Open the search terms report, sort by cost, and the top of the list looks like the whole problem: a few expensive queries that obviously should not have run. Add them as negatives and it feels like the job is done. It is not, and the reason is a quiet arithmetic problem. The biggest single terms are rarely where most of the money goes. Most wasted spend is spread across hundreds of small-cost queries that no sort brings to the top, plus a whole layer of terms Google will not show you at all.
This is the part of wasted-spend work the search terms report walkthrough sets up but does not exhaust: once you know how to read the report and write a negative, the next skill is finding the waste that a cost sort structurally cannot surface. That is what this post is about.
Why the top of the list lies to you
Wasted spend follows a long-tail distribution, and a cost sort only reads the head of it. The top twenty terms by cost usually contain your most obvious offenders, and dealing with them is real work worth doing — Optmyzr notes that in many accounts the highest-cost search terms hide a large share of the waste. But "a large share of the visible waste" is the catch. Below those terms sit hundreds of queries each costing a pound or two, individually too small to notice and collectively often larger than the head.
The reason this matters is that broad match and Performance Max both generate long tails by design. A broad keyword can match thousands of distinct queries, most of them appearing once or twice, and each landing near the bottom of a cost sort. Search Engine Land's account audits routinely find that 20-30% of most accounts' spend is quietly underperforming, and very little of that shows up as a single expensive term. It accumulates as drips. If your entire audit is a cost sort, you fix the drips you can see and keep paying for the ones you cannot.
The layer Google will not print
Below even the low-cost long tail is a layer you cannot see at all. Google withholds search terms that were not searched by enough different people, for user privacy, so those impressions and clicks happened and cost you money but never appear as a row. The straightforward test is to total the cost of every visible row and compare it to the campaign total for the same period — the visible rows always come to less, and the gap is the hidden portion. In many accounts that gap is substantial.
Worse, the report has drifted from a record of queries toward a summary of intent. Practical Ecommerce has documented Google decreasing search-terms visibility, and the terms that remain increasingly describe a pattern Google inferred rather than the exact text a person typed. The practical upshot is the same either way: you cannot audit the hidden layer query by query, because there are no queries to audit. The only tool that reaches it is a thematic negative — a broad negative on a wasteful word blocks every query containing it, visible or withheld. That is the whole case for thinking in themes rather than strings.
Reading the long tail with an n-gram pass
The way to find waste a cost sort misses is to stop reading queries one by one and start counting the words inside them. An n-gram analysis takes every search term, breaks it into individual words and short phrases, and totals the cost and conversions attached to each. A single query for a competitor's name costs pennies and hides at the bottom of the report; rolled up across every query that mentions that competitor, the same word might show hundreds of pounds of spend and no conversions. The theme is invisible query by query and obvious word by word.
You do not need a tool to start. Download the search terms report to a spreadsheet, split the query column into words, and build a pivot that sums cost and conversions by word. Sort that by cost with zero conversions and the wasteful themes rise to the top on their own: the job-seeker words, the free-and-DIY words, the wrong-product words, the geographies you do not serve. Each high-cost zero-conversion word is a candidate thematic negative, and because it is a theme it will also work the hidden portion of the report the same way. This is the technique that turns "there are ten thousand queries, I cannot read them all" into a half-hour job.
Filtering for spend that returned nothing
Before the n-gram pass, one filter does most of the work on its own: everything with real spend and zero conversions over a meaningful window. Set the date range longer than your typical conversion lag so you are not punishing terms that were about to convert, then filter to conversions equals zero and sort the survivors by cost. This is the provable-waste view — terms that took money and returned nothing — and it deliberately ignores the terms that converted, however oddly they are worded, because a term that pays for itself is not your problem.
The judgement call is how much no-return spend is enough. A single click is never a pattern; a term that has spent something approaching your target cost per acquisition with no conversion is a real signal. The exception is intent that can never convert regardless of volume — the jobs and salary and free-tier searches for a product that is none of those things — which you block on sight without waiting for the spend to prove it. Everything you decide to block here should still be checked against your active keywords first, and placed at the right level, exactly as the Performance Max negatives and Search workflows both require.
A routine that catches the drip, not just the spike
The fix for structural blindness is two passes on different cadences, not one pass done harder. Run the quick top-of-report pass often — weekly for serious spenders, monthly for smaller accounts — to catch new high-cost offenders before they run up a bill. That pass is the cost sort everyone already does, and it is worth doing; it just is not sufficient on its own.
Then run the deeper pass monthly to quarterly: download the full report, run the n-gram rollup, and work the zero-conversion themes into shared negative lists. This is the pass that catches the low-cost drip and, through thematic negatives, the hidden layer underneath it. Keep the same written log the rest of your negatives use — what you excluded, at what level, and why, with a date — because the deep pass adds broad thematic negatives that touch a lot of traffic at once, and if volume falls the log is the only way to find and reverse the one that went too far. Waste that hides is still waste; it just needs a method that does not rely on it being expensive enough to notice.