Yes. Siftsy calculates sentiment scores along with its Siftsy Score so you can view and export sentiment analysis and the % positive / neutral / negative breakdown. In Version 3 you also get Content Score and Topic Score as separate breakouts.
Understanding Siftsy Score - Methodology
This page is the full Siftsy Score and sentiment scoring methodology, version 2.5.2 (October 2026). You can also download the PDF version, or view it right here:
How Siftsy analyzes comments
Comment sections are reactions to content, not standalone text, so every comment is analyzed within its context:
Social post data: description, transcript (when available), and additional metadata
Comment section data: comments are analyzed alongside the rest of the comment section at the same time
Additional context data: data from search partners (like Perplexity) for up-to-date relevant information
The scoring system is model-agnostic and built for continuous evolution. It works with any modern AI model (OpenAI, Anthropic, Google, Meta, and others), applies Siftsy's scoring parameters and Siftsy's updates to them, and pulls in additional context when needed.
What Siftsy Score measures
Siftsy Score is a 1–10 temperature check of a comment section in response to content. It is not a like count, not a virality metric, and not "how people feel about the post" alone.
Comments are rated on three signals (Sentiment, Relevance, and Consensus) that are blended into one score, so a loud, off-topic, or fighting thread cannot look healthy just because one axis is high. Each version blends the same signals with different weights:
Version 1 (Equal Average): all three signals count equally, prioritizing holistic engagement
Version 2 (Sentiment-biased): Sentiment counts double (2 : 1 : 1), so the mood of the section moves the score most
Version 3 (Split sentiment, recommended): the Version 2 blend with Sentiment split evenly into Content and Topic, plus reply hierarchy, a negativity cap, and gentler engagement weighting
Siftsy aims to rank and score comments within a Siftsy Score accuracy of ±1 point of how a human would rank them using the scale definitions on this page.
The three signals
Sentiment (1–10)
How commenters feel about the post, creator, or topic:
Low scores: intentionally negative, bitter, angry, mean, or insulting tones
Moderate scores: a neutral, matter-of-fact, observational tone
High scores: praise, admiration, or good humor in tone
In Version 3, Sentiment is a weighted blend of Content and Topic. Versions 1 and 2 use the direct comment-sentiment average. How each comment is scored is covered under How Sentiment Is Scored below.
Relevance (1–10)
How closely comments relate to the original post content:
Low scores: completely off-topic to the post and the conversation about the post
Moderate scores: general commentary not explicitly about the post, but applicable
High scores: directly responding to or paraphrasing something about the post
Consensus (1–10)
How commenters interact with each other, not whether they agree with the post:
Low scores: explicitly arguing with other commenters
Moderate scores: sharing mixed opinions with other commenters
High scores: in complete agreement and sharing mutual support
A thread can be uniformly negative toward the post and still have high Consensus.
What the Scores Mean
Read the score in two passes. First, the temperature check: a fast, standardized read on the overall vibe of a comment section reacting to content. Use it to rank posts, compare them against each other, and spot the ones that deserve a closer look using the bands below. Then, the substance: the number tells you that something is happening, not what. The breakouts show where the temperature is coming from, and Search and Insights let you read the conversations behind the number before drawing conclusions.
Versions 1 and 2
Range | Label | Meaning |
|---|---|---|
7.0+ | Immaculate | The comments are positive and supportive. |
5.0 – 6.9 | Good | Most of the conversation is in good faith. |
3.0 – 4.9 | Sub-Par | Something here isn't landing. |
< 3 | Off | Negativity and infighting run the comments. |
Version 3
Range | Label | Meaning |
|---|---|---|
7.0+ | Immaculate | The comments are positive and supportive. |
4.5 – 6.9 | Mixed | The comments are split and attitudes are mixed. |
< 4.5 | Off | The comments are substantially negative and critical. |
Reading the score in context
The same number means different things depending on the job:
Campaign performance: track scores across posts, creators, and flights to see which activations landed. A rising trend across a campaign says more than any single post's number
Creative reviews: a high Content score with a low Topic score, or the reverse, separates reaction to the execution from reaction to the subject and tells you whether to revisit the creative or the message
Consumer research: mid-range scores are often the most informative. High Relevance with mixed Sentiment usually marks a section full of unfiltered opinions about the product itself
Brand reputation: watch the trend, not the level. An account that normally scores steady and drops two points is a louder signal than one low-scoring post
Press and communications: a low score with high Consensus means the audience has settled on a shared narrative; a low score with low Consensus means the conversation is still contested. The two call for different responses
Content Score and Topic Score
Traditional sentiment analysis collapses "I hate this video" and "I hate this subject" into one scale. With Content and Topic sentiment differentiation, Siftsy measures sentiment toward what exactly the audience is reacting to. Content and Topic Scores are produced only when Version 3 is selected.
