“Spent over $100 with no conversions” is a useful investigation filter and a reckless negative-keyword instruction. Do not ask Claude to turn last season’s zeros into next season’s blacklist.

I used to run that filter during seasonal cleanups. The mistake was treating the output as a decision rather than the start of one.

That $100 line looks like discipline. It bundles together thin click counts, incomplete reporting, and offers that may no longer resemble what you plan to sell. Turning that pile into negatives does not clean the account. It makes next season smaller before it starts.

Keep the filter. Make Claude build an evidence-ranked exclusion queue, not execute a spending rule.

The $100 rule finds suspects. It does not convict them.

Say you spend $20k a month. You sort last season’s search terms by cost, filter conversions to zero, and hand Claude the top 200 rows. The familiar workflow is simple:

  1. Open the search terms report.
  2. Sort by cost, highest first.
  3. Filter for zero conversions.
  4. Add negatives.

Fast, tidy, and a terrible way to prepare for a different season. A spending threshold tells you where to look, not what you found.

At a 3% conversion rate, the chance of zero conversions in 50 clicks is still about 22%. It falls to around 5% after 100 clicks. Twelve clicks and $110 spent with no sale is not a verdict. It is an expensive question.

The report is incomplete, too. Since September 2020, Google has shown only queries meeting undisclosed privacy thresholds. Queries below those thresholds roll into an aggregate “Other search terms” row that still bills you normally. The Google Ads Liaison has acknowledged cases where 80% of spend and 90% of conversions sat in “Other” on exact match.

Your complete blacklist is therefore a blacklist of the visible survivors. Bulk-negate those and you leave the invisible waste untouched while blocking terms that might convert with another week, another landing page, or another offer.

Use spend to prioritize the investigation. Use evidence to decide the exclusion. Those are different jobs.

Ask Claude to show the evidence before it proposes changes

My first instruction forbids the thing I want to avoid: changes. With Claude connected to the ad account through groas, rather than Claude working on its own, I would start here:

Using groas, review last season’s visible search terms. Return spend, clicks, conversions, conversion value, date range, and campaign context. Make no changes.

That last sentence matters. I want an evidence table, not an execution spree. Before judging a zero, I need to see which dates Claude read, which campaign the term came from, and how many clicks sit behind it.

Then I ask Claude to identify what the table cannot establish:

  • How much spend sits in “Other search terms”?
  • Can Search Terms Insights group those hidden queries by intent, without pretending to reveal each query?
  • Does the window include recent clicks whose conversions may still arrive?
  • For lead and SaaS accounts, have the relevant offline conversion uploads arrived?

Timing matters because Google reports conversions against the click date. Recent dates can show spend before all the associated conversions appear. Sorting that unfinished window by cost does not finish it.

A zero with 14 clicks from three days ago and a 9-day upload lag is a pending delivery, not an exclusion case. Make Claude flag incomplete evidence before it ranks anything for removal.

Make Claude return three queues, not one blacklist

The next instruction separates decisions that the $100 filter jams together. I want every visible term assigned to one of three queues, with the reason attached.

1. Irrelevant intent: propose an exclusion

A wine-glasses query on an eyeglasses campaign does not need another season to reveal its intentions. Clearly irrelevant intent can be blocked immediately.

The evidence here is the mismatch, not the spend. Claude should explain why the query cannot fit the offer before proposing a negative.

2. Relevant intent, weak evidence: watch it

A relevant query with too few clicks belongs in a review queue. Keep it, watch it, or reduce exposure. Do not turn uncertainty into an exclusion simply because the row looks untidy.

Relevant-but-unproven terms need more clicks relative to the normal conversion rate before the zero becomes persuasive. Make Claude state what additional evidence the term needs, rather than treating $100 as a universal stopping point.

3. Offer-dependent intent: reassess it

Queries containing “coupon,” “bundle,” or “next-day delivery” depend on what the business actually offers. Last season’s zero tells me little about next season if the coupon, bundle, or delivery promise changes.

Do not exclude a query for rejecting an offer you no longer sell. Put it in the offer-review queue. If Claude cannot confidently classify a row, it stays unresolved, not approved.

Specify the negative, not just the query

For each irrelevant term, I want Claude to propose the exact negative text, match type, and scope. “Block this” is not a complete instruction.

Negative match is literal, not expansive like positive match:

  • Exact: [free software] blocks only that query.
  • Phrase: "free software" requires those words in that order.
  • Broad: all the words must appear, but they can appear in any order.
  • Variants: negatives do not automatically cover plurals or synonyms. jobs does not block job.

That is where a lazy bulk prompt becomes expensive. A query you dislike and a phrase you never want anywhere in the account are not the same thing.

My default is exact for a seen query I want gone, phrase only for wording that is wrong in every context, broad almost never. If Claude proposes a wider match type, it has to defend the extra reach it would remove. “Zero conversions” is not that defense.

Tell Claude what next season actually means

Claude cannot infer the February offer from November’s search terms. Before it judges whether historical results transfer, I give it the brief:

  • Season dates.
  • Landing pages.
  • Prices and promotions.
  • Stock or capacity constraints.
  • Acceptable acquisition economics.

