AI can sort thousands of search terms by what the searcher wanted in an afternoon. It can't tell you which exclusion will quietly cost you customers, and that decision is the part that stays with you.
Sort search terms by what the searcher wanted before you look at what the click cost. Let AI do the sorting and propose negatives, let the spreadsheet do the arithmetic, and let a named person approve each exclusion. Block the narrowest thing that stops the waste, check it against converting rows first, and put a review date on every negative you add.
Say you’re the marketing manager at a 30-person company that sells home treadmills, bikes and rowers online, and you run the Google Ads account yourself. Search spend is about $9,000 a month. On Tuesday you export the last 90 days of search terms and get 2,400 rows, plus a spreadsheet tab called “Negatives v3” that you started in the spring and stopped keeping up in June.
What you’ve been doing is sorting by cost and reading the top 50. You find a few obvious ones, like “treadmill belt replacement” and “free treadmill giveaway,” and you add them. The other 2,350 rows stay unread, and somewhere in them sit the terms that are costing you quietly: fifteen clicks at $1.20 here, forty clicks at $0.90 there, none of them large enough to reach the top 50.
Here’s the idea this guide is built on. AI is good at reading a few thousand short phrases and saying what each person probably wanted. It’s not good at knowing whether blocking a phrase will cost you a customer you can’t see. So let it sort and propose, let your spreadsheet do the sums, and let one named person approve each exclusion before it goes into the account.
The rest of this page shows what that looks like in practice: five intent buckets with your own rows in them, a prompt to paste, the match-type choices that stop one negative from doing too much, a monthly routine and a log. If your export has under a couple of hundred rows a month, skip the AI and read it yourself over coffee. The approach pays off once the list is longer than you’ll realistically read.
Sorting by cost feels sensible, and it can mislead you. A term that burned $60 and a term that burned $6 can be the same kind of search, and a $60 term that converted isn’t waste at all. The better first question is what the person typing was trying to do, because that decides what you do with the row.
Here are the five buckets, filled in with real-looking rows from the treadmill retailer’s export.
| Bucket | What the searcher wants | Example rows | What happens to the row | Who decides |
|---|---|---|---|---|
| 1. Buyer | To buy what you sell, now or soon | “folding treadmill for small apartment”; “treadmill with incline 2026” | Keep. If it converts and isn’t already a keyword, add it as an exact match keyword. | Paid search specialist |
| 2. Researcher | To compare or learn before buying | “treadmill vs elliptical for bad knees”; “how long does a treadmill last” | Leave alone for now. Lower the bid or move to a content-led campaign after 60 days of no conversions. | Paid search specialist, with the marketing manager on budget |
| 3. Wrong product or job | Something you don’t sell or a service you don’t offer | “treadmill belt replacement”; “treadmill repair near me” | Negative keyword, phrase match, written narrowly. | Paid search specialist |
| 4. Wrong person or intent | No intent to pay you | “free treadmill giveaway”; “treadmill jobs”; “used treadmill craigslist” | Negative keyword after the collateral check in the next section. | Marketing manager signs off |
| 5. Can’t tell | The phrase is ambiguous | “tread 3000”; “nordic 1750” | Type it into Google yourself and look at what comes back. Then move it to 1 to 4. | Whoever opens the search results |
Illustrative rows for a fictional online fitness retailer. Your buckets stay the same, but the examples, and what counts as “wrong” for your range, are yours to define.
Group search terms by what the searcher wanted before you look at what the click cost.
Two things about this sheet are worth knowing before you trust it. First, Google’s help page describes the search terms list as terms that a significant number of people have used, and says some low-volume queries are left out to protect privacy.[1] So you’re sorting the visible part of the picture, and some of your spend sits in rows you’ll never see. Second, the five buckets aren’t equally certain. Bucket 3 is usually a factual call about your range. Bucket 4 is a judgment call about what a customer looks like for you, which brings me to who does what.
Different roles in the account feel this differently, and the bucket definitions are where they need to agree:
The decision rule I’d use here is short. Meaning goes to the model, arithmetic goes to the spreadsheet, and permission goes to a person. A chat assistant is strong at judging what a short phrase means in context. It’s weaker at adding up a column of 400 costs and reporting a total with confidence, and it can’t know your margin, your stock or your service area. Each job goes to whoever is best at it.
