A Google Ads prompt that works gives the AI a role, real data pasted in (your search terms report, the Google results page for a keyword, your offer) and a fixed output format. Without them, ChatGPT or Claude guess: by default, the AI sees neither your account nor today's Google results page. Here are 14 prompts we tried with Claude, ready to copy.
Key takeaway: Fed with your real data, a prompt helps you pay only for clicks that can turn into sales; without it, you get plausible lists that cost money.
Most Google Ads prompts fail because the AI gets no data. 'Give me keywords for my heat pump business': ChatGPT or Claude answer from what they learned, without seeing your account, your search terms or today's Google results page. You get a plausible list. Nothing tells you whether those keywords bring buyers, or what they will cost you.
The second flaw is the lack of a fixed format. With no set categories and no table requested, the answer changes on every try, mixes advice with lists, and fits nowhere. A strict format (columns, categories, word limits) makes the answer comparable from one run to the next and easy to check.
The third flaw is the lack of context about your offer. A keyword is only good or bad relative to what you sell. 'used heat pump' is worthless to an installer who only installs new systems, and valuable to a reseller of refurbished equipment. Saying what you do not sell also stops the AI from keeping those searches by default.
The 14 prompts on this page are built on the same model. It follows the advice in Anthropic's and OpenAI's prompt writing guides, listed in the sources: a role, clear and explicit directions, context, and the data kept apart from the request. Replace each part in square brackets with your own information before you send.
Sorting by buying intent is the most profitable prompt, because it works before you spend anything. Paste your list, your agency's list or the one Keyword Planner gave you, describe your offer and what it leaves out, then ask for one category per keyword. A keyword you drop at this stage costs you nothing.
The AI's sort is a hypothesis that the Google results page confirms or not. On our sample data, the AI put 'used heat pump for sale' and 'hvac technician jobs denver' in Off-target without hesitating, and gave low confidence to 'heat pump rebates colorado' and 'heat pump noise complaints', two searches that can come from a future customer or from someone who is just curious. Those are the rows to check, with the prompts that read the Google results page, further down, or with our method to decide whether a keyword deserves a bid.
You are a Google Ads consultant. Your goal: only pay for clicks from people likely to buy.
My offer: [what you sell, to whom, in which area]
What I do not sell: [products, services or customers you do not serve]
Here is my keyword list, one per line:
[paste your keywords]
Put each keyword in exactly one category:
- Buy: the person wants to buy or hire someone now.
- Compare: they are comparing offers, prices or brands before choosing.
- Learn: they want to understand something, with no visible intent to buy.
- Off-target: they want something other than my offer (jobs, used items, a product I do not sell, another area).
Answer with a table: keyword, category, clue that justifies the choice (12 words max), confidence (high, medium or low).
Keep my order and do not add any keyword. If a keyword is ambiguous, set confidence to low instead of guessing.
End with the number of keywords in each category.What you get: A table that puts each keyword in Buy, Compare, Learn or Off-target, with the clue behind the choice and a confidence level, then a count per category. Low-confidence rows are the ones to check on the Google results page.
You are a Google Ads specialist who builds Search campaigns.
My offer: [what you sell, to whom, in which area]
Here are keywords I have labeled Buy or Compare:
[paste the keywords you kept]
Group them into ad groups. Rules:
- A group holds keywords that one ad and one landing page can serve.
- 2 to 10 keywords per group. A keyword that fits with no other goes into a group called 'Standalone'.
- Each keyword in one group only. Do not add or remove any keyword.
For each group, give: a short name, its keywords, the ideal landing page in one sentence, and the main message of the ad.
Finish by checking that each of my keywords appears exactly once.What you get: Ad groups, each with its ideal landing page and main ad message, then a check that no keyword was lost or added along the way.
You are a Google Ads analyst who is careful with money.
My numbers:
- Average margin per sale: [your margin per sale]
- Conversion rate of my campaigns, from click to sale: [your conversion rate]
- Current average CPC of my account: [your average CPC]
Here are keywords, one per line, in the format 'keyword ; high range'. The high range is the 'Top of page bid (high range)' column in Keyword Planner:
[paste your rows]
1. Work out the maximum CPC I can pay without losing money: margin per sale times conversion rate. Show the math.
2. For each keyword, work out the gap in percent: (high range minus maximum) divided by maximum.
3. Answer with a table: keyword, high range, gap in percent, verdict. Verdicts: within budget (gap below -20%), borderline (gap from -20% to 0%), too expensive (gap above 0%).
