What each part of the prompt does
- Two labels only
- Forcing keep or exclude removes the "maybe" column that usually ends up as a second review. The tie-break rule, keep when unsure, decides the borderline rows for you.
- Intent over wording
- Telling Gemini to read what the searcher wants stops it from excluding clumsy phrasings, typos or mixed-language queries that still come from people ready to buy. Those terms are often cheaper than the polished versions.
- The not-offered list
- Your list of what you do not sell is the evidence for every exclusion. Without it, the model has to guess, and the prompt tells it not to; terms it cannot rule out stay in.
- Negatives with a safety check
- Each negative keyword comes with a keep term it must not block. That one line catches the common mistake where a phrase negative like "free" also removes "free delivery" searches that convert.
Variations to try
Same idea, one thing changed.
Shopping campaign You are auditing the search terms report of a Google Shopping campaign for wasted spend. What the business sells, to whom and where: [BUSINESS] What it does not sell or serve: [NOT OFFERED] Search terms report, pasted with columns term, campaign, clicks, cost, conversions: [REPORT] Sort every term into KEEP or EXCLUDE. Nothing else. How to decide: - Judge what the person is looking for, not the exact wording. Awkward or misspelled searches can still come from buyers. - Exclude only terms that clearly cannot lead to a sale: a product or service we do not offer, a country we do not ship to, job seekers, people looking for a brand or model we do not stock, people who only want something free, or existing customers looking for support. - If you are unsure, keep it. A few wasted clicks cost less than blocking a real buyer. - Any term with a conversion is kept. - Use only the business information above; do not assume more. Output: 1. A table: Term, Decision, What the searcher wants, Reason. 2. The shortest set of negative keywords that removes every EXCLUDE term, each with its match type (phrase or exact) and one KEEP term it must not block. 3. The total cost of the EXCLUDE terms, added up from the report, with the sum shown.
Lead generation You are auditing a Google Ads search terms report for wasted spend. What the business sells, to whom and where: [BUSINESS] What it does not sell or serve: [NOT OFFERED] Search terms report, pasted with columns term, campaign, clicks, cost, conversions: [REPORT] Sort every term into KEEP or EXCLUDE. Nothing else. How to decide: - Judge what the person is looking for, not the exact wording. Awkward or misspelled searches can still come from buyers. - Exclude only terms that clearly cannot lead to a sale: a product or service we do not offer, a country we do not ship to, job seekers, people who only want something free, or existing customers looking for support. - If you are unsure, keep it. A few wasted clicks cost less than blocking a real buyer. - Any term with a conversion is kept. Students and researchers looking for definitions are excluded; small businesses asking how to do the job themselves are kept, since some will hire us. - Use only the business information above; do not assume more. Output: 1. A table: Term, Decision, What the searcher wants, Reason. 2. The shortest set of negative keywords that removes every EXCLUDE term, each with its match type (phrase or exact) and one KEEP term it must not block. 3. The total cost of the EXCLUDE terms, added up from the report, with the sum shown.
If the result isn't right
- Export at least thirty days of terms; a single week hides patterns and makes one expensive click look like a trend.
- Add exclusions as phrase match at the campaign level first and check the next report before moving them to an account-wide list.
- Recalculate the cost total yourself in a spreadsheet; models can drop a row when adding long columns.