In practice that means you export the search terms with the columns the model needs to judge and to quote back: the term, the campaign, the match type that triggered it, cost, clicks, conversions and conversion value. Paste or upload it in batches of a few hundred rows, which also makes each answer easier to spot-check. Before you paste anything, skim for personal details, because people occasionally type names or emails into a search box, and if you’re an agency, strip client names you’re not allowed to share. Our guide to using AI at work without leaking confidential company data covers what’s safe to put into which tool, and the one on analyzing a spreadsheet with AI shows how to do the number-crunching side properly.
Here is the prompt, filled in for the treadmill retailer.
| I run Google Ads Search campaigns for an online retailer that sells new home treadmills, exercise bikes and rowing machines to US households. We don’t sell used or refurbished equipment, we don’t offer repairs or replacement parts, and we don’t hire. Below are 300 rows from our search terms export. Columns: search term, campaign, match type, cost, clicks, conversions. For every row, do these four things: 1. Put it in one bucket: Buyer, Researcher, Wrong product or job, Wrong person or intent, or Can’t tell. 2. If it’s in Wrong product or job or Wrong person or intent, propose the narrowest negative keyword that would block it, and say which match type you’d use (phrase or exact). 3. If you’re not sure what the searcher meant, use Can’t tell. Don’t guess. 4. Copy the cost and conversions for that row exactly as I gave them. Don’t add up totals. I’ll do that in the spreadsheet. Then, for every negative you proposed, list any other row in this batch that contains the same word or phrase AND has conversions above zero. Return a table with these columns: search term, cost, conversions, bucket, proposed negative, match type, conflicting rows. |
Swap in your own business rules in the first paragraph. The instruction about conflicting rows is the line that matters most, because it puts the collateral check inside the prompt.
What comes back should look something like the sample below. It’s an illustration of the shape and the standard, not output from a particular tool.
| Search term | Cost | Conv. | Bucket | Proposed negative | Conflicting rows |
|---|---|---|---|---|---|
| treadmill belt replacement | $61.20 | 0 | Wrong product or job | “belt replacement”, phrase | None in this batch |
| free shipping treadmill | $18.40 | 2 | Buyer | None. Do not add “free” as a negative | This row itself converted |
| used treadmill near me | $44.00 | 0 | Wrong person or intent | “used treadmill”, phrase, and “used treadmills”, phrase | None in this batch |
| treadmill vs elliptical for weight loss | $27.10 | 0 | Researcher | None. Review in 60 days | None |
| tread 3000 | $33.00 | 0 | Can’t tell | None until someone checks the results page | None |
Note the second row. A model that only looked at the word “free” would have proposed blocking it.
When it comes back, spot-check ten rows against the original export, including the costs, and sort the conflicting-rows column in your spreadsheet yourself. Models can miss a conflict that a plain text filter finds instantly, so the filter is the real check and the AI’s column is a head start.
Now the scenario most people learn the hard way. Say the bucket 4 pile includes “free treadmill giveaway,” “free treadmill plans” and a dozen more, and someone, a person or a tool, suggests a broad negative keyword, “free.” It looks tidy. Next month spend on those terms drops, the cost per conversion improves, and nobody notices that “free shipping treadmill” and “treadmill free delivery” have stopped showing your ads at all, because a blocked search doesn’t appear anywhere in the export. The export shows you what happened, and says nothing about what you prevented.
That’s why the match type matters more than the word. Google’s help page describes three behaviours: a broad match negative blocks a search if it contains all your negative terms, in any order. A phrase match negative blocks it if the terms appear in the same order. An exact match negative blocks only that exact search, with no extra words.[2] The same page says negative keywords don’t match close variants, so you need to add singular, plural and synonym versions yourself. It also warns that too many negative keywords can mean your ads reach fewer customers.
| The row | The tempting negative | What it would also block | The narrower version | Level |
|---|---|---|---|---|
| “free treadmill giveaway” | free (broad) | “free shipping treadmill”, “free treadmill delivery” | [free treadmill giveaway] exact, plus “giveaway” phrase | Campaign |
| “used treadmill near me” | used (broad) | Searches such as “treadmill used by physios” | “used treadmill” phrase, and “used treadmills” phrase, added separately because plurals aren’t matched | Campaign |
| “treadmill repair near me” | repair (broad) | “treadmill with 5 year repair cover” | “treadmill repair” phrase | Campaign |
| “treadmill jobs” | jobs (broad) | Almost nothing you’d want | “jobs” phrase | Shared list, all campaigns |
Illustrative. “Level” is where you add it. A single word that is wrong in every campaign can go on a shared list; anything context-dependent stays at campaign level.