4. Also compare that maximum with my current average CPC and say what it means in one sentence.
Use only my numbers. If one is missing, say which one instead of making it up.What you get: The maximum CPC your margin allows, with the math shown, and a verdict per keyword: within budget, borderline or too expensive. The math applies the same conversion rate to every keyword: refine it as soon as you have per-keyword numbers.
The search terms report is the best data you can give the AI, because it shows what you have already paid for. Google describes it as the list of searches that triggered your ads; some rare terms are left out to protect searchers' privacy. Export it, keep four columns (term, clicks, cost, conversions) and paste it into the first prompt.
A badly chosen negative costs more than a useless term, because it also blocks your good searches. Google notes that negative keywords do not cover close variants: singular or plural forms and synonyms get through unless you add them (misspellings, on the other hand, are covered). In our tests we planted a trap: one active keyword contained the word 'price', which also appeared in a term to exclude. The AI refused to exclude 'price' on its own and suggested the narrower phrase 'home depot' instead. To compare the tools that do this work, see our roundup of negative keyword tools.
You are a Google Ads specialist whose job is to cut wasted spend.
My offer: [what you sell, to whom, in which area]
What I do not sell: [products, services or customers you do not serve]
Here is an export of my search terms report (search term ; clicks ; cost ; conversions):
[paste the export]
For each term, pick one action:
- Keep: it matches my offer.
- Exclude: it does not match my offer and cannot turn into a customer.
- Watch: the intent is ambiguous, or there is not enough data to decide.
Rules: a term that has already converted is never excluded. Judge the meaning of the term first, then its numbers.
Answer with a table sorted by cost, highest first: term, action, reason (12 words max), cost.
End with the total cost of the terms to exclude and the total cost of the terms to watch.What you get: A table sorted by cost, with one action per term (keep, exclude, watch) and its reason, then what the terms to exclude and the terms to watch cost you.
You are a Google Ads specialist.
Here are the search terms I have decided to exclude:
[paste the terms to exclude]
Here are my active keywords, which must not be blocked:
[paste your active keywords]
Suggest negative keywords. Rules:
- Prefer one word or a short phrase that covers several terms over one negative per term.
- For each negative, pick the match type: broad, phrase or exact.
- Check each negative against my active keywords. If it would block one, do not suggest it and explain why.
- Google does not extend negatives to close variants: add singular or plural forms and synonyms yourself when they matter.
Answer with a table: negative keyword, match type, terms covered, risk to my active keywords.
Then give the final list to paste into Google Ads, one negative per line: in quotation marks for phrase match, in square brackets for exact match, plain for broad match.What you get: Negatives with their match type, the terms they cover and their risk to your active keywords, then the list ready to paste into Google Ads. Negatives that would have blocked an active keyword are dropped, with the reason.
You are a Google Ads analyst.
My offer: [what you sell, to whom, in which area]
Here are the search terms from the last [number] days (term ; clicks ; cost ; conversions):
[paste the export]
Group the terms with no conversion by intent theme, for example: jobs and training, used or cheap, do it yourself, another area, general information. Create other themes if these do not fit, and leave out the terms that match my offer.
Aim for 3 to 8 themes. For each theme, give: name, number of terms, total clicks, total cost, two examples (one if the theme has a single term), and the negative keywords, one to three, that would block the whole theme without touching my offer.
Sort the themes by total cost, highest first. Use only terms from my export and double-check your sums.What you get: The themes that cost you money without converting, sorted by cost, with the negatives that block them. Useful on a long export, when reading term by term takes too long.
The Google results page is the best proof of intent, and the AI cannot see it unless you paste it. Ads at the top and bottom, organic results, question boxes, business listings: everything shows what the person typing the keyword wants. If local installers fill the ads and the business listings, the search is commercial. If price guides, forums and videos dominate, the person is still researching.
A verdict drawn from a pasted page only holds for that page. Ads change with location, device, time and the person searching. On our sample pages, the AI concluded 'Test with a small budget' for 'heat pump installation cost', where the organic results only informed (cost guide, forum thread, video, brand comparison), and preferred 'heat pump installer denver', dominated by business listings. That is the work Should I Bid's intent analysis does, reading the live Google results page for each keyword and the pages that rank on it.
You are a Google Ads consultant who decides before paying for the click.
