Writing and Testing Ad Copy With Claude

#01What makes Claude useful for ad copy?

Volume isn't the point. Any tool can produce fifty headlines. Claude is useful because it can hold a lot of context at once (your positioning, customer language, compliance rules and past results) and produce variations that are actually different from each other, then explain the thinking behind each.

Anthropic's prompting guidance makes the same point in general terms: be clear and direct, give context and motivation, and use examples to show what good looks like (Anthropic). For ads, that means telling Claude who sees the ad, where, what they already believe and what you want them to do next.

Our view: the best use of Claude in paid media is upstream of the copy. Getting the angle right, meaning the reason someone should care, matters more than polishing words, and Claude is good at generating and stress-testing angles quickly.

#02How do you set Claude up with the right context?

Create a project for paid media with these files:

• Positioning and audience notes, including the problems customers describe in their own words. Sales call notes and reviews are gold here.
• Proof points you're allowed to use, each with its source, such as customer counts, awards or review ratings.
• Brand and legal rules: words to avoid, required disclaimers, claims that need approval.
• Past winners and losers, with the results if you have them.

Anthropic explains that project knowledge is available in every chat in the project (Anthropic), so you set this up once and every request benefits.

The "proof points with sources" file matters most. Without it, Claude may write persuasive claims you can't back up, such as "trusted by thousands". With it, you can instruct Claude to use only approved claims.

#03How do you generate angles before copy?

Ask for angles first, then copy for the ones you choose.

Product: [one line]. Audience: [who, and what they're doing
when they see this ad]. Goal: [click to demo / sign up / buy].
Using our project files, give me 8 distinct message angles.
For each: the core idea in one sentence, the customer belief
it plays to, the proof point it relies on (from our approved
list only), and the main risk (e.g. sounds like competitors,
hard to prove). Make the angles genuinely different; don't
give me 8 versions of "save time".

Then pick two or three and pressure-test them: "Argue against each angle as a sceptical buyer would. Which objection would stop them clicking?" This often surfaces a better angle or a missing proof point.

#04How do you write search ads that fit the limits?

Google's responsive search ads accept up to 15 headlines of 30 characters and up to 4 descriptions of 90 characters, and Google may show assets in any order, so each must make sense on its own and in combination (Google Ads Help).

Write a responsive search ad for the keyword theme [theme].
- 15 headlines, max 30 characters each, including spaces
- 4 descriptions, max 90 characters each
- At least 3 headlines contain the main keyword
- Each headline must make sense in any order with any other
- Use only approved proof points
Return a table with the text and its character count.
Then run code to count the characters exactly and fix any
that are over the limit.

The last instruction is important. Language models process text in tokens rather than individual characters, so counts they estimate can be off. Asking Claude to check with code execution, which is available on all plans, or pasting the table into a spreadsheet with a LEN formula, catches overruns before they reach the ad platform.

#05How do you write for social and display?

Each platform has its own constraints and norms, so tell Claude which one and paste the current format requirements from the platform's help pages rather than relying on its memory, as specs change. Then ask for variations by angle, not by synonym:

Write Meta ad primary text for these 3 angles: [angles].
For each angle, give a short version (1-2 lines) and a longer
version (3-5 lines). Open with the customer's problem in their
words, not our product name. Include one proof point. End with
a clear next step. No emojis unless they replace words.

For display or video, use Claude to write the creative brief rather than the final asset: the single message, the visual idea, the text overlay options and what the viewer should feel. Your designer or video team then has something specific to work from.

#06How do you read test results with Claude?

Export results from the ad platform and ask for analysis with guardrails.

Attached: ad-level results for the last 30 days with
impressions, clicks, conversions and spend per ad.
1. Rank ads by conversion rate and cost per conversion.
2. For each comparison you make, tell me whether the sample is
   large enough to trust, and why.
3. Describe what the winning ads have in common in message,
   not just wording.
4. Suggest 3 new variants that test one idea each.
Don't recommend budget changes.

Claude can run the numbers with code execution and explain them. Asking it to judge sample size stops you from crowning a "winner" based on a handful of conversions. For formal significance testing, ask it to show the method and the inputs so someone can check them.

Claude also works in Excel through the add-in on paid plans (Claude), which suits teams that keep ad reports in spreadsheets.

#07From MentionX data: which messages AI engines repeat about youOriginal data

Method. Read-only aggregate queries over MentionX scans from 22 August to 29 October 2026 (UTC): 2,217 engine runs (2,154 completed) across 18 anonymised projects and six engines.

Ad messages and AI answers influence each other: if AI engines describe you differently from your ads, buyers notice. MentionX's aggregates show how engines describe brands in a category:

Our view: check what AI engines say about you before choosing angles. An angle that matches how engines already describe you reinforces it; one that contradicts it needs stronger proof.