Block the smallest thing that stops the waste, at the lowest level that contains it.
Before any negative goes in, ask three questions. If you can’t answer yes to all three, it doesn’t go in yet.
Scoring: three yeses, add it. Two, rewrite it narrower and re-check. One or none, leave it out and look at the row again next month.
The split between what the tool does and what a person owns should be written down before you start, particularly if the account is shared between a specialist, a manager and an agency. This is the version I’d put at the top of the log.
| What AI does | What the human still owns | How it gets checked |
|---|---|---|
| Sorts rows into the five intent buckets and flags the ones it can’t judge | What counts as “wrong” for this business: products, stock, service area, margin | The marketing manager reads the bucket definitions once a quarter and signs them |
| Proposes the narrowest negative and match type for buckets 3 and 4 | Approving each negative, and the level it’s added at | The paid search specialist runs the collateral check in the spreadsheet before adding anything |
| Drafts a one-line reason for each exclusion for the log | The reason being true, in the account owner’s own words | A second person reads ten random log lines each month |
| Summarises what changed since last month’s export | Deciding whether a change matters, and what to do about budget | Discussed in the monthly marketing review with the numbers on the screen |
| Nothing, for conversion tracking setup and bid strategy changes | Anything that changes how Google counts or buys | Specialist, with written sign-off from the account owner |
A starting split. A solo founder plays every role, so the second-person check becomes “read it again the next morning.” Keep the last row regardless.
Every negative keyword has a person’s name next to it, and so does its review date.
One more trap is timing. Google’s help page on conversion lag points out that there’s a delay between a click and a conversion, and that recent numbers can make your cost per conversion look inflated.[3] For a $1,800 treadmill, where people often compare for a week or two, the clicks from last Thursday may not have converted yet. If you export “last 30 days” and exclude anything with zero conversions, you’ll block terms that are about to pay off.
So add a cooling-off window. Only judge rows whose clicks are at least 14 days old, which is our own starting heuristic and not a Google figure. Check your own lag first: if your buyers typically convert within three days, shorten it, and if you sell something considered, lengthen it. And set a spend line, too. A reasonable first rule for a clear bucket 3 or 4 row is spend above roughly twice your target cost per acquisition with no conversions. With a $40 target, that’s a term that has cost $80 or more and never converted, using clicks older than two weeks.
A single big purge feels productive and then decays. Google has said that 15% of the searches it sees each day are entirely new,[4] which means your list of irrelevant queries is never quite finished and keeps shifting. Better to do a modest pass every month than an enormous one twice a year. Forty-five minutes is realistic once the prompt is saved.
The routine used in this guide. Step 2 is the only step AI does on its own; every other step has a person attached.
The log is the piece teams skip, and it’s what saves you in month six. Without it, nobody remembers why “repair” is blocked, a new hire removes it during a cleanup, and three weeks later your ads are showing to people whose treadmill just broke. Here’s what a filled-in log looks like.
| Negative | Match | Level | Why | Owner | Added | Review on |
|---|---|---|---|---|---|---|
| belt replacement | Phrase | Campaign: Treadmills | We don’t sell parts. 7 rows, $212 over 90 days, 0 conversions. | Paid search specialist | Oct 2 | Dec 1 |
| used treadmill | Phrase | Campaign: Treadmills | New equipment only. Checked: no converting row contains it. | Paid search specialist | Oct 2 | Dec 1 |
| used treadmills | Phrase | Campaign: Treadmills | Plural of the row above. Negatives don’t match plurals. | Paid search specialist | Oct 2 | Dec 1 |
| jobs | Phrase | Shared list: All | Wrong intent in every campaign. | Marketing manager | Oct 2 | Jan 2 |
Illustrative entries. The “Review on” column is what stops the log becoming a graveyard of old decisions.
On review day, look at the rows again. If a term you blocked would now be a buyer, for example because you launched a refurbished range, remove the negative and write why. It’s the same upkeep you’d give any process your team relies on: one named owner and a date in the diary.
