How should I prioritise which negative keywords to add first?

Order the pass by recoverable dollars, not by how many times a term appears. For each candidate cluster of irrelevant or over-target search terms, multiply its spend over the window by the share of that spend you expect to actually reclaim, and negate the clusters with the largest recoverable figure first. A term that appears fifty times but costs a few dollars total ranks below a single expensive query that burned a meaningful slice of budget with no return. Frequency and alphabetical order — the two ways the search terms report sorts by default — are both blind to dollars, which is why a pass driven by them tends to clear a lot of harmless noise while the real bleed sits further down the list untouched.

What does CPA above account average actually tell me?

It tells you a term is dragging the account mean up, which makes it a candidate for review — not an automatic negative. Account-average CPA is the break-even reference the rest of the account is measured against, so a search term running well above it is spending more per conversion than your typical query and, if it converts rarely or not at all, is a net drain. The nuance the raw comparison hides is relevance: a term above average but relevant and converting may still clear your real margin and belongs to Smart Bidding to manage, whereas a term above average and irrelevant is pure waste. Use the above-average flag to build the review pile, then split that pile by relevance before you negate anything.

Why not just sort the search terms report by cost descending?

Sorting by cost descending is the right first move and an incomplete one. It surfaces the single most expensive queries, but waste in most accounts is not concentrated in a few giant rows — it is spread across many small, thematically related terms that each look trivial alone and add up to real money together. A cost-descending sort shows you the tall individual bars and hides the wide shallow ones. Clustering the report by theme first, then summing spend within each cluster, is what makes the distributed bleed visible, and it is also what lets you negate one defining word instead of fifty individual terms. Cost-descending tells you the biggest rows; cluster-and-sum tells you the biggest problems.

How do I estimate the savings from a negative keyword before adding it?

Take the cluster's spend over a representative window, subtract any spend tied to conversions you would lose by negating it, and treat the remainder as the recoverable figure — then discount it, because negating a query rarely returns every cent. Some of the auctions you block would have been bid down by Smart Bidding anyway, and some budget simply reallocates to other queries rather than vanishing from the account. The point of the estimate is not a precise forecast; it is a ranking key. You are deciding order, so even a rough recoverable number per cluster is enough to put the biggest bleed first, which is the whole value of the exercise. Confirm the actual saving after the fact in your wasted-spend metrics, not before.

Does this prioritisation method work for Performance Max search terms?

Partly, and with a caveat about what the report gives you. Performance Max exposes search categories and terms in its own search terms report rather than the full raw query stream you get in Search, and the figures are privacy-thresholded, so your cluster totals there are directional rather than complete. The prioritisation logic still holds — group what you can see, sum the spend, negate the biggest recoverable clusters first via account-level or campaign negatives — but accept that the denominator is incomplete and that some bleed is invisible by design. Treat the Performance Max pass as a second, coarser sweep after you have ranked and cleared the fully visible Search bleed.

Rank Negative Keywords by Dollars Saved, Not Term Count

Most negative-keyword passes are ordered by the wrong thing. You open the search terms report, it defaults to sorting by impressions or sits in whatever order the UI hands you, and you work down the list negating whatever looks irrelevant until you run out of time or patience. The result is a pass that clears a lot of cheap noise and leaves the expensive bleed sitting untouched further down, because nothing in that workflow ever asked the only question that matters for where to spend your limited attention: which negatives give back the most money.

This post is an ordering rule, not another list of terms to block. The decision of which terms deserve a negative is covered elsewhere on the site; this is about the sequence a finite pass should run in so that the budget comes back fastest. The method is to cluster the report by theme, rank those clusters by the dollars you can realistically recover, and negate the highest-bleed buckets first. It is the difference between a tidy list and a recovered budget, and you can apply it to the report open in front of you this week.

Why frequency and cost-descending both mislead

The two orderings the search terms report makes easy — by frequency and by cost descending — each answer a question that is not the one you need. Frequency ranks by how often a term triggered, which rewards cheap, repetitive junk that feels like progress to clear but barely moves the account. You can negate twenty variations of a term that cost a few dollars between them and feel productive while the real drain goes unaddressed. A pass ordered by frequency optimises for the count of negatives added, which is a vanity number; nobody was ever helped by a longer list.

Cost descending is better and still incomplete. It surfaces the tallest individual rows, but waste in a typical account is not concentrated in a handful of giant queries — it is distributed across many small, related terms that are individually trivial and collectively expensive. Sorting by cost shows you the tall bars and hides the wide shallow ones, so a cost-led pass tends to catch a few obvious offenders and miss the thematic bleed that only becomes visible when you add the small rows up. The fix for both failures is the same: stop ranking individual rows and start ranking clusters.