Content Score (1–10)
How commenters feel about this post and its creator.
Range | Label | Meaning |
|---|---|---|
7+ | Supportive | Strongly positive toward the post or creator |
5.6 – 6.9 | Favorable | Reacting in good faith toward the post or creator |
4.5 – 5.5 | Neutral | Not reacting directly to the post or creator |
4.5 – 5.5 | Divided | Reactions to the post or creator are polarized. At least 25% of Content readings lean positive and 25% lean negative. |
3.0 – 4.4 | Critical | Raising critiques of the post or creator |
< 3 | Hostile | Aggressively against the post or creator |
Topic Score (1–10)
How commenters feel about the topics being discussed.
Range | Label | Meaning |
|---|---|---|
7+ | Positive | Strongly positive toward the topics |
5.6 – 6.9 | Positive Lean | Mostly positive toward the topics |
4.5 – 5.5 | Neutral | Not reacting directly to the topics |
4.5 – 5.5 | Divided | Reactions to the topics are polarized. At least 25% of Topic readings lean positive and 25% lean negative. |
3.0 – 4.4 | Negative Lean | Mostly negative toward the topics |
< 3 | Negative | Aggressively against the topics in the post |
Overall Sentiment Score
How commenters feel about the post and its topics.
Range | Label | Meaning |
|---|---|---|
7+ | Positive | Strongly positive about the post overall |
5.6 – 6.9 | Positive Lean | Mostly positive about the post overall |
4.5 – 5.5 | Neutral | Reactions are generally neutral. The average Content and Topic scores are within 2 points of each other. |
4.5 – 5.5 | Divided | Content and Topic scores diverge by 2+ points, or at least 25% of comments lean positive and 25% lean negative. Sentiment sits in the middle because of a polarized response. |
3.0 – 4.4 | Negative Lean | Mostly negative about the post overall |
< 3 | Negative | Strongly negative about the post overall |
How Neutral vs Divided is decided. The tests apply only when a score sits in the mid band (4.5–5.5). Overall Sentiment is Divided if either test fires; otherwise the label stays Neutral. Content and Topic use the split test on their own readings:
Content/Topic divergence (overall Sentiment only): |Content − Topic| ≥ 2. With a mid-band average, a 2-point gap puts one dimension at 5.5 or higher and the other at 4.5 or lower, a positive read on one and a negative read on the other
Split distribution: the smaller side holds at least 25% of readings. The positive side is readings of 7+ plus 5.6–6.9; the negative side is readings below 3 plus 3.0–4.4. Both camps must each hold at least a quarter of readings, so the neutral bucket is at most 50%
Overall Sentiment pools Content and Topic, with each comment counted once per dimension. In Version 3, headline Sentiment is the average of Content and Topic: Sentiment = (Content + Topic) / 2.
How Sentiment Is Scored
Sentiment is the mood signal inside the Siftsy Score, and it is also reported on its own: a 1–10 score per comment, rolled up into a section average and a positive / neutral / negative breakdown. Each comment is scored by context-aware AI applying Siftsy's scoring parameters, not by keyword lists, and never on the comment text alone. This section describes how that scoring behaves; the complete parameters are proprietary and continuously refined.
How a comment gets scored
The post is analyzed first. Before any comment is read, the model builds a picture of the post itself from the caption, transcript, title, account bio, and retrieved outside context: who the creator is, the editorial stance they are taking, the messages they want the audience to walk away with, and the topics commenters are likely to react to.
Comments are scored against that picture, together. Comments are evaluated in the context of the post analysis and alongside the rest of the comment section, so replies, running jokes, and pile-ons are read the way a human moderator would read them: with the thread in view, not one comment at a time.
Each comment gets two sentiment readings. The model scores the comment's attitude toward the post and creator and, separately, toward the topics being discussed. Each reading uses the 1–10 scale and is paired with a confidence level. Version 3 surfaces these as Content Score and Topic Score; Versions 1 and 2 blend them into a single Sentiment score.
