90-Day Search-Term Review

Variables

Replace each placeholder before you run the prompt.

• {{industry_context}}Vertical and target persona — needed to distinguish B2B intent from B2C contamination.

  B2B industrial fluid power, target persona is plant maintenance engineer at $20M+ manufacturers.

• {{search_terms}}Search-term report rows: term, impressions, clicks, cost, conversions. Tab-separated or CSV.

  See attached CSV — 3,200 rows, 90-day window, ~$28K spend.

You are reviewing a 90-day search-term report for a B2B account. Your job is to produce two outputs: a comprehensive negative keyword list, and a list of expansion candidates worth promoting to keywords.

INDUSTRY CONTEXT:
{{industry_context}}

SEARCH TERMS (term, impressions, clicks, cost, conversions):
{{search_terms}}

NEGATIVE KEYWORD LIST:
Group negatives by reason. Standard buckets:
- B2C contamination (DIY, hobby, retail signals)
- Wrong-product (we don't sell this)
- Competitor terms (decide per-account whether to include)
- Job seeker / informational (jobs, careers, what is)
- Geographic mismatch
- Pricing intent we don't serve (cheap, free, sample-only)
- Unclear / low-volume (recommend monitoring, not adding)

For each negative, suggest match type (negative exact / negative phrase / negative broad). Default to negative phrase unless exact is needed.

EXPANSION CANDIDATES:
List terms with conversion rate above the account average, ranked by total conversions. For each: suggest match type for promotion, suggest the campaign / ad group it should go in.

Do not invent terms not present in the data. Do not recommend "let it run longer" — the window is already 90 days.

Expected output shape

A categorized negative-keyword list and an expansion-candidate list, both with match-type recommendations. Typically 50–200 negatives + 10–30 expansion candidates.

Why we wrote it

The standard search-term review pattern in agencies is unstructured — one analyst eyeballs the report and copies negatives into a spreadsheet. This prompt forces categorization and produces a defensible artifact.

How to use

1. Open Claude or ChatGPT. The recommended model for this prompt is claude-sonnet-4-6 — opus when the prompt requires deep reasoning, sonnet for the rest.
2. Replace every {{variable}} with content specific to your account. The examples above are starting points, not templates to ship as-is.
3. Paste the prompt and run.
4. Read the output against the expected shape above. If the model produced a structurally different response, re-prompt rather than accept the drift.