“Pause the keywords that spent over $100 with no conversions” is a useful review queue and an incomplete authorization to stop spend. I chose this instruction from groas’s AI Assistant material because it neatly exposes the gap between asking ChatGPT for recommendations and letting it change an account.

It sounds decisive. It reads like something a busy owner can paste into ChatGPT and be done with. But the $100 figure says nothing about which account the assistant can touch, which dates count, which conversions matter, or how long to wait before calling a click a failure.

The threshold identifies candidates. It does not settle the decision. I’m taking the sentence apart in the order it arrives: the action, the keywords, the spend, and the supposed absence of conversions. Then I’ll rebuild it into a request a decision maker can defend.

“Pause”: the sentence starts with a decision it hasn’t earned

Pause stops spend. It also stops new search-term data and whatever the keyword might reveal next week about intent. That can be the right trade. The original instruction simply assumes it is.

What the deck calls “optimization,” I sometimes call paying $100 for market research and then refusing to read the notes.

Before authorizing a pause, I want an answer to a narrower question: does the evidence support stopping this keyword, or does it support investigating it? A keyword with scattered, irrelevant queries presents a different problem from one with useful queries and a landing page that failed for a day. A negative keyword, bid cut, or structural change may address the cause more directly.

I require a reason for each proposed pause, not one batch reason for ten keywords. The explanation should cover spend, clicks, the mature reporting period, counted conversion actions, search terms, and contrary evidence. “Over $100, zero conversions” repeats the filter. It does not explain the decision.

That distinction matters once ChatGPT has execution access through groas. The useful pattern is evidence first, bounded approval second: inspect the proposed changes before they go live, then retain a log and a way back.

My verdict on the first word: pause belongs in the proposal column, not at the start of an unconditional instruction. If the assistant cannot defend the action for each keyword, I leave it in the review queue.

“The keywords”: name the account and the exact entities

“The keywords” sounds specific until an assistant has to act on it. Which ad account? Which campaigns? Which match types? Enabled keywords only, or everything in the report?

An instruction that feels obvious to the advertiser can still be ambiguous to the system holding editing access. I make the assistant identify scope before it evaluates performance: account ID, campaign, ad group, keyword ID, match type, and current status. Then I confirm that it is looking in the right place.

There is a second trap inside those two words. Keyword spend is not search-term intent. A broad match keyword for emergency plumber might spend $140 across 60 stray queries with zero conversions. Two of those queries might still deserve their own ad group and landing page. Pausing the keyword stops further traffic from that keyword; it does not resolve which queries were worth pursuing.

So my evidence pass includes both levels:

  • Keyword performance: cost, clicks, counted conversions, and current status.
  • Search-term evidence: the queries behind that spend, including anything that argues against a pause.
  • Structure: the campaign, ad group, and landing page needed to interpret the result.

Thin, scattered queries go into a review batch for negatives and structure rather than automatically joining the pause batch. The object of the instruction must be an exact entity, but the evidence cannot stop at the entity’s summary row.

“Spent over $100”: a threshold needs dates, currency, and economics

$100 with no date range is a number without a denominator. Last 7 days, last 30 days, and lifetime describe three different problems. I don’t leave the assistant to choose among them.

I specify a completed reporting period and the account currency. Then I put the spend beside the economics of the account. $100 is an illustrative review threshold, not a universal loss limit.

Say I’m reviewing an account spending $20k a month with a $50 target CPA. In that example, $100 with zero conversions is 2× target CPA, a point practitioners flag for reviewing direct-response keywords. It is not proof that the keyword cannot work.

Click volume changes the interpretation. For a keyword that normally converts at 3%, zero conversions after 30 clicks still happens about 40% of the time through variance alone. After 100 clicks, that chance falls below 5%.

That is why a stricter version of the rule pairs 2× target CPA with 100 or more clicks. Even then, context matters. Guidance may stretch the spend threshold to 4–5× CPA for high-ticket offers or long sales cycles, where upper-funnel value complicates a direct-response verdict.

I would not turn those examples into another universal rule. Their purpose is to show what the original sentence leaves out: the same $100 can represent very different amounts of evidence.

My prompt therefore asks for four things together:

  1. A defined reporting period.
  2. Cost in the account currency.
  3. Spend as a multiple of the agreed target CPA.
  4. Clicks beside cost, with thin-data candidates separated from executable proposals.

The dollar tripwire earns a look. It does not earn a pause by itself.

“No conversions”: define the outcome before counting failures

Name the conversion actions and inspect the tracking evidence

“No conversions” is the vaguest clause in the sentence. A zero means little until I know what the report is counting.

Google Ads distinguishes between Primary and Secondary conversion actions. Before treating a column as the verdict, I want the assistant to name the actions being counted, their sources, and when each last recorded a conversion. For this review, I explicitly identify the Primary actions I intend to use.

No conversion definition, no pause. A thank-you page that stopped firing after a site migration can make broken tracking look like broken keywords. The assistant needs to inspect the available tracking evidence rather than quietly treating every zero as a performance finding.

Connecting ChatGPT to the actual account records gives it evidence to inspect. Without those records, it is guessing about the account. With them, it still needs to explain what the records establish.

Separate mature clicks from clicks still waiting for an outcome

Fresh spend can make conversion performance look worse than it is. Cost and conversion reporting do not necessarily arrive together: Google describes different data-freshness windows for conversion reporting and imported goals. Beyond reporting delays, a customer may convert days after the click.

A keyword that spent $112 over the weekend with zero conversions on Monday can look different by Thursday if two demo requests arrive from Saturday’s clicks. The Monday zero was real as a reporting observation. It was incomplete as a decision.