Limits. A convenience sample of mostly B2B software projects, run through engine APIs, which can behave differently from the consumer apps. Perplexity, Copilot and DeepSeek figures come from a single day (26 October 2026), so treat them as a snapshot, not a trend.

#08Worked example: Ledgerline's demo campaign

Ledgerline is a fictional accounts-payable tool. Figures are illustrative.

Ledgerline's paid media manager loads positioning notes, twelve approved proof points and three past campaigns into a project. Claude proposes eight angles; she tests "close the month without chasing receipts" against "know your cash position every morning".

Claude writes the responsive search ads and checks lengths with code; two headlines were over 30 characters and get fixed. After four weeks, Claude flags that one ad group's "winner" has too few conversions to call. She extends that test and pauses only the clear loser. Ledgerline is fictional and the details are illustrative.

#09What should you try before the next lesson?

Pick one live campaign. Give Claude your approved proof points and ask for eight angles, then have it argue against each. Compare the strongest one with what you're running now.

Next: Email and lifecycle copy with Claude. Or return to the Claude for Marketers track.

#10How do you keep variants genuinely different?

The most common problem with AI-generated ad copy isn't quality, it's sameness. Ask for twenty headlines and you often get the same idea with different verbs. A test between near-identical ads tells you very little, because whichever wins, you haven't learned anything you can use next time.

Fix it in the prompt by naming the dimension you want to vary. You might vary the motivation (save money, save time, reduce risk, look good to the boss), the format (question, statement, number, customer quote), the specificity (broad promise or concrete detail) or the stage of awareness (problem-aware or solution-aware). Ask Claude to label each variant with the dimension it tests. When results come in, you learn something about your audience, not just about a phrase.

It also helps to ask Claude to review its own list before you see it: "Group these headlines by underlying idea. If more than three share an idea, replace the extras with new ideas." That single step usually improves the spread more than asking for more options.

#11How do you keep ads compliant?

Regulated categories such as finance, health, alcohol and employment have rules on what ads can say, and platforms add their own policies on top. Claude can help you follow them but shouldn't be your final check.

Put your legal team's rules into the project as a plain list: phrases that are banned, claims that need a disclaimer, and words that trigger review. Then add a compliance pass to every copy request: "Before returning the copy, check each line against our rules file and list any line that might break a rule, with the rule." Ask for the check as a separate list so reviewers can see what was flagged and why.

For anything that will run in a regulated category, a qualified person still signs off. The value of Claude here is that legal reviewers see fewer problems and spend their time on the genuinely difficult ones.

#12How do you localise ads with Claude?

Claude can translate, but ads need adaptation rather than translation. A pun, an idiom or a cultural reference that works in one market can fall flat or offend in another. Ask Claude to adapt rather than translate, and to explain each change: "Adapt these ads for the German market. Keep the angle, change the wording to sound natural to a German business buyer, flag any claim that may need different wording under local rules, and explain your main changes in English."

Have a native speaker in that market review the result before launch. Claude gets you a strong first version quickly; local review catches tone and nuance problems that are easy to miss from outside.

#13How do you build a test log Claude can learn from?

Keep a simple file in the project with every test: the date, the angle, the variants, the result and what you concluded. After a few months, that file becomes your most valuable context. Ask Claude to read it before proposing new tests: "Based on our test log, which ideas have we not tried yet, and which have we tested enough to stop testing?"

This turns Claude from a copy generator into something closer to a strategist with a memory of your account, without relying on any automatic memory feature. It also makes onboarding easier when someone new joins the team.

#14How do you match ads to landing pages?

Message match between an ad and its landing page affects whether people stay once they click. Claude is good at spotting the gaps because it can compare both at once. Paste the ad variants and the landing page copy, or give it the page URL with web fetch, and ask: "For each ad, does the landing page's headline and first screen deliver what the ad promised? List mismatches and suggest either an ad change or a page change, whichever is smaller."

Run the same check when you launch new angles. A strong new angle that points at an old page often underperforms, and the reason isn't the ad. Fixing the page headline to echo the angle is frequently the quickest win in the whole test plan, and it costs nothing in media spend to try.

Keep a short record of each mismatch you fix and what happened afterwards. Over time that record shows whether your problem tends to be the ads or the pages, which is useful when planning where to spend effort next quarter. Share it with whoever owns the website, too, so page fixes get prioritised alongside new campaign work rather than after it.

#15Common questions

Can Claude count characters for ad limits? Its estimates can be off. Ask it to verify with code execution, available on all plans, or check lengths in a spreadsheet.

Can Claude launch or change my ads? Keep launches and budget changes in your ad platform under human control. Use Claude for copy, briefs and analysis.

Will Claude invent claims in ads? It can if you don't give it approved proof points. Provide a sourced list and instruct it to use only those.