brand-listening

Imported from brightdata/skills (skills/brand-listening/SKILL.md). Install upstream with npx skills add brightdata/skills --skill brand-listening. Copyright stays with the author.

Brand Listening

Find out what people are actually saying about a brand across social platforms, news, and reviews — powered by live web data, not stale training knowledge. Combines the Bright Data CLI (bdata) for collection with a sentiment + theme analysis layer to deliver a cited, actionable digest.

Never answer brand-sentiment questions from training knowledge alone. Public sentiment changes daily. Always collect live mentions first, then classify and synthesize.

Prerequisites

1. Bright Data CLI installed:

   curl -fsSL https://cli.brightdata.com/install.sh | bash

2. One-time login completed:

   bdata login    # or: bdata login --device  (SSH / headless)

Verify before collecting:

if ! command -v bdata >/dev/null 2>&1; then
    echo "bdata CLI not installed — see skills/bright-data-best-practices/references/cli-setup.md"
elif ! bdata zones >/dev/null 2>&1; then
    echo "bdata not authenticated — run: bdata login"
fi

Halt and route to setup if either check fails.

Core Workflow

1. Clarify scope — Which brand/product/person? Which platforms? What time window (default: last 30 days)? What does the user want to do with it (general health check, launch monitoring, complaint triage, advocacy hunting)?
2. Discover, then collect — Use bdata search to find where the brand is being discussed, then bdata pipelines to pull structured mentions from each platform. Parallelize independent calls.
3. Normalize — Collapse every raw result into the single mention schema in references/sentiment-and-output.md before any analysis.
4. Classify & cluster — Assign sentiment per mention (with a reason), then group mentions into themes. Follow the sentiment guardrails — never inflate either side.
5. Deliver — Produce the cited digest (Output A), optionally with the structured dataset (Output B). Every report ends with a "So what — recommendations" section.

Data Collection Rules

• Discovery first. You rarely have the right URLs up front. Run bdata search "<brand> site:reddit.com" --json (and per-platform variants) to find threads, profiles, and articles, then feed those URLs to pipelines.
• Prefer bdata pipelines over bdata scrape whenever a pre-built scraper exists for the platform. Pipelines return clean structured JSON (author, date, engagement, text).
• Look up scrapers, never guess them. For each platform: bdata pipelines search "<platform> <data type>", then bdata pipelines info <dataset_id> <method> to see the input fields, then bdata pipelines run. Never guess a dataset_id or field names. Full workflow: the brightdata-cli skill, pipelines reference.
• Always pass --json when you need to parse or pipe output.
• Be cost-efficient — a standard sweep is ~6–12 bdata calls, not 50. Pull the highest-signal threads/profiles, not everything.
• Parallelize independent calls across multiple Bash tool calls in one response.
• Every mention needs a source URL. No unattributed quotes, ever.
• Never fabricate sentiment or fill gaps. If a platform returns nothing, report it in "Gaps & caveats".

Platform Modules

Pick the platforms that fit the brand. Consumer/cultural brands skew TikTok/Instagram/Reddit; B2B/SaaS skews Reddit/X/review sites; local businesses skew Google Maps reviews.

Each run below uses the dataset_id you picked from the search line above it. Check bdata pipelines info <dataset_id> <method> first: extra inputs (comment count, days back) are --<field> <value> flags named there.

Reddit — honest, unfiltered sentiment

# Discover relevant threads
bdata search "<brand> site:reddit.com" --json
bdata search "<brand> review reddit" --json

# Pull structured post + comment data from the threads found
bdata pipelines search "reddit posts"
bdata pipelines run <dataset_id> collect_by_url "<reddit-thread-url>" -o reddit.json

Reddit is the single best source for candid opinions brand channels hide. Prioritize it.

X / Twitter — real-time reaction

bdata search "<brand>" --json                      # find recent discussion
bdata pipelines search "x posts"
bdata pipelines run <dataset_id> collect_by_url "<x-post-url>" -o x.json

Instagram — brand aesthetics, comments, advocacy

bdata pipelines search "instagram posts"         # profile posts; pick a method with info
bdata pipelines search "instagram comments"
bdata pipelines run <dataset_id> collect_by_url "<instagram-post-url>" -o ig_comments.json

TikTok — cultural relevance, viral sentiment

bdata pipelines search "tiktok posts"            # profile posts; pick a method with info
bdata pipelines search "tiktok comments"
bdata pipelines run <dataset_id> collect_by_url "<tiktok-video-url>" -o tt_comments.json

YouTube — reviews, tutorials, long-form opinion (comments are gold)

bdata search "<brand> review youtube" --json
bdata pipelines search "youtube videos"
bdata pipelines search "youtube comments"
bdata pipelines run <dataset_id> collect_by_url "<video-url>" -o yt_comments.json

Reviews — structured customer sentiment

# App-based products, local businesses, Facebook pages: find the review scraper
bdata pipelines search "google play"             # or "apple app store"
bdata pipelines search "google maps reviews"
bdata pipelines search "facebook reviews"
bdata pipelines run <dataset_id> collect_by_url "<listing-or-page-url>" -o reviews.json

# SaaS / software — discover then scrape (no pipeline)
bdata search "<brand> site:g2.com" --json
bdata search "<brand> site:capterra.com" --json
bdata scrape "<g2-or-capterra-url>"

News & press — coverage and tone

bdata search "<brand>" --json                  # general SERP, scan for news
bdata scrape "<article-url>"                    # pull full article text for tone

Never hardcode a scraper. The catalog changes, and two scrapers can share a name. Always pick the dataset_id from bdata pipelines search and the fields from bdata pipelines info in the same session.

Choosing What to Collect

Sentiment, Themes & Output

Read references/sentiment-and-output.md for:

• The normalized mention record schema (use it as the row shape for both analysis and the dataset output).
• The sentiment method + guardrails (classify from text as written; sarcasm/ambiguous → mixed; always report counts and percentages with the denominator).
• Theme clustering — where the actionable insight lives.
• Output templates — A (digest report), B (structured dataset), C (both).

Output Quality Standards

1. Every mention has a source URL — no unattributed claims.
2. Facts separate from interpretation — the quote is the fact; your sentiment label is interpretation. Keep them distinct.
3. Percentages always carry N — "61% positive of 142 classified mentions", never a bare percentage.
4. Be honest about gaps — list platforms that returned nothing this window. Note small samples that can't generalize.
5. Date-stamp everything — "Data collected on · window: last days".
6. End with "So what" — every report closes with 2–4 actionable recommendations, each grounded in a theme from the data. Raw mentions without interpretation are not a deliverable.