Add negative keywords in Google Ads only when three things are on the table for each candidate: a search terms report showing the query actually triggered your ads, a business-relevance judgment explaining why that query cannot convert for you, and an overblocking review confirming the negative will not cut real demand. That is the whole evidence rule. It also answers how to make a negative keyword list: build it from your own search terms report, never from a copied starter list. Everything else in negative keyword management, harvesting cadence, broad-match monitoring, account-level lists, exists to feed those three inputs.
Most wasted-spend problems I see do not come from missing negatives. They come from negatives added without evidence: copied from a "starter list" blog post, guessed from intuition, or bulk-imported from another account. Those negatives feel like work getting done, and some of them quietly block queries that were converting.
Why generic negative keyword lists fail
A generic negative list answers a question nobody asked. It encodes assumptions about a business the list author never saw: that certain modifiers signal low intent, that certain audiences never buy, that certain phrasings mean research rather than purchase. In one account those assumptions hold. In the next one they block the exact query a buyer types.
Two failure modes show up repeatedly:
Overblocking. A phrase or broad-match negative catches more than the term you meant. You added it to kill one bad query, and it silently suppresses a family of good ones. Because blocked queries never appear in your search terms report, the damage is invisible unless you specifically review what a negative could match before adding it.
Stale assumptions. A negative that was correct at the time outlives the reason it was added. Nobody remembers why it exists, nobody owns reviewing it, and the account carries an invisible ceiling on reach. This is a governance problem, not a keyword problem.
The Claude Ads Google audit treats both as first-class problems. Its catalog includes negative-keyword governance, account-level negatives, material irrelevant search-term spend, and search-term audit recency as distinct checks, each evaluated only from current account evidence, never from a fixed platform-wide threshold. You can see the full control coverage on the Google Ads platform page.
The negative keyword evidence rule, spelled out
Before any negative goes in, each of these must exist:
| Evidence | What it proves | Where it comes from |
|---|---|---|
| Search terms report | The query really triggered your ads and spent money in the window you care about | Google Ads search terms report or API export for a defined date range |
| Business-relevance review | The query cannot plausibly convert for this specific business | Operator knowledge: the offer, the audience, the sales motion |
| Overblocking review | The negative, at its chosen match type and level, will not block converting or plausible queries | Checking the negative against existing converting terms and likely query families |
If any row is missing, the correct action is to go get it, not to substitute a guess. This is exactly how the product behaves: Claude Ads will not generate, suggest, or illustrate specific negative keywords without a search terms report plus business-relevance and overblocking review. It requests the missing evidence instead of naming sample or "commonly excluded" terms as a workaround. That constraint annoyed me the first time I hit it. Then I audited an account where a copied starter list had been blocking a converting query family for months, and the constraint made sense.
How to harvest negative keywords from search terms
Harvesting is the supply side of the evidence rule. The search terms report is the only place Google shows you what you actually paid for, so review it on a schedule, not when performance drops.
A workable routine:
1. Pull the report for a meaningful window
Use a window long enough to capture query volume, short enough that the account state it reflects still exists. Keep the export tied to its date range, campaign scope, and pull date, because a search-term decision without lineage is just an opinion. Claude Ads normalizes every export with retained lineage for exactly this reason.
2. Segment terms by decision, not by metric
Sorting by cost is a start, but the decision categories matter more:
- Terms that converted: candidates for adding as keywords, not for negatives.
- Terms with spend, no conversions, and clear irrelevance: negative candidates, pending the overblocking review.
- Terms with spend, no conversions, but plausible relevance: investigate before blocking. Conversion lag, landing page mismatch, or thin volume can all explain zero conversions on a good query.
- Terms you cannot see: rising percentages of hidden or low-volume terms shrink your evidence base, which is itself an audit finding (search-term visibility is its own control in the audit catalog).
3. Run the overblocking check before committing
For every negative candidate, decide match type and level (ad group, campaign, or shared account-level list) by asking what else it could catch. Check the candidate against your converting search terms first. A negative that overlaps a converting query family needs a narrower match type or should not go in at all.
Broad match is a harvesting machine, treat it like one
Broad match and close variants are the main reason harvesting never ends. Broad match trades control for reach, and the search terms report is where you find out what that trade cost this month. Close variants do the same thing to your exact and phrase keywords at smaller scale: Google matches queries it considers equivalent, and equivalence is Google's judgment, not yours.
The Google audit scores this directly: broad-match control and close-variant pollution are separate controls in the catalog, alongside material irrelevant search-term spend. The audit does not apply a universal rule like "never use broad match," because with strong conversion data and smart bidding, broad match can be the right call. What it checks is whether the leakage is measured, whether negatives are governing it, and how recently anyone looked.
In practice, the loop runs through two commands. /ads audit surfaces the problem:
/ads audit
It returns findings with confidence graded high, medium, low, or none, and evidence coverage graded (80%+ inputs), provisional (60-79%), or insufficient (below 60%). If the search terms report is missing from your inputs, the related controls return as unknown instead of being guessed at.
Then /ads optimize turns findings into action:
/ads optimize --draft
Search-term actions in optimize require the same three pieces of evidence, and every proposed negative stays a draft until you explicitly approve it through the mutation gate. Applied changes retain a rollback. Nothing touches the account by default.
The honest limits
The evidence rule does not make negative keyword decisions easy, it makes them defensible. Business relevance is still a human judgment about your offer. Overblocking review is bounded by the queries you can see, and Google hides part of the query stream. And a well-governed negative list still needs periodic review, because the business changes underneath it. What the rule removes is the specific, common failure of blocking demand based on someone else's assumptions.
Try it
Claude Ads runs this workflow, harvest, evidence check, drafted actions behind an approval gate, as part of its Google Ads coverage across a 12-platform, 33-skill stack. Install it from the installs page or directly in Claude Code:
/plugin marketplace add AgriciDaniel/claude-ads
/plugin install claude-ads@ai-marketing-hub-claude-ads
Run /ads audit on your Google account first. The search-term findings will tell you whether your negative list is governing spend or just accumulating history.