What is considered
Input | How it is used |
|---|---|
The post | Caption, transcript, title, who posted it, and the creator's stance: what they are arguing, celebrating, or announcing. The same comment can score differently under a different post. |
The conversation | Comments are scored together, not one-by-one in a vacuum. A reply is resolved through the parent comment it answers, so "exactly" and "nah" carry the meaning of the exchange they sit in. |
Language as people use it | Slang, emoji, sarcasm, and idiom are scored for what they mean in use. Non-English comments are read in their own language and cultural register, not through a literal translation. |
Outside context when needed | Recent events, controversies, and background retrieved from search partners. This is the context that makes a reaction legible when the post alone doesn't explain it. |
Reading the 1–10 scale
The scale is anchored to intent, not word choice. Per-comment scores are continuous (a comment can score 7.2), and the same bands apply to both sentiment readings:
Range | Read | What lands here |
|---|---|---|
7+ | Positive | Praise, excitement, buying intent, admiration, delight |
5.6 – 6.9 | Positive Lean | Interest, genuine questions, friendly humor, good-faith engagement |
4.5 – 5.5 | Neutral | Matter-of-fact, informational, or not addressing this at all |
3.0 – 4.4 | Negative Lean | Skepticism, critique, disappointment, unfavorable comparisons |
< 3 | Negative | Hostility, insults, mockery, protest |
Neutral means "no read", not "mixed feelings." A comment that never addresses the creator or the topic sits at 5 rather than pulling the average in either direction. Genuinely mixed comments ("love the concept, not the execution") are weighted toward their dominant take instead of defaulting to the middle. At the section level, a 4.5–5.5 average can still be Divided when the readings are polarized. These are the same bands behind the Content and Topic Score labels and the positive / neutral / negative breakdown percentages.
How intent is read
Alignment with the post's stance, not word polarity. On a post criticizing something, "this is terrible, I hate it" agrees with the creator, so it counts as positive engagement toward the creator even though the words are negative. Disagreeing with praise works the same way in reverse
Intended meaning over literal words. Sarcasm, mock praise, and rhetorical questions are scored for what the commenter meant. Playful teasing and laughing-with are distinguished from mockery and laughing-at
Replies inherit their thread. A short reply agreeing with a critical comment carries that criticism; a reply pushing back against it flips the stance. Commenter-versus-commenter fights lower the section's Consensus rather than distorting sentiment
Participation counts as engagement. Answering a call-to-action, tagging friends, or dropping a hype reaction is treated as real positive engagement rather than discarded as contentless
Redirected praise is read as deflection. Celebrating a competitor, predecessor, or alternative on a post about something else is not support for that post, however positive the words sound
Story and real world are scored separately. Hating a fictional villain is engagement with the work; criticizing a real actor, brand, or production choice is scored literally as criticism
Images and GIFs
The scorer reads text only; it never sees an image or GIF itself. Media in a comment is turned into text in three steps, then scored exactly like a written comment, alongside the comment's own words and its thread:
The image or GIF attached to the comment is picked up
It is captioned as it appears, with no post, comment, or thread context
Perplexity searches the caption for a known meme, character, or origin
The caption plus any meme context is scored with the thread, like a written comment
The caption comes first without context so a reaction GIF is described for what it shows, for example "a man slowly clapping." Only after the meme search does the scorer work out from the thread whether that applause is sincere or sarcastic, just as it does with a normal text comment.
In one worked example, an Instagram post about a cat that looks almost perfectly round drew many GIF replies. A GIF-only comment captioned as "a shocked man reacts to a broken plate" was identified through meme search as a well-known reaction meme and scored 7.0 (Positive) read against a delighted thread. A second GIF-only comment captioned as "a fluffy cloud floating against a muted gray background" found no known meme and scored 7.5 (Positive), reading as a compliment next to a round, cloud-like cat.
Confidence and consistency
Every score carries a 1–10 confidence level. Ambiguous comments (subtle irony, rare dialects, garbled text) lower confidence rather than forcing a certain read, and spam is excluded entirely. Per-comment scores then average into the section-level Sentiment score, with each comment counting once in the positive / neutral / negative breakdown.