My offer: [what you sell, to whom, in which area]
Keyword: [the keyword]
Here is the text of the Google results page for this keyword, copied as is (ads, organic results, boxes):
[paste the results page]
Analyze only this text, without assuming anything else:
1. How many ads do you see, and from which advertisers?
2. What do these ads sell: the same thing as me, something else, or nothing?
3. What do the organic results show: sales or service pages, comparisons, articles that explain, forums, videos?
4. What does the person typing this keyword most likely want to do?
Finish with one verdict, Bid, Test with a small budget, or Skip, and the reason in two sentences.
If the pasted text looks incomplete (no ads, cut-off page), say so before concluding.What you get: A count of the ads and advertisers, what each one sells compared with you, the type of organic results, then a reasoned Bid, Test or Skip verdict. The verdict only holds for the page you pasted, on the day you copied it.
You are a Google Ads copywriter.
Here are the ads copied from the Google results page for [the keyword]:
[paste the ads: advertiser, headline, description]
My verifiable strengths: [your verifiable strengths]
1. For each ad, record in a table: the advertiser, the main promise, the proof offered (numbers, reviews, certifications), the call to action, and what it says about price.
2. List the promises that several advertisers repeat.
3. Suggest two angles that no ad uses and that my strengths can back up.
Do not make up anything that is not in the pasted ads or in my strengths.What you get: A table of each competitor's promises, proof, calls to action and price mentions, the list of promises everyone repeats, and two open angles your strengths can back up.
You are a Google Ads consultant with a limited budget.
My offer: [what you sell, to whom, in which area]
Google results page for keyword A, [keyword A]:
[paste page A]
Google results page for keyword B, [keyword B]:
[paste page B]
Compare the two pages on four points: number of ads, type of advertisers, type of organic results, likely intent of the searcher.
Answer with a three-column table: point compared, keyword A, keyword B.
Then say which one should get the budget first and why, in three sentences at most. If the two pages are too similar to decide, say so.What you get: A table that compares the two pages point by point, then a short recommendation: which keyword gets the budget first, and why.
A responsive search ad is easy to draft with AI, as long as you give it Google's rules and your real strengths. According to Google Ads help on responsive search ads, you can enter up to 15 headlines of 30 characters and 4 descriptions of 90 characters, which Google combines on its own. Google advises including at least one of your keywords in your headlines and providing as many unique headlines as you can.
Recount the characters anyway, and check every claim. In our tests, the lengths the AI reported were right and no strength was made up: the prompt forbade it and listed the real ones. That is not guaranteed, because a language model splits text into chunks of words, not letters, and can be off by a character; the LEN function in a spreadsheet is enough to check. The review prompt caught the headline that was too long, the duplicate 'NATE-Certified Technicians' and 'Certified NATE Technicians', and claims to prove such as 'The #1' or 'Cut Your Heating Bill by 70%'.
You are a Google Ads copywriter.
My offer: [what you sell, to whom, in which area]
My verifiable strengths: [turnaround times, guarantees, certifications, reviews, prices if you show them]
Ad group: [ad group name]
Keywords in the group: [the keywords in the group]
Write a responsive search ad:
- 15 headlines of 30 characters or fewer, spaces included.
- 4 descriptions of 90 characters or fewer, spaces included.
- At least 3 headlines use a keyword from the group.
- Varied headlines: keyword, strength, proof, call to action. Never two headlines that say the same thing.
- Use only my strengths. Do not make up numbers, promises or prices.
Answer with two numbered tables, headlines then descriptions: text, character count. Recount each text before answering.What you get: 15 numbered headlines and 4 descriptions, with their length, built only on your strengths. Recount before you publish: Google Ads fields take no more than 30 characters per headline and 90 per description.
You are a demanding Google Ads ad reviewer.
My offer: [what you sell, to whom, in which area]
Keywords in the group: [the keywords in the group]
Here is my responsive search ad, headlines first, then descriptions, one per line:
[paste your headlines and descriptions]
Flag:
- headlines over 30 characters and descriptions over 90 characters;
- headlines that repeat each other or say nearly the same thing;
- claims I must be able to prove (superlatives, numbers, guarantees);
- keywords from the group that appear in no headline.
Answer with a table: text concerned, problem, suggested fix that stays within the character limit. Do not list texts with no problem.What you get: A list of only the texts with a problem (too long, duplicated, claims to prove, missing keywords), each with a fix that fits the character limit.
You are a Google Ads copywriter.