The easy number to pull is how many negatives you added. That tells you the routine ran, and nothing about whether the account is better. The measurement lens we use at Future Factors has three steps: use, persistence and impact. Use is whether the routine is happening. Persistence is whether it’s still happening a few months on, without a reminder. Impact is whether the account is actually better for it.
| Step | Number to pull | Who pulls it, and when | What it tells you |
|---|---|---|---|
| Use | Was the monthly routine completed, and what share of total search spend sat in rows that were reviewed? | Paid search specialist, monthly | Whether the list is being read at all, which is the first thing to fix if you only ever read the top 50. |
| Persistence | In month three, is the log still being updated, and did the 60-day reviews happen on time? | Marketing manager, at the quarterly review | Whether it became a habit or a burst. A log with no entries after month two is the signal. |
| Impact | Share of spend landing in buckets 3 and 4, cost per conversion on the affected campaigns, and impressions on the terms you kept. Watch for an impression drop on buyer terms. | Marketing manager with the specialist, quarterly | Whether spend moved toward buyers without losing reach. These move for other reasons too, so read them as direction, not proof. |
A measurement sheet for a small paid search team. It needs nothing beyond the Google Ads account and a spreadsheet.
It helps to keep the four stages of change separate. A specialist who has been shown the prompt and the three questions has capability, which is the training part. The five buckets and the monthly routine give that capability a recurring place to be used, and that’s workflow design. The second-person log check and the review dates are reinforcement, which is what turns a good first month into something still running in month six. Adoption is what you see when that’s happened, and it doesn’t arrive on its own.
A team can start this internally with the tables on this page, and many will. Expert-led training shortens the learning curve in a few specific ways: people practise on your own export rather than a tidy example, everyone on the account works to the same bucket definitions, and the judgment calls, like what a wrong-person row means for your business, get worked through with someone who has seen them go wrong. If you’re also building out the ads themselves, our guide to writing Google Ads with AI covers the copy side, and running a campaign retrospective with AI is a good home for what you learn from the monthly pass.
Here’s a first step for this week. Export 90 days of search terms for your biggest campaign, save the prompt above with your own business rules filled in, and run one batch of 300 rows. Sort the conflicting-rows column yourself, pick the three negatives you’re most sure of, write them in a log with your name and a review date, and stop there. Three careful entries will show you where the routine needs tightening before you scale it up.
It depends on your organisation’s rules and on what’s in the export. Search terms are typed by people, so skim for names and emails, and agencies should remove client names they’re not allowed to share. If your company has an approved AI tool, use that one. Campaign-level cost and conversion numbers are commercially sensitive too, so check the policy before you upload the file, not after.
Monthly is a sensible default for most accounts, and weekly makes sense when spend is high or you’ve just launched a campaign. The frequency matters less than the pattern. Whatever you choose, put it in the diary as a recurring slot with a named person, because a cleanup that depends on someone remembering tends to stop by the third month.
We wouldn’t, at least not to the live account. A wrong negative is invisible in the export, so the failure shows up weeks later as lower volume. If you use a script or integration, make it write proposals to a pending list that a named person approves, and keep the same log. The approval step is cheap and the cost of skipping it is hard to see.
This guide is written for Search campaigns. Performance Max, Shopping and Microsoft Advertising report search data differently, and the negative keyword rules differ by platform and have changed before, so check each platform’s current help pages before applying it. The sorting idea, grouping terms by intent and approving each exclusion, carries over. The match-type details may not.
Google doesn’t publish a safe number, and its help page warns that too many can reduce the customers you reach. So count isn’t the test. The test is whether each one has a reason in the log, passed the conflict check and still has a review date. A long list of well-documented negatives is healthier than a short list nobody can explain.
The description of the search terms list, the negative keyword match types and the warning about too many negatives come from Google’s own Ads Help pages; the note on conversion lag comes from Google Ads Help as well; and the figure on new daily searches comes from a 2022 post by Google’s head of search. All four were read on Google’s pages in September 2026. The 14-day wait, the two-times-target-CPA line and the 60-day review date are Future Factors starting heuristics, not published benchmarks. Claims about how much ad spend is wasted, which circulate widely in agency and tool-vendor blogs, were left out because none could be traced to a primary source. The company, the numbers and the people in the examples are illustrative, not real cases.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.