Features
Part of Social media feeds: a practical guide to reading them well
Testing a social media feed decision against the evidence
Social media feed example: audit a seven-day sample, classify sources and recommendations, adjust controls, compare results, and state what remains unknown.
What to take away
- A personal feed audit describes one account and period, not a platform-wide rule.
- Define categories before sampling and preserve the same method after changes.
- Separate followed posts, recommendations, ads, and reposts.
- Record desired outcomes before using controls.
- Report uncertainty, news-cycle effects, and unresolved recommendation reasons.
This fictional case follows Maya, an adult who uses one social platform for local news, illustration, and friends. Names, counts, posts, and platform features are invented. The method does not expose proprietary ranking logic or assess mental health.
The question
Maya says, "My feed has become mostly suggested short videos. Can I make local updates and friends easier to find without closing my account?"
The question contains a goal and a constraint. The audit defines success as a larger share of posts from chosen local sources and friends in the first 30 feed items, with fewer unrelated recommendations. It does not require eliminating all discovery. That is step one of the feed process: a purpose sentence with measures attached.
Why classification matters
Platforms can categorize users and content in ways that people do not see or fully understand. Pew Research Center's study of Facebook algorithms and personal data asked users to inspect the interests the service assigned to them and found mixed accuracy and comfort. That dated Facebook study provides context for cautious interpretation, not a description of Maya's fictional platform.
Baseline method
Maya samples the first 30 posts in the Home feed at 8 a.m. and 7 p.m. for seven days. She does not include stories, notifications, search, or direct messages. She records only category counts and removes names from the worksheet. Removing names is audience management; the widest plausible audience includes whoever the worksheet eventually reaches.
Each post receives one source-path code:
- F: followed friend or family account
- L: followed local institution or reporter
- C: followed creative account
- R: recommended account not followed
- A: labeled advertisement
- S: share or repost from another source
She also records post age, format, topic match, and whether a recommendation label appears. Ambiguous items go into an unresolved field instead of being forced into a category.
Baseline findings
The 14 sessions contain 420 observed positions:
Show the numbers
| Friends and family | 63 |
|---|---|
| Local followed sources | 42 |
| Creative followed sources | 84 |
| Recommended accounts | 168 |
| Labeled ads | 42 |
| Shares and unresolved | 21 |
Short video occupies 231 positions, or 55 percent. Maya's purpose rating marks 176 positions useful, 92 partly useful, and 152 not useful for the three stated goals.
The baseline supports her observation that recommendations and short video are prominent in this sample. It does not show why each post ranked where it did. Why each post ranked where it did stays unknown; the checkable causes, from source behavior to app mismatch, are the practical substitute.
Changes
Maya makes four dated changes after the baseline:
- Adds eight local sources and ten friends to a favorites view.
- Unfollows twelve inactive promotional accounts.
- Uses "not interested" on unrelated recommended videos during normal sessions.
- Turns off nonessential recommendation notifications.
She does not block broad viewpoints or interact with unwanted posts for the sake of the test. She keeps the Home feed for discovery and checks Favorites once each morning for essentials. Splitting surfaces this way follows the use-case comparison: essentials in a maintained list, discovery on its own.
Follow-up method
After a three-day settling period, Maya repeats the same Home-feed sample for seven days. She also records the first 20 positions in the morning Favorites view, but reports that surface separately because it has a different candidate set.
A University of Rhode Island repository paper on teaching feed awareness describes a week-long lesson using social media diaries and platform comparison charts. Maya's case borrows the general practice of a dated diary, not the paper's participants, teaching outcomes, or conclusions.
Follow-up findings
The Home feed produces:
| Source path | Baseline | Follow-up | Change in positions |
|---|---|---|---|
| Friends and family | 63 | 76 | +13 |
| Local followed sources | 42 | 55 | +13 |
| Creative followed sources | 84 | 88 | +4 |
| Recommended accounts | 168 | 135 | -33 |
| Labeled ads | 42 | 45 | +3 |
| Shares and unresolved | 21 | 21 | 0 |
Useful ratings rise from 176 to 222 positions. Unrelated recommendations fall, but ads remain close to the prior share. Short video falls from 231 to 189 positions.
In the separate Favorites sample, 121 of 140 observed positions come from the selected accounts. Nineteen are interface notices, repeated posts, or other eligible items. Maya decides that Favorites is the better surface for a quick morning check.
What changed
The observed mix moved toward Maya's chosen sources, and the favorites surface gave her a more predictable route. She keeps the changes and schedules a monthly follow-list review.
She changes her wording from "the app ignores everyone I follow" to "in my first baseline sample, followed friends and local sources occupied one quarter of Home-feed positions, while recommendations occupied two fifths. After the listed changes, the followed share rose and the recommendation share fell."
Limits and alternative explanations
Several factors prevent a causal claim:
- the sample covers one fictional account
- the observed weeks had different local events
- followed accounts may have posted at different rates
- the platform could have changed or tested ranking
- category judgments include human coding error
- some shares and labels were unresolved
- a three-day delay may not isolate the timing of effects
The audit can guide Maya's personal setup. It cannot estimate what other users see or prove which internal feature caused the change.
A reusable evidence log
| Field | Entry |
|---|---|
| Question | Can chosen sources be easier to find? |
| Account and surface | One adult account, Home and Favorites |
| Baseline | 14 sessions, 420 positions |
| Changes | Favorites, unfollows, feedback, notifications |
| Follow-up | 14 sessions after three-day gap |
| Main result | More chosen-source positions, fewer recommendations |
| Unknown | Internal weights and platform changes |
Common questions
Why sample fixed positions instead of total session time?
Fixed positions make sessions more comparable. Time can vary with video length, reading, and interruption.
Why not test several changes separately?
That would improve causal interpretation but take longer. This personal audit evaluates a practical bundle and states the limitation.
Are percentage changes enough?
Counts, denominators, coding rules, and raw observations are also needed. Small samples can move sharply.
Did the ads increase?
The count rose by three positions, which may be ordinary variation. The case makes no causal claim.
Can Maya publish account screenshots?
Only after considering consent, privacy, copyright, platform rules, and whether text and names can be removed. Aggregate tables are safer here.
What would strengthen the test?
Longer sampling, preregistered categories, independent coding, platform-change records, and one change at a time would support narrower causal questions.