How to read this page
Thirty-seven prompts, in two halves. The first twenty-two are one-line questions, for when you want an answer now. From number twenty-three on they are full templates, written to be pasted whole — the parts in square brackets are yours to replace, and the rest is deliberate: the role line, the data window, the output shape and the closing clause about uncertainty all change the answer you get back.
Every prompt is marked one of two ways, and the difference decides whether it is worth anything.
Needs account access means the prompt asks about your data. In a plain chat window it produces a polite guess, so connect the account first — step by step for ChatGPT, for Claude, or watch it done in two minutes on Google Ads and ChatGPT, connected. Working in Claude instead? The Claude prompt library is written for its connector, projects and artifacts.
Works in any chat means the prompt is about writing, structuring or reasoning rather than about your numbers.
Each prompt is followed by a line on what a good answer looks like. That line matters more than the prompt: it is how you tell a real answer from a fluent one.
Diagnosis: what is actually happening
1. "Show me last week's spend, clicks, conversions and cost per conversion by campaign, with the change versus the previous week." (Needs account access) A good answer is a table with both weeks side by side, not just last week's totals. Comparison is the whole point.
2. "Which of my campaigns are limited by budget right now, and how much spend am I leaving on the table?" (Needs account access) Should name specific campaigns and quantify the gap. "Some campaigns may be budget-limited" is a non-answer.
3. "What changed in this account in the last 14 days, and did each change do what it was supposed to?" (Needs account access) Should list changes with dates and the metric that moved after each. This is the prompt that catches your own mistakes.
4. "Compare my conversion rate this month against the same month last year, by campaign, and flag anything that dropped by more than 20%." (Needs account access) A good answer respects seasonality by using the year-over-year window you asked for rather than quietly substituting month-over-month.
5. "My cost per conversion went up 30% this month. Walk through the possible causes in order of how likely they are for this specific account, using the data." (Needs account access) The phrase "for this specific account" is doing the work. A good answer cites your numbers; a bad one recites a checklist.
6. "Which campaigns are spending the most and converting the least? Rank by wasted spend." (Needs account access) Should define what it counted as wasted and show the ranking, so you can disagree with the definition.
7. "Review this campaign structure and tell me what a Google Ads specialist would criticise." (Works in any chat, better connected) Adversarial framing produces far better feedback than "is this good?".
Search terms and negatives: the money leak
8. "List search terms from the last 30 days with more than €20 of spend and zero conversions." (Needs account access) A plain list with spend per term. Set the threshold to something meaningful for your budget.
9. "Group last month's search terms by intent and show spend and conversions per group." (Needs account access) Good answers produce a handful of legible groups — research, comparison, brand, competitor, irrelevant — not fifty micro-buckets.
10. "Which search terms are triggering my ads but are clearly a different product or audience? Explain why for each." (Needs account access) The explanation is what makes this reviewable. You are checking its judgement, not accepting it.
11. "Add these as negative keywords at the campaign level: [paste your reviewed list]." (Needs account access, writes) Should confirm exactly what it added and where. Never approve this from a list you have not read.
12. "Suggest 30 negative keywords for a [your business] campaign, grouped by theme, before I have any search-term data." (Works in any chat) Useful for a launch. Expect obvious themes — jobs, free, DIY, competitors, wrong geography. Naming the services you do not sell is what makes the answer specific, and the cleaning-business version shows the same prompt written that way. The auto-repair version adds the vehicles a shop turns away, which the account has no other way to know, and the landscaping version adds the crew that is already booked for the same reason. The moving-company version adds the destinations you do not drive to, which is the one thing a mover's account can never infer.
13. "Are any of my match types working against me — broad terms picking up traffic that a phrase or exact version should be catching?" (Needs account access) Should name the keyword pairs in conflict, not lecture about match types generally.
Ad copy: five things to ask
14. "Here is my landing page: [URL]. Write ad copy that promises exactly what the page delivers — no claims the page does not support." (Works in any chat) The constraint is the point. Message-match failures are a top cause of high cost per conversion.
15. "Rewrite these three headlines to lead with the outcome instead of the feature." (Works in any chat) Should change the sentence structure, not swap adjectives.
16. "My click-through rate is below 2% on this ad group. Here are the ads and the keywords — what is the mismatch?" (Needs account access) Should point at specific keyword-to-headline gaps.
