40 ChatGPT Prompts That Read Your Google Ads Account

Most published Google Ads prompt lists assume the AI cannot see your account, so they optimise for pasting — and the results read like a consultant who never opened the account. Every prompt below assumes the opposite: a connector is attached, the model reads and runs the live account, and every write is gated by your approval. That single assumption is why these return findings rather than frameworks. Prompts marked (write) will trigger an approval step; everything else is read-only.

Substitute your own account ID for 123-456-7890 throughout, and adjust date ranges to your volume — small accounts need longer windows before patterns are real.

Why connected prompts beat copy-paste prompts

A pasted CSV is a snapshot with a paste limit. A connected prompt runs against the account itself: nothing truncated, nothing stale, and every follow-up question free. The craft shifts from choosing what to paste to specifying scope, window and evidence — which is what the patterns at the end of this page teach. If you have not connected a client yet, the how-to-use-ChatGPT walkthrough covers it in five minutes.

The forty are organised the way work actually arrives: audits when something feels off, terms and negatives as standing hygiene, reporting on a cadence, copy when refreshing, builds when expanding — and patterns last, because they upgrade everything above them. Every prompt is written to be edited: swap in your thresholds, your service area, your naming conventions, because specificity is what converts a template into a finding.

Audit and wasted-spend prompts

The highest-leverage opening moves — each reads wide and returns a ranked list, which is the shape a working session wants to start from.

1. "Audit account 123-456-7890 top to bottom: settings, structure, keywords, ads, budgets. List the ten most consequential problems in descending order of monthly cost, each with the evidence and the fix."
2. "Find all spend in the last 60 days with zero conversions — by search term, keyword and placement. Group into: confidently waste, possibly lagged, needs a human call. Explain each grouping." — the grouping is the point: raw zero-conversion lists overstate waste by ignoring lag.
3. "Which campaigns have search partners or Display Network expansion enabled? Show performance for those surfaces separately from core search results."
4. "Check every campaign's location options and geo targets. Flag anything reaching people outside the regions we can actually serve, with spend attached."
5. "List ad groups with keywords but no enabled responsive search ad, and ad groups where every ad has been unchanged for over 90 days."
6. "Read change history for the last 30 days and summarise what changed, who or what changed it, and which changes coincide with performance shifts."
7. "Which auto-applied recommendations have been applied to this account in the last 90 days? What did each one change?"
8. "Compare my keyword list against the search terms actually matched last month. Where has matching drifted furthest from intent?"

Run prompt 1 quarterly and prompt 2 monthly; the rest earn their place during a full AI audit session.

Search terms and negative keyword prompts

Standing hygiene. These few minutes a week are the difference between matching drift costing you pocket change and costing you a campaign.

9. "Pull search terms for the last 30 days, all search campaigns. Rank the worst by spend with zero conversions, grouped by theme, and propose negatives with match type and level for each theme." — the per-theme grouping turns forty rows into four decisions.
10. "Run a 1-, 2- and 3-word n-gram analysis on the last 90 days of search terms. Which fragments aggregate meaningful spend without conversions?"
11. "Which search terms convert well but are not yet keywords anywhere in the account? Propose where each should live, at which match type."
12. "Check all negative keywords across campaigns and shared lists for conflicts: negatives that block keywords we bid on, and duplicates at different levels." — run it after any bulk negative import; conflicts are how good traffic dies quietly.
13. (write) "Add these as phrase-match negatives at campaign level to campaign X: [list]. Show me the full list before applying."
14. "Which queries are being matched by more than one campaign? Tell me which campaign served each and whether the cheaper one lost."
15. "Build me a proposed shared negative list for obvious non-buyer intent — jobs, free, DIY, tutorials — based on what has actually appeared in my search terms, not a generic template."

The judgement layer for turning fragments into negatives — and the levels to add them at — is covered in the negative keyword automation guide.

Reporting and pacing prompts

The readouts that replace exports. Fix the definitions once — window, conversion actions that count, materiality thresholds — and these become the weekly rhythm.

