# How Buzzsift reports are made

## What it reads

- **TikTok**: Searches videos from the last 90 days, keeps those whose caption names the subject, and reads 2 pages of comments under the 6 with the most comments.
- **Reddit**: Searches posts (2 pages) and, separately, comments that mention the subject inside any thread (3 pages); then reads 2 pages of comments in the 4 busiest threads about it.
- **YouTube**: Searches this year's videos, keeps those whose title names the subject, and reads 2 pages of top comments under the 5 most viewed.
- **Instagram**: Searches reels (2 pages), keeps those from the past year whose caption names the subject, and reads 2 pages of the most liked comments under the 5 with the most comments.
- **X**: Searches the latest posts (3 pages) and the top ones (1 page) that name the subject, keeps those from the past year, and reads 2 pages of replies under the 4 with the most replies.

## How a report is made

1. **Read**: The posts and comments above, public only, with no platform login. A few hundred in all.
2. **Mark each one**: An AI model reads them in batches of 120 and marks every post: not about the subject (another meaning of the name, spam), or positive, negative, mixed, neutral. Posts marked as not about it are dropped.
3. **Count**: The sentiment shares, overall and per platform, are counted by code from those marks. The model never estimates a percentage.
4. **Find the points**: In each batch the model lists the points people make, each with the ids of the posts that make it. A final pass merges points that say the same thing, keeps their posts and ranks them by how many posts back them (up to 7 a section). Ids that don't exist are thrown away.
5. **Link**: Every point shows how many posts make it and up to 4 of them, linked, so you can read them yourself.

## Limits

- It reads a sample, not everything: the most viewed and most commented posts that platform search returns, so a quiet complaint on a small account can be missed.
- Recent only: TikTok the last 90 days, YouTube, Instagram and X the past year. Reddit search favors recent and popular threads.
- The model can misread sarcasm or slang. That's why each point links to its posts.
- Facebook, LinkedIn, news sites and blogs aren't read.
- If fewer than 5 posts are about the subject, the report stops and the points go back.

## The sample reports' numbers

| Report | Kind | Run | Searches and page reads | Read | About the subject | Time |
|---|---|---|---|---|---|---|
| [Stanley](https://sociallisteningai.com/samples/brand-stanley/) | Brand | 2026-11-14 | 55 | 716 | 397 | 67s |
| [Owala](https://sociallisteningai.com/samples/competitor-owala/) | Competitor | 2026-11-14 | 58 | 1022 | 606 | 104s |
| [standing desk](https://sociallisteningai.com/samples/painpoints-standing-desk/) | Pain points | 2026-11-14 | 57 | 753 | 479 | 88s |
| [Oura Ring](https://sociallisteningai.com/samples/brand-oura-ring/) | Brand | 2026-11-14 | 56 | 877 | 567 | 98s |
| [Yeti](https://sociallisteningai.com/samples/brand-yeti/) | Brand | 2026-11-14 | 59 | 1110 | 472 | 93s |
| [Hydro Flask](https://sociallisteningai.com/samples/brand-hydro-flask/) | Brand | 2026-11-14 | 57 | 867 | 610 | 88s |
| [Whoop](https://sociallisteningai.com/samples/brand-whoop/) | Brand | 2026-11-14 | 55 | 752 | 387 | 72s |
| [Notion](https://sociallisteningai.com/samples/brand-notion/) | Brand | 2026-11-14 | 55 | 843 | 531 | 88s |
| [Obsidian](https://sociallisteningai.com/samples/brand-obsidian/) | Brand | 2026-11-14 | 57 | 1049 | 620 | 93s |
| [Dyson](https://sociallisteningai.com/samples/brand-dyson/) | Brand | 2026-11-14 | 58 | 1086 | 851 | 112s |
| [Shark](https://sociallisteningai.com/samples/brand-shark/) | Brand | 2026-11-14 | 58 | 1289 | 201 | 67s |