> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://api.docs.siftsy.com/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://api.docs.siftsy.com/_mcp/server.

# Scores

> What Sentiment, Content, Topic, Relevance, and Consensus actually measure — and how they roll up into the Siftsy Score.

The **Siftsy Score** is a 1–10 temperature check for a comment section: how people reacted to the content. It is not a like count and not a generic “positive or negative” label. It blends three independent readings, each scored by context-aware AI on the same 1–10 scale:

* **Sentiment** — how commenters feel about the post and the subjects it discusses
* **Relevance** — how on-topic the conversation actually is
* **Consensus** — how commenters are reacting to **each other**

Sentiment carries the most weight. Relevance and Consensus keep a high score honest: a cheerful but off-topic thread, or a pile-on fight, should not read as a clean win.

Every number you get from the Siftsy Analyst API is already stored. The endpoints never rescore. Read `GET /team` → `analysisSettings.scoreVersion` (and `engagementWeight`) so you interpret the rollup the same way the dashboard does.

> **Warning**
>
> **Consensus is not agreement with the post.** Sentiment (and Content / Topic) is how commenters feel about the post, creator, or subject. Consensus is how they treat other commenters. A thread can be uniformly negative toward the post and still have high consensus.

## Content vs. topic

A single sentiment number collapses two different stories: “I don’t agree with this video” and “I hate this subject.” On score version 2 and 3, Sentiment splits into two readings on every comment, then averages them 50/50 into overall `sentiment`:

| Score                       | Asks                                                   | High means                      | Low means                            |
| --------------------------- | ------------------------------------------------------ | ------------------------------- | ------------------------------------ |
| **Content** (`content`)     | How do they feel about *this post and its creator*?    | Backing / echoing the publisher | Pushback against the post or creator |
| **Topic** (`topic`)         | How do they feel about *the subjects being discussed*? | Enthusiastic about the subject  | Hostile toward the subject           |
| **Sentiment** (`sentiment`) | Both, combined                                         | The section landed              | The section did not                  |

The average is compatible with older versions. The split is the diagnosis.

A **high Content score with a low Topic score** means the audience is with the creator and against the subject — the execution worked; the message or the thing being discussed did not. The reverse means the subject resonates and the post is what is missing. Own-social teams usually read Content first. PR and crisis teams usually read Topic first. The underlying scores do not change when you switch that lens; only which story you lead with does.

## Relevance and consensus

These are supporting measures, not a second sentiment score.

**Relevance** is whether the comment is actually about the post, the creator, or what the post is talking about. High is a specific, on-topic reply. Low is spam, solicitation, or a comment that could sit under anything. A relevant criticism still scores high on Relevance.

**Consensus** is commenter-to-commenter. High is agreeing with or echoing other comments. Low is arguing with them — debate, contradiction, fighting. Never treat “agrees with the post” as high Consensus or “disagrees with the post” as low Consensus.

| Score         | Measures                                                         |
| ------------- | ---------------------------------------------------------------- |
| **Sentiment** | How commenters feel about the post and the subjects it discusses |
| **Content**   | Alignment with the post / creator (score version 2+)             |
| **Topic**     | Reception of the subject itself (score version 2+)               |
| **Relevance** | How on-topic the comment is                                      |
| **Consensus** | Alignment among commenters, not with the post                    |

## How they become a Siftsy Score

Per-comment scores are the same across versions. A version is a different way of rolling those comments up — switching versions on the team recalculates instantly and does not rescore.

|           | Version 1                                                      | Version 2                                       | Version 3                                                   |
| --------- | -------------------------------------------------------------- | ----------------------------------------------- | ----------------------------------------------------------- |
| Blend     | Sentiment, Relevance, and Consensus equally (1/3 each)         | Sentiment 66%, Relevance 16.5%, Consensus 16.5% | Same weights as v2                                          |
| Sentiment | One reading per comment                                        | Same as v1                                      | Content + Topic, averaged 50/50                             |
| Reading   | Immaculate 7.0+ · Good 5.0–6.9 · Sub-par 3.0–4.9 · Off below 3 | Same as v1                                      | **Immaculate** 7.0+ · **Mixed** 4.5–6.9 · **Off** below 4.5 |

Version 3 also hardens the rollup against the ways comment sections mislead an average:

* **Negativity cap.** If a meaningful share of comments are very negative (below 3), the score’s ceiling drops. A section that is about 12% very negative can score at most 8.5. A loud positive majority cannot paper over a hostile contingent.
* **Engagement weighting.** Comments with more likes and replies count more on a curve: every comment keeps at least 1× weight and the most-engaged comment counts up to 5×. That multiplier applies to the sentiment readings only. Relevance and Consensus describe the comments themselves, so likes do not scale them.
* **Reply hierarchy.** Top-level comments carry full weight. Replies share a capped pool, so a long argument under one comment does not outweigh the rest of the thread.

`scores.siftsyScore` is that rollup. `scores.sentiment`, `content`, `topic`, `relevance`, and `consensus` are the components.

## Mix buckets

`GET /posts/{postId}/metrics` (and the `mix` object on an analysis) returns five-bucket percentages for each dimension. The JSON keys are the same; the dashboard labels change with the score.

| Key                   | Sentiment     | Content    | Topic        | Relevance | Consensus  |
| --------------------- | ------------- | ---------- | ------------ | --------- | ---------- |
| `positivePercent`     | Positive      | Supportive | Enthusiastic | On-topic  | Agreeing   |
| `leanPositivePercent` | Lean positive | Receptive  | Favorable    | Relevant  | Discussing |
| `neutralPercent`      | Neutral       | Split      | Mixed        | —         | —          |
| `leanNegativePercent` | Lean negative | Skeptical  | Critical     | General   | Debating   |
| `negativePercent`     | Negative      | Opposed    | Hostile      | Off-topic | Fighting   |

`neutralPercent` is typically unused for Relevance and Consensus. Filter comments with the dashboard labels or the JSON keys — both work. See [Retrieving comments](/guides/core-concepts/retrieving-comments).

## Numeric cutoffs

These are the mix-bar ranges. A comment in a bucket is the same comment counted in that bar on `GET /posts/{postId}/metrics` and `GET /posts/{postId}/analysis`, and the same comment that can appear in a comments **sample** filtered by `?sentiment=…` (or `content`, `topic`, `relevance`, `consensus`).

| Key             | Five-bucket (v3 overall sentiment; always for content, topic, relevance, consensus) | v1/v2 overall sentiment |
| --------------- | ----------------------------------------------------------------------------------- | ----------------------- |
| `positive`      | ≥ 7                                                                                 | ≥ 7                     |
| `lean_positive` | > 5.5 and \< 7                                                                      | ≥ 5 and \< 7            |
| `neutral`       | 4.5–5.5                                                                             | unused                  |
| `lean_negative` | ≥ 3 and \< 4.5                                                                      | ≥ 3 and \< 5            |
| `negative`      | \< 3                                                                                | \< 3                    |

Metrics are **materialized**. You get `404` with `metrics_not_available` until scoring has finished.