17. "Read my landing page and list the objections a buyer would have that the page never answers." (Works in any chat) The best of these prompts. Every unanswered objection is a conversion you paid for and lost.
18. "What sitelinks, callouts and structured snippets should this campaign have, based on what my landing pages actually offer?" (Works in any chat) Should reference real page content, so paste the page or give the URL.
Budgets and bidding
19. "Which campaigns should get more budget based on the last 60 days, and which should get less? Show the reasoning." (Needs account access) Should account for whether a campaign was budget-limited. Reallocating away from something that never had room to prove itself is a classic error.
20. "Cut the daily budget on any campaign with a cost per conversion above [threshold], and tell me what you would change before changing it." (Needs account access, writes) "Before changing it" is the safety clause. Read the proposal, then approve.
21. "Is my bidding strategy right for this campaign's conversion volume? Be specific about how many conversions a month I actually have." (Needs account access) Smart Bidding needs volume. A good answer names your number and says plainly whether it is enough.
22. "If I increase this campaign's budget by 50%, what does the data suggest happens to cost per conversion? Say what you cannot know." (Needs account access) The "say what you cannot know" clause separates useful forecasting from confident nonsense.
Launch: building the campaign
23. Search campaign from a one-paragraph brief
You are a Google Ads specialist building a Search campaign. Business: [what you sell, in one sentence] Ideal customer: [who buys, and what they type when ready to buy] Goal: [leads | sales | calls], measured as [conversion action] Budget: [amount] per day Geography: [countries, regions or cities] Excluded: [audiences, regions or products we do not want] Build the campaign with: - 4 to 6 tightly themed ad groups, named by theme - 8 to 15 keywords per ad group, each marked exact, phrase or broad - one responsive search ad per ad group: 12 headlines, 4 descriptions - a starting negative keyword list, grouped by theme - the bidding strategy you would start with, and why that one Create it paused. Before you create anything, show me the full structure and wait for my go-ahead. Say what you cannot know.
Needs account access to create it; the structure alone works in any chat.
A good answer shows you the whole tree before touching the account, names the bidding strategy with a reason attached to your conversion volume, and confirms in writing that the campaign is paused. If the word "paused" is missing from the reply, open the account and check before doing anything else.
24. Performance Max asset-group plan
Plan a Performance Max campaign for [product or service]. Audience signals I can supply: [customer list | website visitors | search themes | none yet] Creative I already have: [images, videos, logos — be honest about what is missing] Budget: [amount] per day. Target: [conversion action]. Give me: - how many asset groups, and the theme of each - for each: the search themes, the headline angles, and the exact image and video sizes I still need to produce - what you would exclude, and how - the single reporting view I should check weekly, and the metric in it that tells me PMax is working rather than absorbing budget Explain what each asset group is for in one sentence. Say what you cannot know until it has run.
Works in any chat.
Performance Max hides its own structure, so the deliverable here is the explanation, not the campaign. A good answer tells you which creative you are missing before you launch, and names a weekly check you can actually run.
25. Shopping campaign structure review
Review this Shopping setup and tell me what a specialist would criticise. Products: [how many, and the categories] Current structure: [one campaign for everything | split by category | split by margin | split by best sellers] Feed problems I know about: [missing GTINs, thin titles, ...] Margin range: [low % to high %] Cover, in this order: 1. what the current structure makes impossible to control 2. the split you would use instead, and the reason 3. the five feed fields worth fixing first, ranked by revenue impact 4. what to leave alone Be specific to this catalogue. No generic feed advice.
Works in any chat.
The ranking is the point. Feed advice is abundant and mostly interchangeable; a good answer says which five fields matter for your catalogue, and what it would deliberately not touch.
Keywords and search terms
26. High-intent keyword plan with match types
Build a keyword plan for [product or service] in [market]. Buyer intent I want: someone ready to [buy | book | request a quote], not someone researching. Exclude: [free, DIY, jobs, competitors, wrong geography — edit this] Return a table: keyword | ad group | match type | intent (ready to buy / comparing / researching) | why it belongs. Rules: - 4 to 6 ad groups, each tight enough that one ad fits every keyword in it - no keyword that only makes sense for a different product - mark anything where a phrase or exact version should carry the traffic instead of broad Do not estimate search volume or CPC. If I need those, say so and tell me where to get them.
Works in any chat.
The last clause matters more than the rest. Assistants invent volume and CPC figures with total confidence, and telling it to refuse is what keeps the plan usable. Judge the rest by whether each ad group could share a single ad.
