manifoldmcp/marketing-skills / plugins/manifold-ai-search/skills/brand-monitoring
brand-monitoring
VerifiedWhen the user wants brand monitoring, brand mentions on Reddit, social and the web, mention alerts, social listening, share of voice against competitors, or a brand awareness check. Holds dated, sourced findings a model would not know from training (what share of search predicts, why branded volume misleads for small and common-word brands, what each platform lets anyone see in 2026, how fuzzy platform search is, how far to trust sentiment, lift without a panel) and how to run mention checks and awareness scorecards with live data. Also use when the user mentions alert me when someone mentions us, monitor Reddit for our brand, brand monitoring, mention alerts, track mentions every day, who talked about us this week, a brand awareness plan, nobody knows who we are, share of voice against competitors, share of search, grow branded search, brand lift, get our name out there, or getting the brand known. AI answers alone go to ai-search, rivals' moves to competitor-intel, client reporting to agency.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 12.3 KB |
| references/awareness.md | markdown | 6.0 KB |
| references/brand-lift.md | markdown | 3.9 KB |
| references/coverage.md | markdown | 4.9 KB |
| references/field-notes.md | markdown | 4.3 KB |
| references/monitoring.md | markdown | 10.4 KB |
| references/sentiment.md | markdown | 3.3 KB |
| references/share-of-search.md | markdown | 6.3 KB |
references/sentiment.md
View on GitHub# Sentiment and what a mention needs
How far to trust "positive, negative, neutral", and the classification that serves a brand better. Checked October 2026.
## What the evidence says
- **Models are fine on plain polarity and weaker on everything else.** Across 13 tasks and 26 datasets, large language models performed well on simple sentiment classification but lagged on tasks needing deeper understanding or structured sentiment, such as which aspect of a product the opinion is about (Zhang et al., "Sentiment Analysis in the Era of Large Language Models", May 2023). Brand mentions are mostly the hard kind: "switched from Acme, the API was fine but support never answered" is one mention with two opinions.
- **Sarcasm is still hard.** SarcasmBench tested eleven models and found sarcasm understanding well behind sentiment (2024); on a seven-way sarcasm-type task the best macro F1 was 0.37 (Sarc7, Gemini 2.5, 2025). Social posts about brands are often sarcastic.
- **Lexicon tools are worse on social text.** Word-score methods (VADER and the like) are inspectable but blind to context, slang and mixed languages.
- **Nobody has published a benchmark on brand mentions.** Vendor sentiment accuracy claims come without a labelled test set. Treat a dashboard's sentiment share as a rough signal, not a measurement.
## Classify by what the mention asks for
A sentiment score does not say what to do. A kind does:
| Kind | Example | Action |
|---|---|---|
| Complaint | "Acme's export broke again" | Pass to support; reply if public and the community allows |
| Question | "Does Acme work with Shopify?" | Answer, or note the gap in the docs |
| Comparison | "Acme vs Beta for a small team?" | Watch; answer only with an honest comparison |
| Recommendation | "We use Acme, it's great for X" | Thank; save as proof, with permission |
| Praise | "Acme's new dashboard is lovely" | Thank; save as a testimonial, with permission |
| Churn or switch | "We moved off Acme to Beta" | Read in full; it is the most useful mention there is |
| News or launch | A post about the brand's own announcement | Count separately; it measures reach, not opinion |
- **Mixed is a valid answer.** A switch story that praises one feature and leaves over another is both; record both aspects rather than forcing one label.
- **Read the thread, not the line.** A comment's meaning often depends on what it answers. Reddit comment rows carry `post_title` and `parent_id`; read those before labelling.
- **Leave out what is not organic.** The brand's own accounts, partner posts (#AcmePartner), paid posts (`is_ad` true) and seeded "review" threads (see [monitoring](monitoring.md#reading-the-rows)) are not public opinion.
- **Report counts by kind with examples, not a percentage positive.** "Four complaints this week, three about exports, linked" is actionable; "62% positive" is not, and on a sample of twenty it moves 20 points with four posts.
## When a model labels the mentions
- Give it the thread context and the kinds above, ask for the kind, the aspect and a one-line reason, and spot-check twenty labels by hand before reporting any share. Disagreement on more than a few means the labels need a human pass.
- Keep the label with the row, so a later run can be compared on the same scheme. A changed scheme starts a new baseline.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack." | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read. | Stars: 196k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report. | Stars: 196k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt. | Stars: 196k | Updated | Claude CodeCodex | Verified | Install |