Open your Google Ads account and pull the search terms report. Somewhere in that list is a query that has spent more than any keyword you chose, and it has never produced a single conversion. The search term report analysis is the closest thing paid search has to an audit trail, and every search that triggered your ads shows up in it with the cost attached.
Most accounts look at this report once a quarter, or when someone asks why conversions dropped. The queries people actually type have drifted a long way from the keywords you chose. Black Propeller, a paid search agency, reports advertisers now see three to five times as many unique search queries as they did two years ago, and their audits found negative keyword hygiene to be the fastest budget reclaim lever in over half the accounts they review. PPC Chief, another agency, estimates the average account wastes 20 to 40 percent of its monthly budget on clicks that never convert.
Those numbers belong to the people who wrote them. Your own report gives a better figure for your own account. The numbers come from your data, and Gemini handles the part that normally eats an afternoon, the sorting.
Why search term report analysis catches budget leaks first
A search term is the phrase a person actually typed. A keyword is the phrase you bid on. Google's search terms report documentation draws that line explicitly, and it explains how close variants, misspellings and broad match bring in queries you never wrote down.
The report is also deliberately incomplete. It only shows searches that a significant number of people used, and low-volume queries get folded together or hidden. The terms you can see are the ones that spent enough to surface, which makes them the ones worth acting on.
North Country Growth, a PPC consultancy, applies a simple rule of thumb. Any term that has spent more than one to two times your target cost per acquisition with zero conversions is a candidate for negation. The same report shows the opposite opportunity, because queries that convert without being keywords yet are the cheapest wins in the account.
We covered the retargeting side of the same waste problem in our guide to AI retargeting that skips the wrong people.
What you need to build the workflow
- A Google Ads account with a Search campaign that has run for a few weeks, plus access to edit keywords, which means an advertiser or admin role. The report needs data before it is useful.
- Gemini. The free Gemini app handles a few hundred terms easily. For bigger exports the Gemini API with structured output works better, and the Gemini API text generation docs cover both.
- A spreadsheet to hold the export and the classification.
- Fifteen minutes for the first pass, then thirty minutes a week.
Step 1: Export the search terms report from Google Ads
In Google Ads, go to Insights and reports, then Search terms. Google's walkthrough for adding negatives from the report starts the same way. Set the date range to 90 days, or 30 if you see thousands of terms a month. Add cost, conversions and impressions to the columns before you export, then sort by cost, highest first.
Export the CSV, and drop it into Google Sheets if you have more than a few hundred rows so Gemini can read the columns cleanly. You will come back to this report in Step 4 to apply the changes.
Step 2: Build the classification prompt in Gemini
The prompt does the work that normally eats an afternoon. You are going to ask Gemini to read every search term and sort it into one of three buckets, keep, add or negate, with a reason for each call.
Give Gemini the context it needs
Gemini classifies well when it knows what it is classifying. Before the terms, tell it what the business sells, who the customer is, what a qualified lead looks like, your target cost per acquisition, and the brand terms. In the Gemini API this goes in the system instruction, which the text generation docs describe as the way to configure model behaviour. In the app it is simply the first part of your message.
The classification prompt
You are a Google Ads analyst reviewing a search terms report. Business: [what you sell, one line] Customer: [who buys it, one line] Qualified lead: [what makes a lead worth pursuing] Target CPA: [the number from your account] Brand terms: [your brand and product names] For each search term, classify it as one of: KEEP - the term matches what we sell and a keyword already covers it ADD - commercial intent, close to what we sell, not yet a keyword NEGATE - irrelevant, informational, freebie-seeking, job-related or a competitor Return a table with these columns: search term, classification, reason, suggested keyword or negative keyword. Be conservative with ADD. When in doubt, choose NEGATE.
That is the whole prompt. The context lines stop the model from guessing, and the conservative instruction keeps the negatives safe by default.
Run it
Paste your terms under the prompt in the Gemini app and it will return the table. For a few hundred terms that works well. For thousands, use the Gemini API with structured output so you get JSON or a clean CSV back instead of prose. gemini-3.6-flash is the current default model on the text generation endpoint. WE-DO's AI search term mining workflow follows the same pattern, exporting 90 days of data, classifying with a prompt, reviewing, then implementing.
Step 3: Sort the terms into keep, add and negate
Gemini will hand you a table, and your job is to check it. The model can be wrong in specific ways, missing your market's slang, missing a competitor you care about, or flagging terms that already have keywords against them.
KeywordMe's three-bucket framework is the same one the prompt uses. Add as negative when the intent is wrong, promote to keyword when a query converts without being targeted, and monitor the borderline cases.