Why use Siftsy Score
Sentiment analysis tells you if comments are positive, neutral, or negative in tone and intent. Siftsy Score tells you why, how it relates, and what it means for your content, and Version 3 adds what people are reacting to: the post and creator (Content) versus the subjects being discussed (Topic). Siftsy offers sentiment-only analysis for compatibility with existing workflows, but the Siftsy Score reveals patterns invisible to sentiment-only analysis, tracks audience reception trends faster, standardizes comparison across posts, and lets teams switch score versions instantly without losing history.
In practice that difference shows up in real scenarios: marketing teams see not just whether people liked an ad but exactly how they received it, product launches distinguish artificial praise from real feedback and genuine concerns, content strategy identifies which posts truly resonated rather than which simply avoided negative reactions, community teams recognize authentic engagement versus superficial reactions, and crisis response shows whether people hate the video or hate the topic, and whether commenters are aligned or fighting.
What is not in the score
Like count, view count, or follower count as a direct input. Likes and replies only matter if engagement weighting is on, and even then they only scale comment weights
"Agrees with the post" as Consensus
Hidden-from-analysis comments
Custom sentiment rules as a separate headline formula. They change per-comment scores before this math
Comparing Siftsy Score versions
Version 3 is an evolution, not a replacement. It uses the same scale, signals, and weights as V2, and adds a Content/Topic split and guardrails so reply pile-ons, hostile minorities, and viral outliers can't write the headline.
Switching versions is a re-calculation, not a re-run. Select V3 as the team score version and every score updates instantly, including posts already analyzed. You can switch back to V2 or V1 at any time. Nothing is lost, because all versions are built from the same per-comment AI scores.
Version 1 | Version 2 | Version 3 | |
|---|---|---|---|
Version label | Equal Average | Sentiment-biased | Split sentiment-biased |
Prioritizes | Holistic reactions | Reaction sentiment | Targeted reaction sentiment |
Scale | 1–10 | 1–10 | 1–10 |
Signals | Sentiment, Relevance, Consensus | Same as V1 | Same as V1 |
Signal weights | 33.3% / 33.3% / 33.3% | 66% / 16.5% / 16.5% | Same as V2 (66 / 16.5 / 16.5) |
Sentiment | One 1–10 reading per comment | Same as V1 | Content + Topic, averaged 50/50 into Sentiment |
Per-comment scores | Context-aware AI; versions are rollups only | Same as V1 | Same as V1 |
Score bands | Immaculate 7.0+, Good 5.0–6.9, Sub-Par 3.0–4.9, Off < 3 | Same as V1 | Immaculate 7.0+, Mixed 4.5–6.9, Off < 4.5 |
Top-level vs reply | Every comment counts as 1 | Same as V1 | Top-level = 1. Combined replies = 1 |
Negativity cap | None | None | Yes |
Engagement weighting | Optional | Optional | Default |
What V3 adds to V2
Addition | What it adds |
|---|---|
Content vs Topic | A split read: reaction to the post and creator versus the subjects discussed |
Mixed band | A first-class label for mid-range scores as mixed reactions, not mild failures |
Negativity cap | A ceiling that gives a hostile minority visibility even when the average is high |
Engagement weighting | Now the default, with a 1× floor so quieter comments still count, and likes lifting Sentiment only |
Top-level vs reply | Full weight for top-level comments; replies share a capped pool so the score tracks breadth |
Real-world validation
These cases show the Siftsy Score on real comment sections. Scores in these case studies were produced with Version 1 (Equal Average) for demonstrative purposes.
Post | Platform | Comments | Siftsy Score | Reception |
|---|---|---|---|---|
Comedian @veronika_iscool's comedy music video featuring Vitamin Water and NYC summer heat | TikTok | 758 | 8.6 | High praise across almost all commenters for the humor, the catchy tune, and the charming product integration |
Tech creator Ben Kaluza demonstrates the Shade app, a video file app for creators | TikTok | 128 | 6.4 | Generally positive engagement and interest, with questions and skepticism about data privacy and local AI functionality bringing sentiment down |
Entrepreneur Sara Blakely introduces Sneex, a luxury high-heel sneaker | TikTok | 613 | 4.4 | Excitement about the concept, but many concerns about the design and price |
A TV brand promotes itself with a post about an international sporting event | 4 | 1.7 | Significant dissatisfaction with the brand's products; the sporting event is not referenced |
Frequently asked questions
Does Siftsy report on sentiment?