16. "Compare last week to the week before: spend, conversions, cost per conversion, by campaign. Flag big movers and give the most likely cause for each, citing data." — the causes are the deliverable; the table is just the evidence.
17. "Write a client-ready monthly summary for account 123-456-7890: results versus last month, the three causes behind the movement, and the three actions for next month. Plain language, no jargon." — name the audience and the model writes for them; name nothing and it writes for another marketer.
18. "Month-to-date spend by campaign, current 7-day run rate, and projected month-end at that rate versus a monthly plan of [number]. Which campaigns land over or under?"
19. "Show conversion lag: of last month's conversions, what share were reported within 1, 3, 7 and 14 days of the click? Use it to say how much of this week's dip is probably just lag." — run this once and recent-week dips stop looking like emergencies.
20. "Which campaigns are limited by budget, and what did that cost in lost impression share last week?"
21. "Build a table of the last 8 weeks: spend, conversions, CPA, impression share. Mark the weeks where anything changed by a lot, and annotate each with what changed in the account that week."
22. "Sanity-check yesterday's numbers against the same weekday last week and the weekday average this month. Anything anomalous enough to investigate now?"

RSA and ad copy prompts

Copy prompts work best after a read prompt has loaded the account's real language — draft from queries, not from the brand deck.

23. "Read the top converting search terms for ad group X over 90 days, then draft 15 headlines and 4 descriptions in that language. Vary structure: keyword, proof, offer, question." — the reading step is what separates this from every prompt list written for unconnected models.
24. "Which of my RSAs have the weakest asset performance labels, and what do the strong assets in this account have in common?"
25. "Rewrite the ads in ad group Y to reflect the landing page's actual claims. List where current ad copy promises something the page does not deliver."
26. (write) "Create this RSA in ad group X, paused, with headline 1 pinned to position 1 for compliance: [content]. I will enable it after review."
27. "Compare my ad copy against the queries that triggered it last month. Where is the biggest mismatch between what people searched and what the ad says?"
28. "Draft three headline variants under 30 characters that include the phrase [keyword] naturally, and flag any that risk policy problems."
29. "Which ad groups show a Google Ads ad strength below Good, and for each, which input — headline variety, keyword coverage, description length — is the limiting one?"

Campaign build and restructure prompts

The write-heavy set — this is where the agent builds campaigns end to end. Every one of these surfaces an approval before anything lands, and anything new can arrive paused.

30. (write) "Draft a search campaign for [product]: campaign settings, two ad groups by intent, keywords with match types, negatives, one RSA each. Show me the full plan first; create everything paused." — plan-first keeps the approval legible: you approve a structure you have already read.
31. "Propose a restructure of campaign Z: which ad groups to merge, which keywords to move, what the new structure looks like, and what it fixes. Do not change anything yet." — restructures are decisions worth sleeping on, which is why this one reads and proposes but does not touch.
32. (write) "Create an ad group [name] in campaign Z with these keywords at phrase match, and mirror the negatives from ad group W."
33. "Which of my campaigns overlap in targeting enough that they should be merged, and what would a merged structure preserve from each?"
34. (write) "Set up geo targeting for campaign Z: target [regions], exclude [regions], and show me the before and after."
35. (write) "Apply an ad schedule to campaign Z serving [days/hours], and tell me what share of last month's conversions occurred outside that window so I know what I am giving up."

Prompt patterns: scoping, date ranges, and asking for evidence

Five habits, five upgrades to every prompt above — these separate findings from filler:

36. Scope first. Open with the account or campaign: "For account 123-456-7890…". Ambiguity in, ambiguity out — and in multi-account setups, scoping is what keeps the right client's data in the answer.
37. Date range always. "Last 30 days" versus "this month" versus "last month" produce different tables. Name the window; on small accounts, widen it before trusting patterns.
38. Demand evidence. Append "citing the data you used" or "show the rows behind each claim". It converts the model from commentator to analyst, and makes verification a glance instead of a hunt.
39. Set thresholds. "Meaningful spend" means nothing; "spend above one day's average budget" means something. Give the model your materiality line and it will stop reporting noise.
40. Separate reading from writing. Ask for the analysis, review it, then issue the write as its own instruction. Two steps cost seconds and keep every approval prompt legible — a practice covered in depth in the safe prompting guidelines.

Where to start

Forty prompts are a library, not a curriculum. A first week that works: run 1 to see the account the way the model sees it, then 9 and 13 to bank the fastest win, then 16 on Monday as the standing readout, then 18 near month-end for pacing. Add the copy and build prompts once the hygiene loop feels routine, and fold the five patterns into everything from day one — they cost nothing and compound immediately. One temperament note: when a prompt returns something surprising, resist acting inside the same breath — ask the follow-up that would prove it, then act. Surprise plus verification is insight; surprise alone is just novelty.

The deeper point transfers even if none of the wording does: scope, window, evidence, threshold, and a clean split between asking and acting. Prompts age as the platform moves; the pattern behind all forty does not.

Start a free pilot and run prompt 1 against your own account.