My offer: [what you sell, to whom, in which area]
My verifiable strengths: [your verifiable strengths]
Group 1, people ready to buy: [buying keywords]
Group 2, people comparing options: [comparison keywords]
For each group, write 5 headlines of 30 characters or fewer that answer what the person wants at that moment: to act for group 1, to feel reassured and compare for group 2.
Give the character count of each headline. Do not make up any strength.
End with one sentence that explains the difference in tone between the two sets.What you get: Two sets of 5 headlines, one that pushes to act, one that reassures and helps compare, with their length and one sentence on the difference in tone.
An ad that promises what the page does not deliver pays for clicks that leave right away. The person clicked on 'In-Home Estimate in 72 Hours' and reads 'we call you back within 5 business days': they leave, and the click is paid for. Google takes landing page experience into account when it assesses ad quality, as our guide to Google Ads Quality Score explains.
Copy the page text from your browser, top to bottom, and paste it with the ad. On our sample data, the first prompt caught the two promises missing from the page (the 72-hour estimate and the rebate paperwork help) and the conflict on timing, quoting the exact sentence each time. The second rewrote the top of the page using only facts from the page, and refused to add the promises the page did not keep.
You are a Google Ads auditor who checks what people see after the click.
Keyword: [the keyword]
My ad (headlines and descriptions):
[paste the ad]
Text of the landing page, copied from the browser, top to bottom:
[paste the page text]
Check point by point:
1. Is each promise in the ad kept on the page? Quote the sentence on the page that keeps it, or write 'missing'.
2. Does the keyword, or its idea, appear at the top of the page?
3. Can the visitor take the action the ad offers (quote, call, purchase) from the top of the page?
4. Does anything on the page contradict the ad (price, timing, area, offer)?
Answer with a table: point checked, finding, severity (blocking, to fix, fine). End with the three fixes to make first.What you get: A table of the promises the page keeps, misses or contradicts, with the exact sentence from the page when there is one, then the three fixes to make first.
You are a landing page specialist for Google Ads.
Keyword: [the keyword]
What the ad promises: [the headline and main description of your ad]
Current text of my landing page:
[paste the page text]
Rewrite only the top of the page, what people see without scrolling: a headline, a subheadline, three key points and the button text.
Rules: the headline carries the idea of the keyword; the top of the page keeps the promise of the ad; every fact comes from my page, do not make any up.
After each element, quote in parentheses the sentence from my page that the fact comes from. If the ad promises something the page does not say, flag it instead of writing it.What you get: A headline, a subheadline, three key points and a button built only from facts on your page, each with its source sentence, and the list of the ad's promises the page does not keep.
We tried every prompt with Claude, on made-up sample data, never on real accounts. For the English version, a fictional US company: a heat pump installer in Denver, with 20 keywords, a 19-row search terms export, two Google results pages, four competing ads and a landing page, all written for the exercise. The French version went through the same steps with a fictional French company in the same trade, in Montpellier.
Filled with the sample data, the prompts were given to Claude as is, one after another and set by set, not each in a fresh conversation. We read every answer, redid the sums and recounted the ad characters with a script, and found no errors. The prompt that groups wasted spend by theme had to be fixed: it asked for a single negative per theme, and the AI split the export into ten themes, eight of them with a single term. It now asks for one to three negatives and aims for 3 to 8 themes: on the new run, it returned 7. Two others were tightened afterward, the budget one and the one that turns excluded terms into negatives, then run again on the same data.
Your results will depend on the data you paste. A short list, a truncated export or a cut-off page gives a poor answer. We did not try these prompts in ChatGPT: they use no feature specific to Claude, but the answer can change from one assistant to another, and from one run to the next with the same assistant.
The manual path works with any assistant and only costs you time. For the search terms report, open Google Ads, then Campaigns, Insights and reports, and Search terms. Pick the date range, download the export and keep the four columns the prompt needs. If the export is very long, start with the terms that cost the most.
For the results page, run the search yourself and copy all the text. Open a private browsing window, type the keyword, select the whole page and paste it into the prompt. Ads are labeled as sponsored; there is no need to click them, the copied text is enough. To see your own ads without skewing your stats, Google offers the Ad preview and diagnosis tool.
The limit of Path A is time, and the moment of the copy. Copying the Google results page for 30 keywords, one by one, quickly takes an evening, and each copy only shows one location, one device and one moment. For a long list, or to redo the sort every month, Path B saves you that work.
