingAudio-playingForwardingYear-searchingName-searchingGuess-searchingCallingRecallingArchivingSavingVentingOverthinkingHourly-checkingFailing-to-findGiving-upRecheckingCheckingReply-tracingSortingDownloadingAssumingExpectingBacktrackingLocatingNight-checkingRetryingCross-checkingPriority-checkingSelective-checkingPinningStarringScreenshottingRefindingMarking-readSelf-forwardingBulk-deletingDeletingUnstarringUnpinningReactingEmoji-reactingSticker-reactingCommon-group-searchGroup-switchingNot-sortingNot-captioningCaptioningNot-labelingAccumulatingNot-organisingHoardingRestoringBacking-upRedownloadingResavingScrollingFast-scrollingSlow-scrollingEndless-scrollingDate-scrollingThe wall after clustering, at the positions it was left in. Two groups that were my own hypothesis restated came off it.
Survey
Testing whether eleven people were a coincidence
The qualitative work gave me mechanisms. It could not tell me whether they generalised past one campus. I wrote a survey instrument straight out of the observed behaviours rather than out of opinion, and ran it to 59 respondents.
What came back, read off the response sheet
59 responses
- Long-term habituation
- 44 of 59have used WhatsApp 5 years or more75%27 of 59have used it for more than 8 years
- Intensity
- 29 of 59describe their usage as almost constant49%
- Primary retrieval attempt
- 39Search bar39Media, Links and Docs15Manual scrolling
- Recall cue depended on
- 38Keywords33Sender identity23File name
- Workarounds in active use
- 37Starring23Forwarding to a personal group21Screenshots15Pinning chats12Saving to phone storage
First thing they reach for
Primary retrieval attempt, out of 59
- Search bar39
- Media, Links and Docs39
- Manual scrolling15
This is what people reach for first, not whether it worked: the survey asked about the attempt, and the outcomes are in the retrieval task above. The tie is the finding. The media tab is opened exactly as often as search is used, and fifteen people start by scrolling the chat by hand. Respondents could name more than one, so the counts run past 59.
Workaround stacking
Workarounds in active use, out of 59
- Starring37
- Forwarding to a personal group23
- Screenshots21
- Pinning chats15
- Saving to phone storage12
The bars sum to 108 across a sample of 59. Almost everyone runs more than one workaround at a time, which is the clearest evidence that none of them works.
The two rated scales, against each other
Feeling buried and feeling able to find it are not the same thing
Every respondent rated both: how strongly they agreed their information gets buried, and how confident they were of retrieving a file shared weeks ago. The two were rated separately, which is why these can be plotted against each other at all.
One circle per answer pair. 56 of 59 responses rated both scales; the other three left the grid blank.
More respondents →37 of 59
star messages, the most adopted workaround in the survey
0 of 11
returned to the starred list during a live retrieval task
Those two numbers come from two different instruments, which is why they are set side by side rather than plotted on one scale: one is a count from the survey, the other is an observation from the room. Starring is the most adopted workaround and the least used one. Adoption is not use, and that gap is only visible because the two methods were run against each other rather than in sequence.
I took this for a surprise until I went looking. It is not one: Whittaker’s study of 345 mailboxes found no correlation between how diligently people filed and how well they found, and the gap I measured in chat is that result again in a different app. That is the better reading of it. A three week student study with eleven participants cannot discover a new law of behaviour, but it can be sharp enough to detect a real one, and the methods disagreeing with each other is what made it visible here.
Synthesis
Three passes, each producing a different artefact
Pass one: empathy mapping
Say, Think, Do, Feel across all eleven participants, with retrieval strategy, key challenges and participant suggestions carried alongside. The Feel column clustered into five emotional patterns that I used as the working labels for the rest of the study.
The Storage Illusion
Belief that files are safely in WhatsApp, with no retrievability to support it.
“Most important files must be somewhere in my chats.”
Zain, 26, Kerala
Effort Outsourcing
Retrieval load pushed onto other people or other apps.
Social Entrapment
Cannot leave or restructure groups for social reasons, so volume is permanent.
“Deleting WhatsApp could mean losing everything.”
Vidushi Sharma, 26
Search Pessimism
Learned expectation that search will fail, so it is not attempted for media at all.
Attention Fragmentation
Retrieval attempts abandoned mid-way because the interface itself is the distraction.
Pass two: pattern finding and the coded taxonomy
Every participant was coded across usage, saving behaviour, retrieval behaviour, what they retrieve, feelings about the chronological interface, tool failure, workarounds and proposed improvements. Repeated codes became the study’s vocabulary, and this is the single most reusable thing the project produced.
