Most Google Ads accounts I have audited have at least 15 percent of spend on keywords that should not be running. Some are old A/B winners that are no longer winning. Some bid on broad-match queries that fire on irrelevant searches. Some have a Quality Score so low that the auction is paying 4x the necessary CPC just to show. Pruning them is one of the highest-ROI activities in PPC, and it almost never gets done because account managers are busy on launches and campaign builds.

An AI agent for Google Ads keyword pruning does the boring part: pull last week's performance, identify the underperformers, draft the changes, hand the manager a CSV that imports straight into Google Ads Editor. The manager spends 30 minutes reviewing, applies, and 15 percent of monthly waste goes away. The agent does not push live changes for the first three months, because PPC errors are expensive and reversibility matters.

What this agent does

The agent pulls last week's keyword and search-term performance from the Google Ads API every Monday. For each keyword it computes the four pruning signals against campaign-level baselines configured by the manager. It generates a change packet with three classes of action: pause underperformers, add negative-match keywords from clearly-irrelevant search terms, and flag low-Quality-Score keywords as rewrite candidates. The change packet is exported as a Google Ads Editor compatible CSV.

What the agent does not do (during the calibration period): it does not apply changes through the API, does not adjust bids, does not change match types, does not move keywords between ad groups, does not edit ads. The write surface during calibration is zero. Auto-apply, when earned, is bounded to a small allowlist.

For the broader pattern of read-then-recommend agents, see what an AI agent can actually do.

Sources of truth

The agent does not browse Search Console, GA4, or landing-page analytics. The Google Ads data is the unit. For the broader rationale, see how to limit agent actions.

The four pruning signals

Every keyword is evaluated against these four. Failing one is not enough; the agent looks at the pattern.

  1. Cost per conversion against target. If the keyword's CPA is more than 1.5x the campaign target over a statistically meaningful window (configurable, default 30 conversions or 30 days, whichever first), the keyword is a pause candidate.
  2. Conversion rate against campaign baseline. If the keyword's conversion rate is below 30 percent of the campaign baseline, the keyword is suspect. Low conversion rate plus high CPA almost always means the keyword is intent-mismatched.
  3. Search-term relevance. For each keyword, the agent samples the triggered search terms and classifies relevance. If more than 25 percent of triggered queries are off-intent, the agent proposes negative keywords to filter them. Broad-match and phrase-match keywords are usually the offenders.
  4. Quality Score trend. A keyword whose Quality Score dropped two points or more in the last 30 days, with low impression share, becomes a rewrite candidate. The agent does not rewrite the ad copy itself; it flags the ad-group ad for the account manager to revisit.

The pause decision requires three of four signals to agree (or two of four plus high spend impact). This is conservative on purpose; false-pausing a winning keyword costs the next two weeks of recovery.

Output: the weekly Editor CSV

The agent's primary artifact is a CSV that imports into Google Ads Editor. Every change is annotated with the supporting signals.

Pause-candidate sheet. Each row: keyword, ad group, campaign, last 30 days spend, CPA, conversion rate, why-flagged. The account manager reviews and decides whether to apply.

Negative-keyword sheet. Each row: proposed negative keyword, match type, target campaign or ad group, supporting search-term examples. The negatives are scoped tightly (typically to the ad group or campaign that triggered the off-intent queries, not the account level), because account-level negatives can cascade.

Rewrite-candidate sheet. Each row: keyword, ad group, current Quality Score, prior Quality Score, why-flagged. No copy is drafted; the manager rewrites with the agent's context.

Summary report. Estimated weekly savings if all changes are applied. Estimated risk (false-positives based on prior weeks' overrides). For the broader monitoring pattern, see how to monitor agent activity.

Guardrails

For the broader safety frame, see AI agent safety and guardrails.

Common mistakes

Frequently asked questions

Can an AI agent prune Google Ads keywords?

It can recommend pruning. The agent pulls weekly performance from Google Ads, identifies keywords whose cost per conversion is well above the campaign target or whose search-term reports contain irrelevant queries, drafts the corresponding pause and negative-keyword changes, and hands the account manager a packet. The manager reviews and applies the changes. The agent never pushes changes live without that step in the first three months.

What does the agent look at?

Four signals per keyword. Cost per conversion against the campaign target. Conversion rate against the campaign baseline. Search-term relevance: percent of triggered queries that match the keyword's intent. Quality Score trend. A keyword that fails on three of four is a pause candidate; a keyword that triggers irrelevant queries becomes a negative-keyword candidate; a keyword with poor Quality Score and low impression share becomes a rewrite candidate.

Does the agent apply changes automatically?

Not by default. The agent generates a Google Ads Editor compatible CSV of proposed changes. The account manager imports it, reviews the diff, and applies. After three months of consistent accept rates above 90 percent, the team can opt specific change types (pausing keywords below a strict threshold, adding negatives from explicit-mismatch search terms) into auto-apply. Everything outside that allowlist stays manual.

What about brand-protection and competitor keywords?

The agent recognises brand keywords (your own brand name and trademarks) and never proposes pausing them based on cost-per-conversion alone, because brand bidding has defensive value beyond direct conversion. Competitor keywords get flagged for human attention but are not auto-classified; the strategy on competitor bidding is too account-specific to delegate.

How often should the agent run?

Weekly is the sweet spot. Daily produces too many small change recommendations against noisy data; campaign performance fluctuates day to day for reasons unrelated to keyword quality. Weekly captures real signal. Monthly is too slow; you lose two to three weeks of wasted spend per cycle.

Three takeaways before you close this tab

Sources