Most advice about the search terms report treats negating a query as a pure win: you found waste, you blocked it, the money comes back. That is true for an irrelevant term and dangerously untrue for a relevant one. Under Smart Bidding, a negative keyword does two things at once — it stops the spend, and it removes the query from the pool the bidder is allowed to compete for. When the query was producing conversions, the second effect is a cost, and it is invisible in the place you are looking. The waste number in your report improves while the account's return quietly falls.
This post is the decision rule for the case the rest of the site's negation advice assumes away: the search term that looks wasteful but is pulling marginal, break-even-clearing conversions the bidder is relying on. We will separate the two kinds of “bad” term, explain why average ROAS is the wrong number to negate on, show how a big negation batch can disturb Smart Bidding's learning, and end with a rule you can apply to a specific row in your own report this week. The goal, as always here, is to leave you able to act — not to add a list of negatives, but to know which ones are free and which ones are a trade.
Two kinds of bad term, one list
The search terms report mixes two completely different problems under one word: “bad.” The first is irrelevance — a query whose intent has nothing to do with what you sell. Someone searching for a free version, a job, a competitor, or a product tier you do not offer will not convert for you no matter how the auction is bid, and the bidder cannot know that because relevance to your business is not something it can infer from conversion data alone. These terms are the free negatives. Blocking them returns spend with no offsetting loss, and you should negate them on sight without consulting their ROAS at all.
The second kind of bad term is merely expensive: relevant intent, real conversions, but a cost-per-conversion that looks ugly next to your target. This is where people make the costly mistake, because the two kinds sit on the same screen and get treated the same way. An expensive-but-converting term is not waste — it is a pricing problem the bidder exists to manage. Negating it does not fix the economics; it removes the conversions entirely. The whole skill of reading the report for ROAS is refusing to let a high cost column trigger the same reflex that an irrelevant query should. The over-negating failure mode is exactly this reflex at scale: a list built on cost rather than relevance that throttles traffic the account needed.
Negate on marginal value, not average ROAS
The number that should drive a cost-based negation decision is marginal ROAS — the return on the next dollar — not the average ROAS splashed across the row. A term can show a poor average while its marginal contribution is still positive, and those are different questions with different answers. Average asks “how did all the spend on this term perform?” Marginal asks “if I remove this term, do its conversions come back through other queries, or are they simply gone?” If they are gone and they cleared your real break-even, the negative is a loss dressed up as a cleanup. Allocating on average while ignoring marginal is, as practitioners managing Target ROAS campaigns put it, how you end up excluding terms that were still pulling their weight.
Translating that into something you can do at the row level: before you negate a relevant term for being expensive, answer whether its conversions are profitable after your actual margins, not after a target ROAS you set as a round number. A term converting at a cost-per-acquisition above your headline target can still be above water once real margin and lifetime value are in the picture, and killing it shrinks the account for no gain. This is the same judgement the how many clicks before negating question turns on — statistical significance before you act — extended one step: significance tells you the numbers are real, marginal value tells you whether acting on them helps.
Smart Bidding has already negated most of your waste
Before you add a manual negative for cost, it is worth internalising that Smart Bidding is itself a suppression engine. Target ROAS and Target CPA already bid down or decline the auctions they predict will convert poorly, so for a large share of genuinely weak queries the algorithm has effectively negated them on your behalf by pushing their bids toward zero impressions. A manual negative on one of those is redundant — you are blocking a query the model was already starving, which changes nothing except the size of your list. The redundancy question is worked through in full in negatives versus Smart Bidding auto-suppression, and the short version is that cost-based negatives are the ones most likely to be doing work the bidder has already done.
The friction appears when your read of a term and the model's disagree. If the bidder is still spending on a query, it is because it sees conversion probability there that justifies the bid — signal that may not be legible to you from a handful of rows. Negate it and you overrule the model on evidence it has and you do not, and you remove that query's data from what it learns on. This is why the strongest practitioners argue that negatives, used carelessly under automated bidding, can do more harm than good: not because negatives are wrong, but because cost-driven ones fight the bidder instead of informing it. Keep your manual negatives to the thing the bidder cannot judge — relevance — and let it keep managing the thing it can.
The hidden cost of a big negation batch
A single negative is a small change; a batch of a hundred added in one sitting is a different event for the campaign. Smart Bidding models the relationship between queries, auctions, and conversions, and removing a large slice of the query surface at once changes the inputs the model has been optimising against. In practice a heavy negation batch can nudge a campaign into a learning-like phase where bids are recalibrated and performance is noisier than usual while things settle. The waste you removed is real, but so is the temporary instability, and reading the result in the first few days will tell you more about the disruption than about the decision.
This is not an argument against ever cleaning up in bulk — a neglected account with a backlog of obvious irrelevant terms is worth a decisive first sweep, and the noise is a price worth paying once. It is an argument against two specific habits: judging a batch too early, and dumping large negative lists repeatedly so the bidder is perpetually re-learning against a moving target. Stage your big cleanups, keep the irrelevant-term sweep separate from fine cost-based tuning, and give the campaign a full conversion cycle before you decide whether the batch helped. The cadence and the metrics to watch while it settles are laid out in the wasted-spend metrics scorecard, which is the right instrument for separating a genuine improvement from learning-phase noise.
A decision rule for the row in front of you
Here is the rule, applied to a single search-term row. First ask whether the query is irrelevant by meaning. If it is — wrong product, wrong audience, free or DIY intent, job seekers, an unrelated brand — negate it immediately and do not look at its ROAS, because it will never convert for you and the bidder cannot work that out. Relevance-based negatives are the free ones, and they are the bulk of a healthy negative list. This is the same clustering discipline that makes a first pass efficient: negate the defining word of an irrelevant theme rather than chasing individual terms.
If the query is relevant but expensive, switch rules entirely: do not reach for a negative. Check whether it converts above your real break-even over a window long enough to be significant. If it does, leave it and let Smart Bidding manage the bid — that is the bidder's job and it is better at it than a static negative. Only negate a relevant, converting term when its conversions are genuinely unprofitable after your actual margins and you have a reason to believe the freed budget earns more elsewhere. In one line: negate for irrelevance freely, negate for cost rarely, and never negate a converting term without first answering the marginal-ROAS question. A negative list built that way protects return instead of quietly eroding it.
Related reading
For the volume question that tells you when a term's numbers are real enough to act on, see how many clicks before negating a search term. For the failure mode this whole post is trying to prevent at scale, see over-negating and throttling good traffic, and for whether a given negative is doing anything the bidder has not already done, see negatives versus Smart Bidding auto-suppression.