- Code responsesAll eleven participants
- Find patternsGroup the repeated codes
- Build the taxonomyOne shared vocabulary
- Used across the studyGuides analysis and design
The coded failure taxonomy: seven ways retrieval breaks
And four repairs people built for themselves
The same behaviours, in the field's own words
Information retrieval has studied this in email for twenty years, under the name known-item search: re-finding one particular item you know exists. Mapping my names onto that literature is what tells me which of these are a chat interface lacking something mail has had for years, and which are a replication of a result already published.
Behaviour
What it means
Findings from literature
Chronological Burial
Known-item search under time-ranked retrieval
Across 3,678 email re-finding tasks, time ranking was the weakest of four conditions and caused most of the failures on older targets. WhatsApp offers time ranking only.
Mackenzie et al., WWW ’19
Keyword Rigidity
Exact match against short queries
Email queries average about 1.5 terms, and advanced operators appear in under 1% of them. People arrive with two words and no syntax.
Carmel et al., 2017
Star-Vaulting
Foldering versus finding
Across 345 mailboxes, filing behaviour showed no correlation with success at finding. Organising and retrieving are not the same skill, and one does not buy the other.
Whittaker et al., 2011
Non-Linear Nav Deficit
Faceted filtering and relevance ranking
Hybrid interfaces, relevance results above a chronological list, gained 12 to 18% MRR over chronological alone in two production mail clients.
Carmel et al., 2017
Media Non-Indexability
Classification and information extraction
Mail clients reach content search cannot by labelling it and extracting from it. A chat stream does neither, so media is outside the index rather than low in it.
Bendersky et al., 2021
Me-Only Group
Personal archiving as a PIM task
20 to 35% of email activity is organising, against 10 to 20% searching. Building the archive is the larger cost, and my Archivists are paying it in an app that gives them nothing back for it.
Narang et al., 2017
Pass three, the information storage and retrieval framework. A hierarchical tree of the retrieval process converted into a tracking table: every branch, self-chat versus group, search versus scroll, found versus abandoned, became a column. This did four things worth naming, because this is the step that turns interview notes into data: it provided a logical structure, it revealed what to measure, it surfaced limitation patterns, where branches such as audio unsearchable and thumbnail too small became both a pain point category and a tracked variable, and it created a taxonomy consistent enough to compare all eleven participants on the same axes.
Information storage and retrieval in WhatsApp
35 branches
The tree before it became a table. Every branch here turned into a column, which is what made eleven interviews comparable.
Personas
Behavioural variable mapping, then personas
Participants were positioned on continuous behavioural spectra, Activities, Attitudes, Motivations and Skills, in the Cooper sense, and clustered on where they landed rather than on who they were. Three clusters held.
The clusters the participants fell into
Cluster
Definition
Key behaviours
Participants
The Personal Archivists
Users who try to fix the retrieval problem by building their own manual database outside the standard chat flow.
- Self-only groups and personal chats
- Forwards media and documents to specific people to tag them
- Saves credentials in starred messages
- High storage use
- Ashish
- Abhijith
- Srilekshmi
- Zain
- Parag
The Social-Manual Searchers
Users who find the interface too difficult to navigate alone and rely on social connections or manual scrubbing.
- Asks friends for summaries or resends
- Scrolls manually for 10 to 15 minutes through thumbnails
- Uses the media tab rather than the search bar
- Calls people when the system fails
- Nila
- Niveditha
- Samreen
- Zain
The Selective Minimalists
Users who manage overload by cutting out noise and only interacting with high-relevance text.
- Ignores and archives groups to avoid mental clutter
- Relies on exact keyword search
- Uses @ mentions as the filter for importance
- Gives up on retrieval inside 30 seconds
- Ashish
- Achyuth
- Bharath
- Parag
Three coping strategies for one structural absence
Persona
Core behaviour
Dominant failure modes
The Systematic Archivist
Builds and maintains multiple topic-based self-groups, tags with captions and hashtags, values efficiency over social interaction. Treats WhatsApp as a vault.
- Keyword Rigidity
- Media Non-Indexability
The Nostalgic Searcher
300 plus groups, emotionally attached to the archive, will not leave groups, retrieves by manual scrolling and by asking people, 15 minutes plus per search.
- Chronological Burial
- False-Storage Assumption
The Selective Minimalist
Opens the app 100 plus times a day, abandons retrieval after roughly 30 seconds, ignores anything they are not tagged in, mutes and archives aggressively.