What is the difference between Content Score and Topic Score?
Content is how commenters feel about this post and its creator. Topic is how they feel about the subjects being discussed. Both are produced only in Version 3, where headline Sentiment is the average of the two.
Does "high Consensus" mean people agree with my post?
No. Consensus measures how commenters react to other commenters. A thread that is uniformly negative toward the post can still have high Consensus. Low Consensus means arguing, fighting, or dunking on other people in the section.
How do likes and replies affect the score?
Only if the team setting is Weighted. Equal mode gives every comment the same weight, plus the Version 3 reply hierarchy. Weighted mode uses the formulas in Appendix C. Distribution bars never use likes; they always count each comment as 1.
Does Siftsy understand sarcasm?
Yes, in both positive and negative connotations. For example, "Following for the JUST KIDDING video that surely has to be coming next" on the Sneex launch video was identified as layered sarcasm and scored 3.8 (Negative Lean).
Can I use Siftsy for positive/negative reactions only?
Yes. Siftsy search and filter lets you view and interpret comments for the sentiment you are researching.
Does Siftsy understand the tone of emojis?
Yes, whether they are standalone emoji comments or emoji added to text for emphasis. For example, "man this is a godsend. apple's search is so crap..." on the Shade app demo was scored 7.2 (Very Positive) with its frustrated emojis read in the context of an overall positive message.
Does Siftsy understand modern internet slang?
In most cases yes, and recent references not yet seen by the AI are an area of ongoing improvement. For example, "Okay but why does this EAT tho" on the Vitamin Water video was understood as a compliment and scored 9.1 (Very Positive).
Supported languages
Tier | Languages | What to expect |
|---|---|---|
Best support | English, Spanish, French, German | High accuracy in sentiment analysis and context understanding, strong handling of nuance, sarcasm, and idioms, and strong support for brand, product, or service mentions and reactions |
Good support | Chinese (Simplified), Dutch, Italian, Russian, Portuguese, Japanese | Generally reliable sentiment analysis that can miss very subtle nuances, good handling of casual language, and adequate support for brand, product, or service mentions |
Medium support | Korean, Norwegian, Arabic, Danish, Turkish, Thai, Hindi, Bulgarian, Vietnamese, Ukrainian, Swedish, Serbian, Czech, Croatian, Finnish, Slovak, Hungarian, Lithuanian, Romanian, Latvian | Moderate performance with potential misinterpretation of tone, possible struggles with slang and regional dialects, but good support for brand, product, or service mentions |
Low support | Swahili, Malayalam, Hmong, Hebrew, Tamil, Amharic, Persian (Farsi), Telugu, Mongolian, Indonesian, Kannada, Kazakh, Malay, Bengali, Uzbek, Filipino (Tagalog), Punjabi, Tatar, Slovenian, Urdu, Azerbaijani, Icelandic, Sinhala, Somali, Basque, Pashto, Quechua, Catalan, Kurdish, Guarani, Macedonian, Armenian, Maori, Estonian, Georgian, Galician, Tigrinya | High likelihood of errors in understanding sentiment and context, with frequent misinterpretations of informal language or cultural references |
Appendix A · How the score is calculated
Every scored comment already has Sentiment, Relevance, and Consensus on a 1–10 scale (and in V3, Content and Topic as well). Hidden-from-analysis comments are skipped. The remaining comments are averaged, then blended:
Siftsy Score = (Sentiment × wS) + (Relevance × wR) + (Consensus × wC)Version | Sentiment | Relevance | Consensus | Label |
|---|---|---|---|---|
1.0 Equal Average | 33.3% | 33.3% | 33.3% | 1 : 1 : 1 |
2.0 Sentiment-biased | 66% | 16.5% | 16.5% | 2 : 1 : 1 |
3.0 Split sentiment | 66% (33% Content + 33% Topic) | 16.5% | 16.5% | 2 : 1 : 1 |
Display values are floored to one decimal (5.94 becomes 5.9). Version 3 may then limit the result to the negativity-cap maximum for that very-negative share (Appendix D).
Worked example
Sentiment 6.0, Relevance 7.0, Consensus 5.0. In Version 3, Sentiment splits into Content 8.0 and Topic 4.0, which average back to 6.0.