27. Search-term triage into negatives
Pull my search terms from the last [30] days and triage them. Sort every term into one of five buckets: ready to buy, comparing, researching, wrong product or audience, junk. For each bucket give me spend, clicks, conversions and cost per conversion. Then list every term with more than [20] in spend and zero conversions, and for each one say which bucket it is in and why. Finally: propose negative keywords grouped by theme, and say whether each belongs at campaign or ad group level. Propose them — do not add anything yet.
Needs account access.
Five legible buckets, not fifty micro-groups, and an explanation beside each proposed negative so you can disagree with it. Read the list before approving it: that reading is the only thing standing between a negative list and a blocked money keyword.
28. Keyword-to-ad-group remapping
My ad groups have drifted. Audit the mapping. For each campaign show me: - ad groups whose keywords no longer share one theme - keywords that would perform better in a different ad group, with the destination named - keywords competing with each other across ad groups, and which one should win - ad groups thin enough to merge, and ad groups broad enough to split For each move, say what you expect to change and which metric would show it. Propose the plan; change nothing.
Needs account access.
A good answer names specific keywords and specific destinations. "Consider tightening your ad groups" is what you get when the assistant cannot see the account — treat it as a sign the connection is not working.
Ad copy and landing pages
29. Responsive search ad written from the landing page
Here is my landing page: [URL] Write a responsive search ad for the ad group [name], whose keywords are [list]. Hard rules: - every claim must be supported by something on that page - 15 headlines, each a genuinely different angle, not a rewording - 4 descriptions - the primary keyword appears naturally in at least 3 headlines - respect the limits: 30 characters for headlines, 90 for descriptions, and count them After the ad, list every claim you made and quote the line on the page that backs it. If a claim has no backing, say so instead of softening it.
Works in any chat.
The claim-to-quote list is what makes this reviewable in a minute. Count the distinct angles rather than the headlines: five rewordings of "best price" is one headline wearing five hats.
30. Refresh weak copy using performance data
Find my worst-performing responsive search ads over the last [60] days — below account-average click-through rate, with enough impressions to judge. For each one: 1. show the current headlines and descriptions with their asset performance ratings 2. name the mismatch between the ad and the keywords triggering it 3. rewrite only the assets rated Low, keeping the ones rated Good or Best 4. tell me what you expect to improve, and how long I should wait before judging it Do not rewrite an entire ad that is working.
Needs account access.
Point 3 is the whole prompt. Left to itself an assistant rewrites everything, which throws away the only asset-level evidence the account had.
31. Assets and extensions from what the page actually offers
Read [URL] and tell me which Google Ads assets this campaign should have. For each of sitelinks, callouts, structured snippets, images, promotions and lead forms: - whether to use it here, and why or why not - if yes, the exact text, within the character limits - the line on the page that justifies it If the page does not support an asset type, say so rather than inventing the content. Then list what the page is missing that would make a strong asset if I added it.
Works in any chat.
That last sentence turns an ads task into a landing-page to-do list, which is usually where the cheaper win is sitting.
Audit: finding the leak
32. Diagnose a performance drop
My [cost per conversion | conversion rate | conversion volume] moved by about [30]% since [date or period]. Find out why, using this account's data. Work through these in order of how likely each is for this specific account: 1. what changed in the account, with dates, and what each change did to the metric afterwards 2. auction pressure: impression share lost to rank and to budget, before and after 3. traffic mix: new search terms, shifted match types, changed device or geography split 4. seasonality, checked year over year rather than month over month 5. tracking: whether the conversions themselves stopped being recorded For each, say whether the data supports it, contradicts it, or cannot answer it. End with the single most likely cause and the one thing you would check next.
Needs account access.
Ranked and cited, with one data window applied consistently. If the answer is a checklist that never quotes your numbers, the assistant is reciting rather than looking.
33. Verify conversion tracking before changing anything
Before I optimise anything, audit my conversion tracking. Check and report: - every conversion action, its category, and whether it is Primary or Secondary - which ones are actually counted by the bidding strategy - duplicates: two actions recording the same event - conversion counting, Every vs One, and whether that fits the action - attribution model and conversion window per action - any action with zero conversions in the last [30] days - a gap between clicks and recorded conversions that looks like a tracking failure rather than a landing-page problem Then, for each problem: what that wrong setting is doing to my Smart Bidding right now. Rank the problems by how much they distort bidding.