An illustrative brief might read: next season runs February 1 to March 31; the landing page is /spring-bundle; the bundle costs $149; free shipping starts at $75; target CPA is under $38; no discount exceeds 15%.

Then comes the instruction:

Reassess the exclusion queue against this brief. Explain which historical conclusions no longer transfer.

A query that struggled with a $199 single-unit page and a 5-day shipping time deserves another look under a $149 bundle with free 2-day shipping. That does not prove it will convert. It means the old result no longer answers the whole question.

Judge next season’s query against next season’s offer. Otherwise, Claude is efficiently enforcing a business you have already changed.

Competitor ads cannot substitute for that brief. Through groas, Claude can pull the ads competitors run on Google and Meta and the landing pages behind them. Useful context. Terrible evidence for a negative keyword.

A rival bidding on “cheap” does not make “cheap” bad for you. A rival avoiding “enterprise” does not make it bad either. For the mechanics behind overbroad exclusions, I would read the negative-keyword mistake that is costing you conversions before approving the queue, not after reach disappears.

A blacklist stamp crushing both irrelevant queries and potential buyers in a Google Ads search terms queue

Approve narrowly enough to recognize a mistake

Campaign, ad group, shared list, or account is not paperwork. It is the blast radius.

I want the full specification before anything gets blocked:

  1. Exact negative text: what, precisely, will be added?
  2. Match type: how much beyond the observed query could it exclude?
  3. Scope: where will the restriction apply?
  4. Reason: what makes the intent irrelevant?
  5. Possible casualty: what relevant traffic could it also block?

That last question forces the review past “this row spent money.” If Claude cannot describe the possible collateral damage, I do not approve the negative.

I start with exact match at campaign level and push shared and account lists to the end of the queue. The same distinction runs through the $100 pause rule as a review queue, not an instruction: a threshold can nominate a row without authorizing the change.

Only approved exclusions get applied through groas, where every change is logged with its reason and can be undone in one step. The log does not make a bad decision good. It makes the decision traceable when I need to reverse it.

A small exact negative with a recorded reason beats a broad phrase negative I have to debug in April. If the mistake would be invisible, the approval was too wide. The prompt-by-prompt version of the evidence review is in find wasted Google Ads spend with one prompt.

Automate the review, not the assumption

I leave autopilot off for this job. In groas, it stays off until switched on, then operates under rules you set, with changes logged and undoable. The workspace costs a flat $19 a month for unlimited teammates, ad accounts, and websites. Paying for execution does not mean every review should become automatic execution.

For a seasonal cleanup, I want Claude to build the queue. I want to approve five to fifteen exact negatives at campaign level. And I want the log to explain each one.

A monthly job that bulk-adds anything over $100 with zero conversions will eventually block a query that needed more clicks, a fixed feed, or a new bundle page. The rule cannot distinguish those cases because it never asks about them.

If I delegate exclusions later, I write permissions like instructions for a careful junior, not a slogan.

Permitted without asking:

  • Exact-match negatives for proven irrelevant intent.
  • One term at a time.
  • Campaign scope only.

Must return for approval:

  • Anything phrase or broad.
  • Anything on a shared or account list.
  • Anything with fewer than 50 to 100 clicks, judged against the normal conversion rate.
  • Anything touching price, coupon, jobs, or location words.
  • Anything where the offer changed since the click.

The narrower permission for jobs and location terms is deliberate. A word that looks irrelevant in isolation can still require context. The rule is not “Claude recognizes a suspicious word.” The rule is “the evidence establishes irrelevant intent.”

That is the difference between automating the review and automating the assumption. The first saves hours. The second saves hours and spends customers.

Keep the filter. Change the instruction.

After the evidence table and next-season brief, this is the approval prompt I would use:

Propose only exact-match negatives for proven irrelevant intent, at campaign scope, with the exact text and one-line reason per row. Rank the queue using clicks, spend, intent, and changes to the offer. List relevant-but-thin and offer-dependent terms separately as watch, not block. For thin-evidence terms, state the click threshold needed for another review. Flag unresolved rows. Make no changes.

That keeps Claude in reviewer mode and leaves approval with me. Seasonal hygiene is deciding what deserves to stay out, not proving how much I can throw away.

Ignoring that distinction creates a familiar mess. Upload 200 phrase negatives in January and the account feels clean for a week. Later, reach narrows, and someone asks why a relevant query no longer triggers the ads. The answer sits on row 147 of a sheet where nobody kept the reason.

Now the cleanup is not just removing negatives. It is reconstructing which exclusion affected which ad group, what the original concern was, and whether it still applies. The review did not disappear. It moved to a worse moment.

A narrow garden gate letting one buyer through while a bulldozer blocks the surrounding field

Keep the $100 filter as a triage list. Send ambiguous terms back for more evidence. Approve narrow exclusions with reasons attached. The advertiser who blacklists last season’s zeros starts smaller and calls it efficiency. The account gets quieter while the buyers go elsewhere.