Personalized Feed Audit Protocol

Personalized Feed Audit Protocol

Purpose

This protocol helps one person observe what a recommendation surface exposed during a bounded period. It does not reverse-engineer hidden weights, diagnose a person, or measure platform-wide effects.

Observation design

Choose one surface, one account state, one device, one time window, and a fixed sample rule. A workable design is the first twenty eligible items seen at the same time each day for seven days. Do not interact for the purpose of changing the feed until the baseline sample is complete.

Record only what is necessary. Do not collect commenters' personal information, protected characteristics, private messages, or speculative mental-health labels.

Controls and intervention

TikTok describes "Not interested," creator or sound controls, and recommendation diversification in its For You explainer. YouTube documents "Not interested," "Don't recommend channel," history controls, and fewer-Shorts feedback in Manage your recommendations and search results.

After the baseline, make one documented change. Examples include following two well-sourced publishers outside the dominant source cluster, using a current platform control, or pausing history during unrelated research. Hold other behavior as steady as practical, repeat the sample, and compare exposure rather than judging success from one surprising item.

Interpretation rule

A narrow sample can show concentration, repetition, and change within that sample. It cannot reveal why the system ranked an item, whether another user saw the same surface, or whether exposure changed beliefs. If the audit causes distress or compulsive checking, stop. The guide is an information-literacy exercise, not health advice.