Version | Blend | Cap | Displayed |
|---|---|---|---|
1.0 | (6 + 7 + 5) / 3 = 6.00 | none | 6.0 |
2.0 | 0.66×6 + 0.165×7 + 0.165×5 = 5.94 | none | 5.9 |
3.0 | same 5.94 | 8.5 at 12% very-negative | 5.9 |
The cap only binds when the blend is above the maximum. 5.94 is already under 8.5, so it still displays as 5.9. A 9.2 blend with the same 12% very-negative share would display as 8.5.
Appendix B · How comments roll up
Per-comment scores: Sentiment, Relevance, Consensus, and (in V3) Content + Topic
Hide rules: comments hidden from analysis are skipped
Structural weight (h): in V1 and V2 every comment = 1; in V3, top-level = 1 and replies share a pool (Appendix D)
Engagement (optional): if Weighted, likes and replies scale the mean
Axis means: weighted average per signal
Headline blend: apply the version weights. In V3, Sentiment is 50/50 Content + Topic first
Version 3 cap: limit the score to the maximum for the very-negative share (Appendix D)
Display: floor to one decimal
Breakdown percentages are never engagement-weighted. Each comment counts as 1 in the distribution bars.
Appendix C · Engagement weighting
If Weighted mode is on, comments with more likes and replies count for more. Version 3 uses a gentle linear-log lift so quieter comments still have weight and the most-engaged comments are lifted, not used as the only weight. Versions 1 and 2 used a steep power curve where the top comment's multiplier is 1× and everyone else is a fraction of that.
Strength w is a team setting clamped to 0–1. With e = max(likes + replies, 1) and eMax the highest e in the comment set, the normalized log ratio is:
normalized = ln(1 + e) / ln(1 + eMax)Version 3 multiplier = 1 + 4w × normalized (range 1× to (1 + 4w)×)
Version 1/2 multiplier = normalized^10wThe range is 1× to (1 + 4w)×. At w = 1 the top comment is 5×; at w = 0.5 the top comment is 3×. If w = 0, the multiplier is skipped and weight = h. A comment with zero likes and zero replies still has e = 1.
Which axes get the multiplier
Axis | Version 1 / 2 | Version 3 |
|---|---|---|
Sentiment / Content / Topic | Yes, when w > 0 | Yes, when w > 0 |
Relevance | Yes, when w > 0 | No, structural weight only |
Consensus | Yes, when w > 0 | No, structural weight only |
Bucket % bars | Never | Never |
Multipliers when eMax = 1,000 and w = 1
e | V1 / V2 | V1 / V2 vs top | V3 | V3 vs top |
|---|---|---|---|---|
1 (floor) | ~10⁻¹⁰× | ~0% | 1.40× | 28% |
10 | 0.00002× | ~0% | 2.39× | 48% |
50 | 0.004× | 0.4% | 3.28× | 66% |
100 | 0.018× | 1.8% | 3.67× | 73% |
250 | 0.11× | 11% | 4.19× | 84% |
500 | 0.35× | 35% | 4.60× | 92% |
1,000 (max) | 1.00× | 100% | 5.00× | 100% |
Appendix D · Version 3 extras
Version 3 adds a reply hierarchy and a negativity cap to keep the score reflective of the overall temperature of comment sections across the most normal and edge cases. Reply hierarchy keeps a long reply chain from outweighing the parent thread. The negativity cap adjusts the maximum score of the scale when a sizeable percentage of comment sentiment is negative.
Reply hierarchy
Top-level comment weight = 1.0
Direct replies share a pool: 0.5 total if there are 1–3 replies, 1.0 total if there are 4 or more
Each reply gets pool / replyCount
If engagement weighting is also on: weight = h × multiplier
Negativity cap
If enough comments score below 3, the maximum the score can reach is lowered, and the displayed score cannot exceed that maximum:
Siftsy Score = min(raw, max)Very-negative share | Maximum |
|---|---|
< 2% | 10.0 (no cap) |
2%+ | 9.5 |
5%+ | 9.0 |
10%+ | 8.5 |
20%+ | 8.0 |
The overall score uses the pooled very-negative share of Sentiment readings, with each comment counted once for Content and once for Topic (readings below 3). Content uses opposed (Content below 3) and Topic uses hostile (Topic below 3). Pooling keeps the overall cap consistent with the dimension caps: 6% opposed and 0% hostile reads as 3% very negative overall.
Related docs
Was this helpful?
Was this helpful?