The Weekly Negative Keyword Sweep With AI

The weekly sweep is how your agent hunts and kills the spend that buys nothing — four steps: it pulls the search terms that cost money without converting, classifies them in seconds, you judge the shortlist, then approve one write per campaign. Ten minutes, once a week, and it compounds — every blocked term stays blocked. This page is the workflow; the concepts behind it — match types, negative levels, which patterns to block — live in negative keywords automation.

Step 1 — pull and classify

One prompt does the reading and the first pass of judgement:

In account <account>, pull search terms from the last 14 days with at
least one click and zero conversions, ranked by cost. Classify each:
- irrelevant (nothing to do with what we sell)
- competitor (a competitor's brand)
- research (free / how to / tutorial / diy intent)
- plausibly fine (could convert with more time)
Show cost and the matched keyword per term. Recommend which to block and
at what match type — but do not add anything yet.

Expect a classified table and a recommendation list. The matched keyword column is the sleeper: when one keyword matched most of the junk, the finding is not "add twenty negatives", it is "that match type is too loose" — a better fix upstream.

The "do not add anything yet" line matters. It keeps the read and the write in separate steps, which is the habit that makes every agent workflow reviewable — the same evidence-before-action rule the whole prompt library is built on.

Step 2 — judge the shortlist

The agent's classification is a draft, not a verdict. Three rules of thumb while you scan it:

• Irrelevant terms: block on sight, exact match. A term that has nothing to do with the offer needs no more evidence than its own text plus the cost column.
• Research intent: block the pattern, phrase match. One "how to" term is noise; five is a pattern. Phrase-match negatives on the pattern word beat twenty exact negatives on its variations.
• Competitor terms: decide, don't default. Sometimes competitor searches convert; sometimes they burn. Ask the agent for the 30-day cost and conversion history on those terms specifically, and make this call once per quarter rather than re-litigating it weekly.

And one restraint: leave "plausibly fine" alone. A sweep that blocks everything that has not converted yet strangles the campaign's ability to learn. You are removing waste, not variance.

Step 3 — approve the write

Add the confirmed list as negatives:
- exact match for the one-off irrelevant terms
- phrase match for the pattern terms
to the campaigns where each appeared. Show me the final per-campaign
list, then make the calls.

Your client surfaces one add_negative_keywords approval per campaign — the account, the campaign, every term and its match type, visible before anything executes. Read the term list one last time in the approval itself, then approve. Negatives added through the connector land at campaign level, attributed to your own Google sign-in in change history.

This is the write worth granting first when you are new to agent-managed accounts: it cannot increase spend, it cannot touch ads or budgets, and it is exactly the job whose manual version — export, pivot, squint, paste — was tedious enough that most accounts simply never did it weekly.

Step 4 — log and stop

Summarise this sweep: how many terms reviewed, how many blocked per
campaign at which match types, total 14-day cost of the blocked terms.

The last number is the sweep's receipt: the spend those terms accumulated in a fortnight is roughly what next fortnight would have cost you. Paste the summary into your notes and stop — the sweep is deliberately a bounded job, one slot inside the 20-minute weekly routine rather than an open-ended optimisation session.

The quarterly counter-sweep

Once a quarter, run the sweep in reverse:

In account <account>, review all existing negative keywords. Which are
phrase-match negatives broad enough that they could plausibly block
traffic we want? Which were added for a problem that no longer exists?

Over-blocking is the sweep's only failure mode, and it is invisible in weekly data — blocked queries never appear in any report. Ten minutes a quarter reviewing the accumulated list keeps years of sweeps from slowly walling off real demand. Your first-ever sweep should also check the inverse conflict — negatives colliding with keywords you actively bid on — which is step five of the 30-minute audit.

Run one now

If an account is connected, the whole first sweep fits in the next ten minutes: paste the Step 1 prompt, judge the table, approve the write, keep the receipt. If not, Start a free pilot — connect with your Google sign-in and make the sweep the first thing the agent does for you. It is the fastest route from "connected" to "found money".