- Non-Linear Nav Deficit
- Contextual Fragmentation
One retrieval, drawn three ways
The Systematic Archivist
1 of 3
A senior asks for an Aadhaar card. Global search returns every other Aadhaar in the account, because the file was never labelled. The route that works is group, then Media, then Docs, which the interface does not suggest. Ends with a resolution to rename things properly next time.
The three are not three kinds of person. They are three coping strategies for one structural absence, and each one fails differently, which is exactly what a design brief needs in order to be scoped.
Solution
The deliverable is the problem, stated precisely enough to design against
Four positions came out of the study.
01
This is a PIM problem, not a notification problem
The instinct with group overload is to fix the inbox: smarter notifications, better unread handling, AI summaries of what you missed. The data says the pain is not in the moment of arrival, it is in the moment of return, days or weeks later. Every workaround in the study, self-groups, screenshots, starring, saving to phone, exists to serve future re-finding. Users are not asking to read less. They are asking to find again.
02
Retrieval is memory-dependent, so the interface fails exactly when memory does
Search in WhatsApp requires the user to already hold the key: the exact word, the sender, the group, the approximate date. Recall cues in the survey were keywords, sender and file name, in that order. This is a system that works best for information you have barely forgotten, and which collapses for anything genuinely old. Contextual Decay is the mechanism, and it means any real fix has to generate retrieval cues the user did not have to store themselves, through metadata, transcription, media indexing or date and type filters.
03
Non-text content is effectively write-only
Text is imperfectly searchable. Images, video, voice notes, stickers and map locations are not searchable at all, and participants had stopped trying. Voice notes were retrieved by playing every audio message in sequence. Map links appeared under Links with no visible place name, so each had to be opened to identify. One participant abandoned the search for her grandmother’s photograph entirely. The most emotionally valuable content in the archive is the least retrievable content in the archive.
04
Any feature competes with a friend, and the friend is currently winning
Social Outsourcing was present in most participants: call a friend, message one person for a summary, ask the sender to resend, have someone post on your behalf. It is fast, context-aware and socially cheap in a close group. A summary or search feature does not have to be good in the abstract, it has to beat thirty seconds of asking someone. That is a far harder bar than the usual framing admits, and it also marks the one place where automation was actively rejected: one participant refused AI summarisation on the grounds that it would normalise a capitalised “OMG” and strip the emotional register she reads groups for. Fidelity of tone is a requirement, not a nice-to-have.
Where a design intervention would start
Four openings, in order of leverage against the coded failure modes
Date and type filtering on global search
Against Non-Linear Nav Deficit. The most requested improvement in the study, and the cheapest.
Media indexing
Captions, transcripts for voice notes, visible place names on location links, and clustering for images, against Media Non-Indexability.
A restructured starred or saved view
With labels and a layout that is not a chat list, against Star-Vaulting. Users have already adopted starring. The retention failure is in the return path, not the save.
Relevance signalling inside groups
So that content matching a declared interest surfaces without the user monitoring the stream, against Chronological Burial. Proposed by a participant, not by me, and the most speculative of the four.
Reflection
What the study taught me, and what I would do differently
What the study taught me about method
Interviews told me what people believe about their tools. Tasks told me what their tools actually do to them. The gap between the two was the entire finding, and I would now design that pairing in from the first session instead of discovering it by accident in the third.
The mistake I am keeping in the case study
Version two of my research question named chronology as the culprit. It felt like progress, because a sharp question looks more rigorous than a broad one. It was actually me closing the frame around a hypothesis the data had not finished supporting, and the card sort with mentors is what caught it. Premature narrowing is a failure mode I now watch for in my own process the way I watch for it in a participant’s.
What I would change
I would run the survey earlier and smaller, as a screener rather than as validation, and use it to recruit for the extremes of the behavioural spectra instead of taking the eleven participants available to me. Recruiting from extremes and mainstreams would have sharpened the persona clusters, which are currently derived from a cohort that is more homogeneous than I would like.
What comes next
The framework is built to be designed against, and the obvious next phase is the one the study scoped but did not reach: prototype the date-and-type filter and the restructured saved view, then run the same live retrieval tasks against them with the same eleven participants. The tasks are already written, the baseline times and abandonment behaviour are already recorded, and the coded failure modes give the evaluation its measures. That is an evaluation design, not a wish, which is the position I wanted the research to leave the project in.
The position the project holds
Billions of people have made their primary messaging app their primary filing system without anyone designing it as one. Reappropriation at that scale is not user error that needs correcting, it is an unmet requirement that has been met badly for a decade. Naming the seven ways it fails is the start of meeting it properly.
Next project
A tabletop pet that guards your phone
Charli · Physical computing · 2026