Needs account access.
Run this before anything else on the list. Every optimisation downstream of broken tracking optimises toward the wrong number, and the ranking tells you which broken setting is costing the most.
34. Rank wasted spend across the account
Rank wasted spend in this account over the last [90] days. Define "wasted" explicitly before you count anything, then show the ranking with the spend attached, across: - search terms with spend and no conversions - keywords with spend and no conversions - ad groups whose cost per conversion is more than [2]x the account average - campaigns limited by budget while a worse campaign spends freely - geographies and devices converting far below average - audiences or placements you would exclude For each line: the spend, what you would do, and what could go wrong if I did it. Total it at the end, and say which part of that total you are least confident about.
Needs account access.
Insist on the definition first — you need to be able to disagree with what it counted. The last line separates a number you can act on from a number that merely sounds alarming.
Optimise and report
35. Budget reallocation without resetting learning
Recommend budget changes for the next [30] days, using the last [60] days of data. For every campaign show: current daily budget, spend, conversions, cost per conversion, and whether it was limited by budget. Rules: - never recommend cutting a campaign that was limited by budget without saying so explicitly - no single change larger than [20]% of the current budget - stay within a total of [amount] per day - flag anything where the change would push a campaign back into the learning period Output a table of proposed changes with the reasoning, then the order to make them in. Propose only; change nothing.
Needs account access.
The budget-limited rule prevents the classic error: moving money away from a campaign that never had room to prove itself. Moving it all in one step is the other classic error, which is what the 20% cap is for.
36. The Monday review, one paste
Weekly review. Give me, in this order: 1. last week vs the week before, by campaign: spend, clicks, conversions, cost per conversion, and the change in each 2. any campaign pacing unusually against its budget 3. search terms above [20] in spend with no conversions 4. new search terms that appeared for the first time last week 5. anything that changed in the account, and what changed it 6. the single thing you would fix first, and why that one Keep it to one screen. No recommendations beyond number 6.
Needs account access.
Save this one and send it every Monday. Consistency beats sophistication here: the value is in the comparison arriving every week, not in the prompt being clever.
37. Client-ready performance summary
Write a performance summary for [client or stakeholder] covering [period]. Audience: [technical | non-technical]. They care about [cost per lead | revenue | volume] above everything else. Structure: - three sentences on what happened, in plain language - the numbers that matter, with the previous period beside each - what we changed, and what each change did - what is not working, stated plainly - what we are doing next month, and what we need from them No jargon without a parenthetical explanation. Do not describe a drop as an opportunity. If a number looks good only because of the window it is measured over, say so.
Works in any chat, and much better connected — then it uses your numbers instead of asking you for them.
The two constraints at the end are there because reporting drifts toward flattery on its own. The summary that names what is not working is the one that keeps a client for a second year.
The formula behind every template
Every template above is the same seven parts. Once you can see them, you can write your own.
A role and a job. "You are a Google Ads specialist building a Search campaign" beats "help me with Google Ads", because it fixes the frame before the model starts guessing at one.
A data window. Last 30 days, last 60 days, year over year. Left unstated, the window becomes whatever is convenient, and two runs of the same prompt stop being comparable.
A comparison, not a number. Last week's spend is trivia; last week's spend beside the week before is information. Nearly every prompt above asks for a change, a ranking or a before-and-after rather than a total, and that habit alone is what turns a report into a decision.
Constraints, stated as rules. "No single change larger than 20%." "Keep the assets rated Good." "4 to 6 ad groups." Constraints are where your judgement enters the prompt; without them you get the average of the internet.
An output shape. A table with named columns, a ranked list, a five-bucket split. Naming the shape is what makes two answers a week apart comparable, and it is the cheapest quality improvement available.
"Say what you cannot know." Assistants are fluent by default and uncertain only on request. That one clause turns a confident guess into a labelled one — most usefully on volume, CPC and forecasts, where an invented number looks exactly like a real one.
An approval clause. "Propose only; change nothing." "Create it paused." "Wait for my go-ahead." Every template here that touches the account is safe for exactly one reason: you read the proposal first. That is also how the connection itself is built — writes arrive as proposals and campaigns are created paused; the documentation covers how approvals work.
None of those seven parts is specific to Google Ads. Facebook ads for wedding photographers writes the same formula against a Meta account, where the evidence is placements, frequency and recorded leads rather than search terms.