Here is what the classification looks like for a fictional account selling project management software to small teams.
| Example search term | Bucket | Why it lands there |
|---|---|---|
| project management software for small business | Keep | Matches what you sell and a keyword already covers it |
| best project management tools for agencies | Add | Commercial intent, no keyword yet |
| free project management software | Negate | Freebie seeker, spends without converting |
| project management software jobs | Negate | Job search intent, wrong audience |
| how to plan a project timeline | Negate | Informational query, rarely converts |
| monday.com pricing vs asana | Negate | Competitor comparison, expensive to win |
Your terms will look different. The classification logic is what transfers.
Keep a human in the loop
Every recommendation goes to a marketer, and only approved changes reach Google Ads. That is the pattern pi-automate uses in their search term manager. A wrong negative hides your ads from searches that might have converted, and Gemini occasionally misreads intent, so the human check is where the mistakes get caught.
Step 4: Apply negative keywords and add high-intent terms
Back in Google Ads, open the search terms report again. Tick the terms you are negating and use Add as negative keyword. Google's guide covers where they can go, ad group, campaign or shared list.
Negative exact match is precise but narrow. For themes, build shared negative lists covering free, cheap, jobs, how to, reviews and competitor names. KeywordMe's wasted spend guide recommends phrase or exact negatives over broad, because negative broad blocks more than you intend. Google's negative keyword documentation warns that over-negating can cut reach, so treat the shared list as the control point and only add terms that have actually spent.
Then add the promoted terms as keywords. Put each one in the ad group that matches its intent, pick a match type deliberately, and let the campaign learn from the traffic.
How to know it is working
Your own numbers decide this, and you already have a baseline in the cost column of the report you exported. Run the workflow weekly for a month, then analyse what the negated terms cost compared with the month before. It should be close to zero while conversion volume holds or improves.
Groas AI's guide to Performance Max negatives documents a worked example. An ecommerce advertiser committed to a weekly 30-minute review, added 1,143 negative keywords over 90 days, and cut their Performance Max CPA by 31 percent while holding conversion volume.
Do not expect a one-off jump. The report changes every week because matching changes every week, which is why the routine matters more than any single pass. Knowing where spend lands is the other half of the job, and our marketing budget attribution pipeline post covers it.
Turn search term mining into a weekly routine
Book thirty minutes, same day every week. Open the report, paste into Gemini, review the table, apply the changes. That is the whole routine, and it compounds as the negatives get more specific and the promoted keywords get cheaper traffic.
The plumbing can automate itself. n8n has a template that pulls the search terms from Google Ads, sends them to Gemini for classification, and posts the result to Slack with approve buttons. The weekly pass becomes a few clicks, with the reasoning staying in Gemini and the approval staying with you.
If you want to see the surrounding workflows, our guide to five n8n templates worth starting with is a good place to look.
This is the kind of workflow we build at Supernodes in a two-week pilot. The report exports itself, Gemini classifies it, and negatives land in a shared list once you approve them. If your Monday looks like a stack of reports that should have run themselves, that is the conversation to have.
Frequently asked questions
Why does my search terms report show queries that have nothing to do with my keywords?
Close variants and broad match bring them in. A search term is what the person typed, while a keyword is what you bid on, and Google matches them loosely on purpose. Google's documentation explains the difference, and the report needs its own review instead of trusting your keyword list.
Should negative keywords go at campaign or account level?
Shared negative lists are the better default for themes like free, jobs or competitor names, because one list applies everywhere and you can review it in one place. Negatives added from the report default to negative exact match. Keep account-wide themes in a shared list, and campaign-specific exclusions in the campaign.
How often should I run search term report analysis?
Weekly is the working cadence for accounts spending real money. North Country Growth's guide recommends weekly attention, KeywordMe's says the same, and the Groas AI example above ran a 30-minute weekly review for 90 days. New accounts or low spenders can start monthly and tighten as volume grows.
Can too many negative keywords hurt my account?
Yes. Google's negative keyword documentation warns that too many negatives mean fewer customers see your ads, and practitioners covered by Search Engine Land warn that aggressive negative lists hurt more than they help in accounts where Smart Bidding has enough signal. Keep the list themed and evidence-based, and remove negatives that stop spending.
Why does my search terms report hide some queries?
Privacy thresholds and volume cutoffs. Google only shows search terms that a significant number of people used, and low-volume or sensitive queries get folded together or omitted. The terms you can see are the ones with spend behind them, which is what makes them worth acting on.