On the connected path, Should I Bid reads the real Google results page for each keyword, with no copy and paste. The Should I Bid connector for Claude gives Claude access to your Should I Bid account. You ask in plain language for keyword ideas for a product or a web page: it gets them with Vision, without using any credit, sorted into branded, non-branded, generic and question ideas.
Before any analysis, Claude shows you the cost and waits for your approval. It requests an estimate (1 credit = 1 keyword analyzed; for Vision ideas, only keywords that score 50 or more are charged), tells you how many credits will be reserved, and only launches the analysis once you say yes. Should I Bid then reads the live Google results page for each keyword, ads included, plus the pages that rank on it, and weighs it all against what your company sells.
The Bid, Test or Skip verdicts come back in the conversation, with a score out of 100 and the reasons behind them. Once the analysis is done, Claude can also give you the recommended negative keywords. The connector is included in Should I Bid paid plans (see pricing) and works in claude.ai, Claude Desktop, the mobile app, Cowork and Claude Code.
The connector does not replace every prompt on this page. It does not read your Google Ads account: for the search terms report, Path A is still the way. It is built for Claude: the connection only accepts Claude apps, not ChatGPT. To write the ads and check the landing page, keep the prompts from the sections above, in the same conversation, on the keywords worth bidding on.
By default, an AI assistant sees neither the Google results page your customers see, nor your numbers, nor Google Ads' numbers. Ask it for a keyword's search volume or CPC: at best, it declines or gives a range with no source; at worst, a precise number that rests on nothing. Paste the Google results page, get volumes and CPC ranges from Keyword Planner, free with a Google Ads account, and your margins from your own books. Should I Bid does not provide volumes or CPCs either: it reads the Google results page and returns a verdict per keyword.
It can be wrong with confidence, on details that cost money. Our tests showed no math errors, but they covered about twenty rows. On an export of several hundred terms, recheck the totals, and check each negative before you add it: a mistake there blocks customers.
What you paste leaves your account. Before you send an export, remove anything that should not leave the company (customer names, email addresses, confidential figures) and follow your company’s rules on AI tools.
It does not decide for you. A prompt helps you sort, write and check. The decision to pay for a keyword is still yours, with your margins and your knowledge of your customers. For a full method, read our guide to choosing Google Ads keywords.
| Tool | Starting price | Details | Source |
|---|---|---|---|
| Should I Bid | Per credit (see Pricing page) | Per credit (1 credit = 1 keyword analyzed). 5 free analyses to try, no credit card, then paid subscription plans with no commitment and credits renewed every month. Plans described on the Pricing page. | Source |
| Google Keyword Planner | Free | Free with a Google Ads account. Exact volumes require an account with active spend, otherwise ranges. Google requires entering billing information and completing account setup to access keyword ideas. | Source |
Prices verified on . Public prices observed on the vendors’ websites on the date shown. They may have changed since: the source prevails.
There is no universal prompt. The one that works gives the AI a role, your offer and what you do not sell, real data pasted in, a precise task, rules and a fixed output format. To start, sorting by buying intent pays off the most: it removes off-target keywords before they cost you a single click.
Both can follow these prompts, which use no feature specific to either one. We tried them with Claude only, on sample data. What changes the result is mostly what you paste: a real search terms report, a real Google results page. To plug in Google results pages read live, the Should I Bid connector works with Claude.
Yes, from your search terms report, not from memory. Paste the export, describe your offer and what you do not sell, then ask for one action per term. Then check each negative against your active keywords: a negative that is too broad blocks your good searches. Google also notes that negative keywords do not cover close variants, such as plurals.
Not without checking. In our tests, the lengths Claude gave were right, but a language model does not read text letter by letter and can be off by a character. Google limits responsive search ad headlines to 30 characters and descriptions to 90. Recount with the LEN function in a spreadsheet before you paste the texts into Google Ads.
Partly. To judge whether a keyword is worth a click, yes: through the connector for Claude, Should I Bid reads the real Google results page for each keyword and returns a Bid, Test or Skip verdict, after you approve the number of credits. It does not read your Google Ads account: for the search terms report, the ads and the landing page, the prompts are still useful. It is included in paid plans.
Not reliably. By default, an AI assistant has no access to Google Ads data: at best, it gives a range with no source. Get volumes and CPC ranges from Keyword Planner, free with a Google Ads account, then paste them into the prompt that checks whether a keyword fits your budget.