I require an explicit maturity window, not an assumption that recent clicks have failed. Excluding the last 3–7 days can be a starting allowance, but it does not settle the question for a long sales cycle. The assistant must state which clicks it treated as mature and separate the recent spend.

The mechanism matters: timing produces the apparent zero, and the apparent zero triggers the pause. A delay allowance breaks that chain before it stops useful spend.

State what the connected records cannot establish

Account records can show cost, clicks, keyword IDs, and recorded conversions. They cannot establish offline closes that never returned to Google Ads, order margin that is not present, or whether sales followed up promptly.

If the CRM holds closed-won outcomes while Google Ads holds form fills, the account view is not the whole commercial story. A keyword with zero recorded conversions may still have produced pipeline that the available records do not reveal.

I make the assistant list both what it checked and what it could not check. Missing evidence is not automatically a reason to keep spending. It is a reason not to present an incomplete finding as a settled one.

My verdict on this clause: name the actions, name the mature period, and name the blind spots. Only then does “no conversions” become usable evidence.

Rebuild the prompt: evidence first, exact-ID approval second

The replacement needs to carry the safeguards, not merely mention them in the surrounding article. A prompt that says “be careful” and then supplies an ambiguous filter has not improved much.

Here is the version I would use in ChatGPT with groas connected. I fill the brackets before running it. This request authorizes analysis and proposals only, not account changes.

Review Google Ads account [account ID / name] through groas MCP.

Scope:
- Campaigns/ad groups: [included scope and exclusions].
- Account currency: [currency].
- Review threshold: [amount in account currency].
- Target CPA: [amount in account currency].
- Minimum click count for PAUSE consideration: [click floor].
- Mature reporting period: [YYYY-MM-DD to YYYY-MM-DD].
- Separate immature period: [YYYY-MM-DD to YYYY-MM-DD].
- Conversion actions to count: [named Primary actions].

1. Confirm the account, currency, and scope. List the counted
   conversion actions, their sources, and their last recorded
   conversion dates. Flag tracking concerns and any requested
   evidence you cannot access.

2. For enabled keywords above the review threshold in the mature
   period, return a table with campaign, ad group, keyword ID,
   match type, current status, cost, clicks, counted conversions,
   and cost as a multiple of target CPA.

3. Include the top 5 search terms for each candidate and available
   landing-page evidence. Identify contrary evidence and explain
   what the table cannot establish.

4. Report immature-period spend separately, including its share
   of spend across the two periods. State the delay allowance used
   and flag whether the available evidence supports treating the
   selected review period as mature.

5. Separate PAUSE proposals from REVIEW candidates. Give each
   keyword a one-sentence reason and its contrary evidence.
   Keep thin-data and unresolved tracking or delay cases in REVIEW.
   Identify alternatives such as negatives, bid cuts, or structural
   changes where the evidence supports them.

6. Make no changes. Wait for my explicit approval of exact keyword
   IDs and actions. Do not infer approval from this review request.

The sequence matters. First I establish what is being measured. Then I ask for the evidence table. Only after that do I ask the assistant to propose an intervention.

The threshold, target CPA, and click floor serve different jobs. The threshold finds candidates; the CPA puts their cost in context; the click floor keeps thin evidence visible. None replaces the maturity check or the per-keyword explanation.

The last instruction is the execution boundary: no changes until I approve exact IDs and actions. The table is still a review queue. My subsequent approval authorizes a defined batch.

I keep autopilot off unless I separately enable it under explicit rules. The same separation applies to the Cursor rules pattern: proposals first, changes only after review. Connecting an assistant is not blanket permission to improvise a spending policy.

After approval: verify the change and retain the reason

Once I approve the exact IDs, I ask ChatGPT to execute only that batch through groas. Then I ask it to re-read each keyword’s status and return the corresponding change-log entry with its reason.

A screenshot alone does not capture the decision. The defensible record is IDs, before-and-after status, reasons, and a way back. groas logs changes with their reasons and supports undoing a change in one step. That lets a boss, client, or colleague reconstruct what happened without relying on someone’s memory.

I keep the record compact, with five groups of fields:

  • Decision context: date, account ID, mature reporting period, and counted conversion actions.
  • Evidence: keyword ID, campaign, ad group, cost, and clicks at decision time.
  • Judgment: the one-sentence reason and any contrary evidence I considered.
  • Execution: verified resulting status and the change-log entry.
  • Follow-up: a revisit date for REVIEW candidates left running.

A review label is not permission to leave an uncertain keyword spending indefinitely. The revisit date makes the unresolved decision visible.

And undo remains a seatbelt, not a budget control. Undo is not a spending policy: scope, thresholds, the reporting period, and explicit approval govern the action. The log records what happened. It does not make the original judgment sound.

Verdict: keep the dollar tripwire, reject the missing assumptions

The original instruction offers a useful starting filter: “Pause the keywords that spent over $100 with no conversions.” I would keep the dollar tripwire because it forces a look. I would reject its leap from a filtered report to an approved change.

The defensible version identifies the account, defines the conversions, fixes the mature period, and shows cost alongside CPA, clicks, and search terms. Then it separates pause proposals from unresolved reviews and waits for exact-ID approval.

Scope, period, conversion definition, delay, and the choice to pause still belong to the advertiser. The assistant can assemble the evidence and propose the batch. It should not silently supply those missing decisions.

That is what I would keep from this artifact: its appetite for action, with a better boundary around the action. ChatGPT reads the connected records through groas, returns the evidence, and changes only the approved batch, with reasons logged.

Let the machine do the clicking. Keep the judgment about what counts as failure for myself.