Automating negative keywords in Google Ads

What a negative keyword actually does

A negative keyword stops your ad entering an auction when the search contains what you excluded. It is not a bid adjustment and it is not a preference — the ad simply does not show. Three consequences follow that people are regularly surprised by.

• It is not retroactive. Adding a negative today does nothing about the money spent on that term last month. This is the whole argument for a weekly cadence rather than a quarterly one.
• It works on the search, not on your keyword. Broad match and Google's own query expansion are what put you in front of those searches; the negative is the only instrument you have for saying no.
• It means something different outside Search. In Display, YouTube and Demand Gen there is no query, so negative keywords exclude content and context rather than blocking a search. The same list will behave differently in a Display campaign than it does in a Search one.

The three negative match types, which are not the positive ones

This is the part that trips people up. Negative match types share the names of positive ones and behave differently, and the default is the loosest of the three.

The practical rule: phrase for most things, exact when you want to remove one specific search and keep its neighbours, broad only when each individual word is disqualifying by itself. A negative broad “free” blocks “is it free to try” along with everything else, and for a freemium product that is a customer you just declined to meet.

Account, campaign, ad group, or a shared list

Four places, and choosing the wrong one is how negative lists become unmaintainable.

Performance Max deserves a note of its own: it takes account-level negatives, and campaign negatives where the campaign supports them, but there is no keyword list to prune in the way a Search campaign has one. What you can and cannot steer there is set out in the Performance Max page.

The patterns worth blocking

Wasted spend is rarely one dramatic mistake; it is a long tail of terms that each look defensible on their own. The recurring categories, with illustrative examples rather than anyone's data:

• Employment. jobs, salary, vacancy, recruitment, internship, careers. The largest single category for anything with a professional audience, and almost always cheap enough to go unnoticed for months.
• Study and DIY. course, certification, exam, tutorial, template, how to build. People intending to do it themselves rather than buy it.
• Existing customers. login, sign in, support number, cancel, refund, app download. You are paying to reach people who already pay you, often outbidding your own organic result to do it.
• Free intent. freeware, open source, crack, torrent. Explicit signals that no money is coming — though note that “free” on its own is not one of these if you sell a free tier.
• Wrong entity. Homonyms, a product in another industry with your name, a well-known person who shares it. Pure broad-match drift, and the most common source of quiet spend on an otherwise well-run account.
• Definition and research. what is, meaning, wikipedia, pdf. Worth having deliberately in some accounts. It should be a decision rather than an accident.

Equally important are the ones that look obviously bad and are not. Competitor names are often the best-converting traffic in a B2B account. “Cheap” is a qualified buyer if you compete on price. “Reddit” and “review” are late-stage research in several categories. This is exactly the judgement that a word list cannot make and a person, or a model with your context, can — which is the argument for interpretation over a static blocklist. The wider set of checks lives on the AI audit page.

Why this is the safest thing to automate first

• It is additive. Nothing is destroyed. Adding a negative creates a criterion; it does not remove your keyword, your ad or its history.
• It can be removed in seconds. Delete the negative and the traffic returns. Compare that with a restructure, which cannot be put back in an afternoon.
• It does not disturb the bidding. Changing a bid strategy throws a campaign back into learning. Adding negatives narrows the traffic the existing strategy bids on, which is a much gentler thing to do to a live campaign.
• It is verifiable. You can see the exact terms that prompted each exclusion, before and after. Very little other automation offers a check that direct.
• It is high frequency and low stakes. The ideal shape for a first automation: it happens often enough that you will find out quickly whether you trust it, and the cost of being wrong once is small.

This is why it is the recommended first job in the sequence on earning autonomy: questions for a week, then one account, then one job — and this is the job.

Doing it with an AI connector

With an assistant connected to the Google Ads API, the whole loop is a conversation. The reading half runs on the read-only tier and changes nothing.

1. Pull the terms. “List search terms from the last 90 days with cost above [threshold] and zero conversions, sorted by cost.” Underneath, that is one query:

   SELECT search_term_view.search_term,
          campaign.name,
          metrics.cost_micros,
          metrics.conversions
   FROM   search_term_view
   WHERE  segments.date BETWEEN '2026-06-01' AND '2026-08-29'
     AND  metrics.conversions = 0
   ORDER BY metrics.cost_micros DESC

2. Group them by intent. “Which of these look like job seekers, students, existing customers or free-tier hunters, and which are too small a sample to judge?” The second half of that question matters more than the first.
3. Ask for the negative, not just the term. “Propose negatives with match types, and say which campaigns each should go on.” A good answer proposes phrase-match exclusions and flags any that risk blocking traffic you want.
4. Approve, then apply. On the full tier the assistant calls the connector's negative-keyword tool, which adds them at campaign level with the match type you agreed.
5. Read the trail. Every call is recorded — which tool, which account, what happened — so “what did it add on Tuesday” has an answer.

How this goes wrong

• Over-broad exclusions. The most expensive mistake in this whole area, and the hardest to see: nothing in the interface reports the conversions you did not get. One-word negative broad terms are where it usually starts.
• Blocking your own keywords. A negative can conflict with a keyword you are actively bidding on, and the campaign quietly stops serving for it. Google flags these conflicts; it is worth asking your assistant to check for them explicitly.
• Acting on too little data. A term with nine clicks and no conversions has told you almost nothing. Automation is very willing to treat it as a finding.
• Starving the algorithm. Aggressive exclusion on a broad-match campaign running Smart Bidding narrows the pool it learns from. Trim steadily rather than in one large pass.
• Lists nobody prunes. Exclusions outlive their reasons. A term blocked during a product you no longer sell is still blocked three years later, and there are caps on how many negatives a campaign or list will hold.

None of this argues against automating the work. It argues for the shape AdCopilot takes: the reading can run on a schedule you set, the proposal is explicit, the approval is yours, and the record of what happened is complete. How the tool list is grouped is on the MCP server page, and connecting Claude takes about four steps.