Run an n-gram pass over your search terms and the output splits into two piles. The first is unambiguous junk — words that are waste in every account that sees them, the ones you negate without a second thought. The second is more dangerous, and most managers treat it exactly like the first: high-spend single words whose intent depends entirely on the account they show up in. "Free" is waste in one account and the highest-converting query in the next. "Template" is a freebie-hunter in a SaaS account and the money query in a business that sells templates. These words do not belong on a negative list. They belong in a review queue.
A practitioner analysis of 31 accounts, covering roughly $133M in spend, found that a small set of recurring tokens accounted for most of the wasted spend rather than list length — naming code, pdf, free, how, for, test, generator, and online among the heaviest, each burning between roughly $0.4M and $2.6M at zero conversions across the sample. The lesson that usually gets taken from that is "add those words as negatives." The better lesson is the opposite: those are precisely the words you must not add blindly, because the same token that leaks money in aggregate is carrying qualified intent in some of the accounts in that sample. This post is about building the discipline that tells the difference — a standing, per-account review queue instead of a reflex negative.
The short answer: some high-spend words are decisions, not junk
The actionable split is between words that are context-free and words that are context-bound. A context-free junk word — a competitor's brand you do not bid on, an unrelated product category, an adult or illegal modifier — means the same thing in every account, so negating it account-wide is safe and correct. A context-bound word changes meaning with the rest of the query and with the business behind the account. "How" signals a research query in one account and sits inside "how much does [product] cost" in another. The n-gram method is excellent at surfacing both kinds because it ranks by spend, but it cannot tell you which pile a word belongs in. That judgement is yours, and it is the whole job.
Treating a context-bound word as junk is the single most common way a well-intentioned cleanup destroys good traffic. It feels productive — you are adding negatives, the list is growing, the account looks tidier — and the cost is invisible because the queries you just blocked stop appearing in any report. This is the same failure mode as over-negating throttling your good traffic, seen from the token level: not too many negatives in general, but the wrong specific words promoted from "review" to "block" without the data check that would have saved them.
Why the same word is waste in one account and revenue in another
A negative keyword does not block a meaning; it blocks a string. When you negate "free" you are telling Google to suppress every search term that contains those four letters as a word, whatever else is in the query. In an account where "free" mostly arrives as "free [product] download" from people who will never pay, that is a clean win. In an account where "free" mostly arrives as "free trial [product]" or "free shipping [brand]", the identical negative suppresses the queries with the clearest purchase intent in the whole account. Same word, same match behaviour, opposite outcome — and the only thing that changed is the business and the queries its market actually types.
This is why published "universal negative keyword lists" are dangerous past the genuinely universal tier. The first few hundred entries — obscenity, unrelated verticals, clear freeloader phrases — generalise. The long tail of common words does not, and dropping someone else's 500-line list into your account imports their intent assumptions as if they were yours. The ambiguous tokens are the ones where that import does the most harm, because they are high-spend by definition: the n-gram pass flagged them precisely because a lot of money flows through them, so getting the keep-or-kill call wrong on a single token moves real budget in either direction.
The standing watchlist: tokens that earn a review, not a negative
Build the queue from your own account, not from a generic list, but the recurring suspects are consistent enough to start from. Freebie signals: free, trial, cheap, discount, coupon, deal. Research signals: how, what, why, guide, tutorial, examples, ideas. Do-it-yourself and employment signals: diy, template, generator, jobs, salary, career, course, training. Format signals: pdf, download, login, app, calculator. Each of these is a word that is waste for some advertisers and the core query for others, which is exactly what qualifies it for the queue. Add any token your own n-gram pass flags as high-spend that you cannot confidently call junk in two seconds — if it needs thought, it needs the queue.
Keep the queue small and account-specific. A dozen or so tokens that your account is genuinely sensitive to is far more useful than a sprawling list you never actually review, because the point of the queue is that you look at each entry rather than rubber-stamp it. The tokens that earn a permanent spot are the ones that keep reappearing in your high-spend terms with mixed intent; a token that resolves cleanly to junk after one review graduates to the negative list and leaves the queue. Over a few weeks the queue converges on the handful of words that are structurally ambiguous for your specific business, which is the set worth the standing attention.
How to run the review in the search terms report
Open the campaign or account, go to the search terms report, and set a window long enough that a zero in the Conversions column is evidence rather than small-sample noise — for most accounts that is 30 to 90 days, longer for low-volume ones. Filter the Search term column to "contains" your token, then add Cost and Conversions (and Conv. value if you optimise to value) to the view. You are now looking at every visible query that carried the word, with the money and the outcomes side by side. Read the pattern, not individual rows: does the word travel with qualified queries that convert, with junk queries that only spend, or with a genuine mix that needs splitting?
Two honesty caveats shape how much you can read into this. First, the report is privacy-thresholded — low-volume terms containing your token never show up, so the true spend on the word is higher than the filtered total, and a token that already looks expensive in the visible rows is worse than it appears. Second, Google now often reports an approximation of the query's intent rather than the verbatim string, so a row grouped under your word may have been triggered by a differently-worded search. Both push the same way: use the filtered view as strong directional evidence for the keep-or-kill call, not as a complete ledger. For the related question of how much spend to require before a zero counts, see how many clicks before you negate a search term.
What to do with each verdict: negate the pattern, keep, or pin it
Each token review ends in one of three actions. If the terms containing the word consistently spend with no conversions and no assisted conversions, negate — but negate the junk phrase, not the bare word. Adding "free download" as a phrase negative kills the leak while leaving "free trial [product]" alive; adding "free" on its own kills both. If the terms convert, keep the word and do nothing to it, because the only correct action on a converting token is to leave it alone. If the terms are genuinely mixed — some convert, some leak — split them: negate the losing phrases and, where the winning queries deserve their own control, pin the intent with an exact or phrase keyword in the right ad group so bidding and creative can treat it deliberately.
Remember that negatives do not match close variants the way positive keywords do, so a phrase negative on "free download" will not automatically catch "free downloads" or a misspelling — you need the plural and close-variant coverage handled explicitly. This is another reason the bare-word reflex is tempting and wrong: negating "free" feels like it covers everything, which is exactly the problem, whereas negating the specific junk phrases is more work but is the only version that does not take qualified traffic down with it. When you do negate, prioritise by the spend the term is actually leaking, the way the CPA-savings prioritisation approach sequences the list so the biggest leaks get fixed first.
Why this is a weekly queue, not a one-time cleanup
Intent drifts, so the review is never finished. A token that reviewed clean last quarter can start bleeding the moment a competitor launches a free tier, a seasonal spike pulls "jobs" or "course" searches into your net, or a broad-match or AI Max expansion starts matching the word in contexts it never reached before. Because the ambiguous tokens are high-spend by construction, that drift compounds quickly — a word that quietly flips from qualified to junk can cost a month of budget before anyone notices, since nothing in the account flags it as changed. The weekly queue exists to catch the flip while it is still a few days old.
Once the queue is built, the recurring work is small: a few minutes to re-run the filtered view for each watchlist token and confirm the verdict still holds. That is a different and more durable discipline than periodic list-building, and it is worth knowing when even this stops paying off — the point of diminishing returns on negative mining applies here too. But the ambiguous-token queue tends to keep earning its keep longer than generic mining, precisely because these are the words where the account's own reality keeps changing underneath you. The managers who hold onto their best traffic are not the ones with the longest negative lists — they are the ones who kept a short list of dangerous words under review and made every call from the account's own conversion data.