How to Audit a Personalized Social Media Feed

How to Audit a Personalized Information Feed

A personalized-feed audit is a private, bounded observation of what one account sees on one surface. It can reveal source concentration, repeated topics, familiar frames, visible disclosure, and changes after a control is used. It cannot reveal hidden ranking weights or prove that a feed changed anyone's beliefs.

The cleanest design is seven days of baseline observation, one documented change, and seven days of resampling.

flowchart LR
    A["Choose one surface"] --> B["Define a fixed sample"]
    B --> C["Observe for seven days"]
    C --> D["Measure concentration and repetition"]
    D --> E["Make one documented change"]
    E --> F["Repeat the same sample"]
    F --> G["Report what changed and what remains unknown"]

Begin with one answerable question

"Is my feed manipulating me?" is too broad to audit. "What percentage of the first twenty eligible Home-feed items came from my five most frequent sources this week?" is measurable.

Choose the platform, surface, account, device, time, and sample rule before opening the app. Home, Following, Search, Shorts, For You, comments, and notifications are different products. Mixing them destroys the boundary.

A practical rule is the first twenty eligible items shown at approximately the same time each day. Decide in advance how advertisements, pinned posts, repeated items, unavailable content, and live material will be treated.

Do not create a fake identity or use another person's account without permission. Do not collect commenters' names, protected characteristics, private messages, or personal details. The goal is to observe your information environment, not build a dossier on people inside it.

Record exposure before interpretation

The log should preserve what was visible without pretending to know why it appeared.

Write the frame as a description, not a verdict. "The post presents the policy as a cost increase" is usable. "The post lies about the policy" requires a separate evidence review.

The E023 comparison that motivated this guide involved roughly twenty TikTok and Instagram posts across two accounts. Dalton often saw different leading comments. That was an informal observation, not a controlled study. A better audit preserves timing, surface, ranking option, and account behavior before drawing a conclusion.

Keep the baseline as quiet as practical

Normal viewing behavior influences many recommendation systems. TikTok's For You explainer describes likes, shares, follows, comments, content creation, video information, and account or device settings among its signals. YouTube's recommendation-system explanation discusses watch history, searches, engagement, dismissals, subscriptions, and satisfaction.

You cannot freeze every signal, and the platform may run its own experiments. You can avoid deliberately reshaping the feed during the baseline. Record unavoidable actions such as opening an item to identify its source.

Do not infer a hidden weight from the order. Even if every item about a topic follows one interaction, the observation cannot isolate timing, broader popularity, similar-user behavior, inventory, moderation, or an experiment.

Calculate concentration and repetition

At the end of seven days, count how many sampled items came from each source and topic. Identify the share supplied by the five most frequent sources. Count how many sources were new, how many items repeated the same frame, and how often visible commercial or synthetic-media disclosure appeared.

Compare days rather than treating the whole sample as timeless. A major news event can legitimately concentrate a feed. A weekday routine can differ from a weekend. Missing a day should be visible rather than backfilled from memory.

The result can say, "In this 140-item sample, five sources supplied 46 percent of observed items." It should not say, "The algorithm gives me only five sources." The first statement names its denominator. The second exceeds it.

Make one documented change

Choose one intervention that matches the question. Follow two well-sourced publishers outside the dominant cluster. Mark a repeated topic as not interesting. Remove an accidental watch-history item. Pause history while conducting unrelated research. Use only a current control that the platform documents.

YouTube's recommendation controls explain "Not interested," "Don't recommend channel," watch and search history, and related settings. TikTok's recommendation explainer describes "Not interested" feedback and other curation actions. Recheck the current interface before giving procedural instructions because labels and availability can change.

Make only one planned change if the objective is comparison. Multiple simultaneous changes may improve the experience, but they make attribution harder.

Repeat the same observation rule for another seven days. Record whether source concentration, topic distribution, new-source discovery, repeated frames, or disclosure changed.

Interpret the result without diagnosing yourself

A narrow audit can show that a pattern existed in the sample. It cannot establish the platform's motive, the complete set of eligible content, or a psychological effect.

The distinction matters because exposure and attitude do not move together automatically. A 2023 Nature field experiment reduced like-minded Facebook exposure for consenting US adult participants during the 2020 election. It changed exposure and engagement but did not produce a measurable change across eight preregistered attitude measures.

Your personal audit is far less controlled. Use it to improve questions and information habits, not to make clinical, political, or platform-wide claims.

If the exercise creates distress, compulsive checking, conflict, or pressure to monitor other people, stop. This is an information-literacy method, not health advice.

Turn the audit into a better research habit

The highest-value result may be noticing when a feed is being asked to do a job it cannot do.

A feed can introduce a claim. It should not be the complete research path for a decision involving health, money, safety, reputation, or civic judgment. When the stakes rise, write the claim in one sentence and continue with [[How to Compare Sources Outside a Recommendation Feed]].

[[Filter Bubbles and Personalized Feeds Explained]] provides the larger model. E080 examines platform-shaped discovery in [[What Is Zero-Click Search and What Does It Measure]]. E084 examines repetitive synthetic supply in [[What Is an AI Slop Feed]].

This audit guide was developed with AI assistance from the preserved E023 captions, current TikTok and YouTube documentation, peer-reviewed Facebook research, and the linked audit protocol. Dalton Anderson remains the author. Platform surfaces and controls can change. Editorial, research, platform, privacy, source, accessibility, and founder review are required before publication or use. Publication is not authorized.