Search term clusterer: themes, spend totals and a keep, negate or watch verdict

The search terms report is where wasted spend hides in plain sight: hundreds of rows, each individually too small to notice, that add up to a real number when grouped by theme. Clustering them by hand is slow, and the temptation is to skim the top ten and stop. This prompt does the grouping, sums the spend per theme, and gives each cluster a plain verdict so you can act rather than admire the list.

It expects the columns you get from a standard export (search term, cost, clicks, conversions) and returns themed clusters, a spend total per cluster, a keep, negate or watch call, proposed negative keywords with match types, and an explicit note of every assumption it made. It changes nothing. You review the negatives and add the ones you agree with.

The prompt

You are a paid search analyst reviewing a search terms report from Google Ads.
Work only from the rows I paste. Do not invent terms, costs or conversions, and
do not assume a conversion is valuable beyond what the data shows. Where you
have to infer intent from a term, say so. Write in British English.

INPUT
I will paste rows with these columns: search term, cost, clicks, conversions.
Currency is the account currency unless I say otherwise.

TASK
1. Group the terms into themed clusters by intent (for example: brand,
   competitor, informational, wrong-product, location, high-intent commercial).
   Name each cluster plainly.
2. For each cluster show: total cost, total clicks, total conversions, and the
   terms it contains (or a representative sample if there are many, and say how
   many you folded in).
3. Give each cluster one verdict:
   - KEEP: converting or clearly on-intent, leave it running.
   - NEGATE: off-intent or spending with no conversions, block it.
   - WATCH: too little data to judge, note the spend threshold at which you
     would revisit it.
4. Propose negative keywords for the NEGATE clusters, each with a match type
   (exact, phrase or broad) and a one-line reason. Prefer the narrowest match
   that does the job, and flag any negative that risks blocking a good term.

ASSUMPTIONS
End with a short list headed "Assumptions" covering anything you inferred:
intent you guessed, terms you could not classify, and any row you ignored and
why. Do not present an inference as a fact.

How to use it

Export the search terms report for the date range you care about, paste the four columns straight into the chat, and run the prompt. This is a reasoning task, so it suits a mid-to-upper tier model; smaller models cluster loosely and mislabel intent. Read the assumptions list first, then work down the NEGATE clusters and add the negatives you agree with at the right level (account, campaign or ad group). Treat WATCH clusters as a note to yourself, not an action. For very large exports, paste in batches by campaign rather than dumping thousands of rows at once, which degrades the grouping.

Failure modes

• It reads a conversion as a good conversion. The report shows that something converted, not that the conversion was profitable or correctly tracked. The prompt tells the model not to assume value, but it cannot see your margins. Countermeasure: sanity-check any KEEP cluster against conversion value or CPA before you act on it.
• A proposed negative is too broad and blocks a winner. A phrase or broad negative can catch terms you want. The "flag any negative that risks blocking a good term" instruction helps, but it is not infallible. Countermeasure: read each negative against your keyword list, and default to exact match when in doubt.
• Truncated or malformed paste. If the export loses a column or the model only sees the first screenful, the totals will be wrong and stated confidently. Countermeasure: confirm the row count it reports matches what you pasted, and split large exports into batches.