demographics-segmentation

Demographics Segmentation (movement attribution)

Goal (job-to-be-done)

Per demographic segment (age range OR gender), explain which segment drove a metric's up/down move, and which segments convert efficiently — within the campaigns Google Ads can demographically identify. Age and gender are one case (the demographic dimension is the parameter), run through the same contribution-to-change engine. Sub-segment of audience/ in the segmentation/ family.

• Who: the marketer / analyst running the account. When: a metric moved and demographics may explain it, or to find the age / gender worth more bid and the creative angle that fits the segment that buys.
• Decision it drives: demographic bid adjustments (up the efficient segments, down the wasteful) + creative / targeting briefed to the segment that actually buys.

Scope

• ✅ Movement attribution by demographics, age and gender, run as two separate analyses.
• ✅ Coverage-aware: state what % of spend the demographic split covers; claims are scoped "within demographic-eligible campaigns…" when coverage is low.
• ❌ Parental status / household income — NOT exposed by google-ads via Porter (0 fields).
• ❌ Trend over time (aggregate flattens trend) → cross-check a flagged segment in the time/ case.
• ❌ Age × gender cross — shreds the already thin sample into noise.

Components (read these references as needed)

• Tools / data plan: references/tools.md — the two separate queries (age, gender) + the coverage-total query.
• Framework / rubric: references/framework.md — the brain: the mandatory coverage check, contribution-to-change, UNDETERMINED rule, efficiency-over-volume.
• Output schema: references/output.md — the JSON this skill emits.

Operate

Input: per row, a demographic segment value (age range or gender enum) + the base counts (impressions, clicks, cost, conversions, conversions_value), pulled for the report period AND its comparison period. Plus the account-total cost (no dimension) for the coverage check.

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

1. Mandatory pre-check — availability + coverage. Confirm google_ads_age / google_ads_gender exist; sum the demographic spend and divide by the account total to get the coverage %. Google reports demographics only for some campaign types, so the split can cover a small slice. State the coverage % up front. If low, every claim is scoped "within demographic-eligible campaigns…", never an account-level statement.
2. Run age and gender as two separate passes. Never request age × gender together.
3. Contribution to change. For a count metric M with ΔM = M(now) − M(prev), each segment s contributes ΔM_s; rank by |ΔM_s| → the segments that explain the move (they sum back to ΔM, exact). For a ratio (CPA, ROAS, conv-rate): not summed across segments — attribute via its numerator/denominator counts.
4. Keep UNDETERMINED visible. The AGE_RANGE_UNDETERMINED / UNDETERMINED gender bucket is often the single largest — report it, never drop it, never optimize on it.
5. Judge on efficiency, not volume. Rank segments by ROAS / CPA, not spend share — the biggest-spend segment is usually just where the platform aimed.
6. Compute every rate/cost from base counts — native ratio fields are wrong at the aggregate.

Emit the JSON in references/output.md: coverage % up front, per-segment metrics with their ΔM share and an efficiency verdict, the best/worst segment, UNDETERMINED kept visible, and a synthesis whose every figure carries vs previous period.

A renderer (porter-reporting, or a chat view) turns that JSON into a bar chart (segment vs spend) with the coverage % as caption. Do not bake emojis/layout into the analysis output.

Example (illustrative — Acme Golf, 13 weeks, NOT rules)

• Coverage ≈ 13% of spend ($4.1k of $32.7k); the UNDETERMINED bucket was the largest → every sentence scoped "within demographic-eligible campaigns…".
• Age 55–64 = the goldmine: ROAS ≈ 11.8 at the lowest CPA (~$17) → bid_up; 35–44 strong (≈7.0) → bid_up; 65+ worst (≈2.7) → bid_down.
• Gender: male ≈ female ROAS (~6.3), but male took ~94% of demo'd spend — a volume skew (golf audience), not an efficiency gap → leave (no real gender lever).