placement-relevance

name

placement-relevance

description

Classify each Google Ads placement — the website, app, or YouTube video the ad actually showed on — as on-topic vs off-topic for the advertiser and brand-safe vs unsafe, by reading the placement name / URL and scraping the page when it's ambiguous, then emit the exclusion list and the % of content spend wasted on irrelevant or unsafe content. Use whenever the user asks where their ads are showing, about placements, brand safety, content exclusions, "what sites / videos am I on", wasted Display / Demand-Gen spend, or wants to read the audience from the content it consumes — even if they don't say "placement". Display / Video / Demand-Gen only (Search and Performance Max do not expose placements).

Placement Relevance & Brand-Safety (content classification)

Goal (job-to-be-done)

Per placement (the website / mobile app / YouTube video the ad actually showed on), judge whether the content is on-topic for what the advertiser sells and brand-safe — by reading the placement name / URL, and scraping the page when the name is ambiguous and the spend is material — then produce the exclusion list and the % of content spend on off-topic / unsafe placements. Placement sub-member of audience/: where / what the audience consumes — the content the ad ran next to reveals who is watching.

• Who: the media buyer / PPC manager running Display / Video / Demand-Gen. When: a content campaign is live and you need to know what your money is showing next to.
• Decision it drives: a placement exclusion list (cut the irrelevant / unsafe), and a read of the realized audience from the content it consumes.
• The differentiator: the verdict comes from reading the content, not from a metric — only AI turns 10,000 raw placement strings into "your golf brand is running next to true-crime videos".

Scope

• ✅ Relevance (on / off-topic) + brand-safety classification of automatic placements; exclusion candidates; % of content spend off-topic / unsafe; the realized-audience signal from content.
• ✅ Availability-gated — Display / Video / Demand-Gen only; an explicit n/a on Search / PMax.
• ❌ Not placement performance / movement attribution — counts feed ranking by spend, not a contribution-to-change verdict (that engine is the count-based siblings).
• ❌ Not managed-placement targeting setup (which placements you chose to target).
• ❌ Search / PMax — no detail_placement_view rows (Performance Max hides placements).

Components (read these references as needed)

• Tools / data plan: references/tools.md — the applicability fingerprint, the placement pull, and the conditional scrape.
• Framework / rubric: references/framework.md — the brain: the availability gate, the spend threshold, the relevance + brand-safety rubric, scrape-when-ambiguous, confidence.
• Output schema: references/output.md — the JSON this skill emits.

Operate

Input: per placement, the display_name / target_url / placement_type + the base counts (impressions, clicks, cost, conversions), pulled for the content campaigns; plus the business context (what the advertiser sells) that the relevance judgment is measured against.

Process (apply the rubric in references/framework.md):

1. Mandatory pre-check — availability + applicability. Fingerprint campaign_advertising_channel_type; if no DISPLAY / VIDEO / DEMAND_GEN, emit verdict: "n/a" (Search / PMax don't expose placements) and stop.
2. Deterministic first. Group by placement, sum spend; drop the long tail below the spend / impression threshold (don't classify 1-impression noise); rank by spend.
3. Then judge — read the content. Classify each above-threshold placement on two axes: relevance (on / off-topic vs the business) and brand-safety (safe / sensitive / unsafe). Judge from display_name when it's clearly enough; when it's ambiguous and the spend is material, scrape the page / channel and re-judge. Attach a confidence; send the genuinely ambiguous to human_review.
4. Roll up. % of content spend off-topic + unsafe; the exclusion candidates ordered by spend; the one-line realized-audience read.

Emit the JSON in references/output.md: synthesis (the % + the action), placements[] (per-placement verdict + confidence + reason + spend), and rollup (% off-topic, the ordered exclusion list). Pure data — no emojis, tables, markdown, or colors in the output.

Voice (link, don't copy): write every narrative line per _framework/writing.md — the heading is a question the data answers; the first sentence answers it then names the driver; plain language for a non-technical owner; the figure carried as data, never spelled out in prose.

Example (illustrative — fictional Acme Golf Demand Gen, NOT rules)

• A YouTube Demand-Gen golf campaign: detail_placement_view returned youtube.com + a list of video ids. Reading display_name: ~half were golf ("Fix Your Putting Stroke in 10 Minutes…", "Final Round Highlights…") → on_topic; the rest were true-crime and finance ("True Crime: The Cabin Case…", "Turn $1,000 Into $100,000…") → off_topic / sensitive. Rollup ≈ the off-topic share of spend → an exclusion list, and the audience read: "the realized audience skews general-entertainment, not just golfers."