Cluster the report before you rank anything

Clustering is what turns a flat list of queries into a ranked set of problems. Group the search terms by theme — the shared word or intent that makes a set of queries the same kind of wrong — and sum the spend inside each group. Now the wide shallow bleed that cost-descending hides becomes a single tall bar: forty “free,” “diy,” and “how to” variations that each cost pennies sum into a cluster worth a real negative. This is also the move that lets you block one defining word instead of chasing every surface form, which is the whole argument of grouping patterns versus single negatives.

You do not need a tool to start — a pivot of the exported report by a theme word will do — but the mechanical version of this clustering is n-gram analysis, which counts the recurring one- and two-word tokens across the report so the themes surface by frequency of spend rather than by eye. The n-gram approach to finding negatives is the engine; the point here is what you do with the output. N-gram grouping tells you the clusters exist. Ranking them by recoverable dollars tells you which one to negate first, and that second step is where the budget actually comes back.

Use CPA above account average to build the review pile

Account-average CPA is the reference your whole account is implicitly measured against, so a cluster whose cost-per-conversion runs well above it is spending more than your typical query to produce a result — and if it converts rarely or never, it is a drain. This is the standard first filter practitioners reach for, and the common guidance is to flag terms whose CPA sits far above the account mean; one widely repeated rule of thumb is to flag terms whose CPA runs several times the account average as review candidates. The flag builds the pile. It does not decide the negative.

The step people skip is splitting that pile by relevance before acting, because “above average CPA” bundles two opposite cases. A cluster that is above average and irrelevant — wrong product, wrong intent, job seekers, free-version hunters — is pure waste and negates on sight. A cluster that is above average but relevant and still converting above your real break-even is a bidding problem, not a waste problem, and negating it throws away conversions Smart Bidding was managing. That distinction is the entire thesis of when a negative keyword lowers your ROAS, and it is why the above-average flag is a sorting key for review, never an auto-negate trigger.

Estimate recoverable dollars per cluster

With the review pile clustered and the irrelevant clusters separated out, you can put a rough recoverable figure on each one and rank by it. Take the cluster's spend over a representative window, subtract any spend tied to conversions you would genuinely lose, and discount the remainder — because negating a query never returns every cent. Some of those auctions Smart Bidding was already bidding toward zero, and some budget reallocates to other queries rather than leaving the account. The number you are left with is not a forecast; it is a ranking key, and even a coarse one does the job of putting the biggest bleed first.

Resist the urge to make this precise. The value of the estimate is entirely in the ordering it produces, and the ordering is robust to a lot of imprecision: a cluster bleeding ten times what the next one does will sort to the top under almost any reasonable discount. What you should not do is confuse the pre-negation estimate with the actual result — that is a measurement you take after, in your wasted-spend metrics scorecard, by watching whether spend fell and conversions held. The estimate decides the order of the pass; the scorecard tells you whether the pass worked.

Run the pass in savings order

Put it together and the pass has a clear shape. Export the search terms report for a window long enough to be significant, cluster it by theme, flag clusters whose CPA runs above account average, split those by relevance, and attach a recoverable-dollar estimate to each irrelevant cluster. Then work strictly top-down by that estimate — biggest recoverable figure first — negating the defining word of each cluster rather than its individual terms. When your time runs out, you will have cleared the bleed that mattered, not the noise that was merely convenient, and the list you built will be short and expensive rather than long and cosmetic.

Two guardrails keep this from backfiring. First, confirm each cluster is significant before you act on it — a high CPA on three clicks is a small sample, not a verdict, which is the point of how many clicks before negating a search term. Second, do not re-run the whole exercise daily; a savings-ranked pass is a cleanup cadence, not a reflex, and the same clusters will not keep topping the list once cleared. For the full context this sits inside — the end-to-end hygiene routine — see the search hygiene audit playbook, and for the raw-dollar view of where the waste hides, the hidden wasted spend in your search terms. The idea of ordering the work by economic impact rather than term count is one several practitioners have landed on independently from search-term pattern analysis; the contribution here is making the recoverable dollar, not the token count, the key you sort on.

For the clustering engine that surfaces the themes in the first place, see n-gram search-term analysis to find negatives. For the relevance-versus-cost split that decides whether an above-average cluster is waste or a bidding problem, see when a negative keyword lowers your ROAS, and for measuring whether the pass actually recovered budget, see the wasted-spend metrics scorecard.