All resources

· 21 min read

50 Best ChatGPT Prompts for Ads and Ad Copy (2026)

The 50 best ChatGPT prompts for ads and ad copy: Facebook, Google, LinkedIn, TikTok, ChatGPT Ads, and Reddit, plus hooks, creative briefs, and landing pages.

Lapis, the YC-backed #1 AI ad generator rated 4.9/5 on G2, uses AI agents to create, launch, and optimize ads on ChatGPT, Google, Meta, and LinkedIn.

What Makes a ChatGPT Ad Prompt Produce Good Output?

A good ad prompt has five parts: context (brand, product, audience, proof), one task, hard constraints (channel, character limit, banned words), a structure to follow, and an output format such as a table. Missing context produces generic copy. Missing constraints produces copy you rewrite by hand.

OpenAI's guidance says to put instructions first, be specific about context, length, format, and style, and show the desired output format, and Anthropic recommends 3 to 5 diverse examples to steer format and tone. Every prompt below follows the same build:

  1. Load context once with the Brand Block below.
  2. Name one task, such as "15 search headlines for this keyword theme".
  3. Set hard limits: character counts, number of variants, banned words.
  4. Force difference by giving each variant a distinct angle, motive, or structure.
  5. Request a table with a character-count column, ready to paste.

How Do You Give ChatGPT Your Brand Context Once and Reuse It?

Paste one Brand Block (brand, product, audience, objections, proof, voice, and rules) into a ChatGPT Project's instructions. Project instructions apply to every chat in that project and override your global custom instructions, so every prompt below inherits your context. Fill in every bracket and upload your proof sheet as a project file.

BRAND BLOCK. Apply this to every ad task in this project.
Brand: [BRAND], [ONE-LINE DESCRIPTION]
Product and price: [PRODUCT], [PRICE], [OFFER TERMS]
Audience: [SEGMENT], who struggle with [PROBLEM IN THEIR WORDS]
Top objections: [OBJECTION 1]; [OBJECTION 2]; [OBJECTION 3]
Proof I can substantiate: [RATING AND REVIEW COUNT]; [STAT AND SOURCE]; [NAMED CUSTOMER OR AWARD]
Voice: [THREE VOICE RULES]
Competitors: [COMPETITOR A], [COMPETITOR B]
Rules: Use only the facts above. If a line needs a fact I did not give, write [NEEDS PROOF]. Count characters for every line and never exceed my limit. No exclamation points, emojis, or unproven superlatives. Never use: unlock, elevate, seamless, game-changer, revolutionize, supercharge.

What Are the Ad Character Limits for Each Channel?

Google responsive search ads allow 30-character headlines and 90-character descriptions, ChatGPT Ads titles max out at 50 characters, TikTok ad text at 100, and Meta, LinkedIn, and Reddit publish recommended lengths before truncation. Put the exact limit in every prompt and ask for a character-count column.

Channel and fieldLimit
Google RSA headline and description30 and 90 characters, up to 15 and 4 per ad
Google sitelink text25 characters
Meta Facebook Feed primary text and headline50 to 150 and 27 characters recommended
Meta, most placements125 primary text, 40 headline, 25 description recommended
LinkedIn single image150 intro text, 70 headline recommended
TikTok non-Spark ad text1 to 100 characters, no emojis
ChatGPT Ads title and description16 to 24 and 32 to 48 recommended; 50 and 100 maximum
Reddit headline300 maximum, under 100 recommended, key point in the first 80

The 50 Ad Prompts: Research and Strategy (1 to 8)

Research prompts turn raw customer language, competitor ads, and objections into the angles every later prompt uses. Run these first and save the outputs as project files, so the copy prompts that follow work from your customers' words instead of category averages.

1. Review miner. Use it when reviews pile up.

Mine these reviews: [PASTE 20 TO 50]. 15-row table: verbatim phrase, problem or desire, emotion, objection, suggested angle. Never paraphrase column 1.

2. Question map. Use it when planning search or ChatGPT ads.

List 25 questions [AUDIENCE] ask Google or ChatGPT before buying [CATEGORY]. Table: question, stage (problem, solution, comparison, purchase), intent 1 to 5, best angle.

3. Competitor teardown. Use it when rivals crowd your category.

Competitor ads: [PASTE WITH BRAND NAMES]. Table: brand, core promise, proof, offer, CTA. Then list 5 promises nobody makes that my Brand Block proof supports.

4. Objection inventory. Use it when clicks fail to convert.

List the 12 strongest objections to [PRODUCT]. Table: objection in the buyer's words, trigger, answering Brand Block proof (or [NEEDS PROOF]), one 10-word defusing line.

5. Awareness matrix. Use it when one message serves everyone.

Build a table across 5 awareness stages (unaware, problem-aware, solution-aware, product-aware, most aware): buyer belief, message, hook, CTA, best channel.

6. Persona split. Use it when "everyone" is your audience.

Split my audience into 4 personas who buy [PRODUCT] for different reasons. Table: persona, trigger moment, top objection, convincing proof, channel, headline. No stereotypes.

7. Offer stress test. Use it when an offer is untested.

Act as a skeptical buyer. Score my offer [OFFER] 1 to 10 on clarity, value, risk reversal, urgency, and believability. Rewrite it 3 ways that raise the lowest score without discounting.

8. Positioning. Use it when messaging is disputed.

Write 5 statements: For [AUDIENCE] who [PROBLEM], [BRAND] is the [CATEGORY] that [BENEFIT] because [PROOF]. Vary the audience or benefit each time. Mark the one with the strongest proof.

The 50 Ad Prompts: Ad Copy by Channel (9 to 22)

Channel prompts bake each platform's limits and reading context into the request. Search answers a query, social interrupts a scroll, LinkedIn speaks to a job title, TikTok pairs text with a spoken hook, ChatGPT Ads read like a helpful answer, and Reddit rewards candor.

9. Google RSA set. Use it when building a search ad group.

Write an RSA for [KEYWORD THEME]: 15 headlines (max 30 characters), 4 descriptions (max 90). Mix 4 keyword, 3 benefit, 3 proof, 2 offer, 2 objection, 1 brand. Table: text, count, type. No repeated first words.

10. Competitor alternatives. Use it when bidding on rival brand terms.

Write 10 headlines (max 30 characters) and 4 descriptions (max 90) for searchers comparing [COMPETITOR] alternatives. Lead with the switching reason my proof supports. Never imply affiliation.

11. Sitelinks. Use it when an ad group has under 6.

Write 8 sitelinks for these pages: [PAGES AND URLS]. Each: unique link text (max 25 characters), two short description lines, URL.

12. Meta feed set. Use it when launching a Meta angle.

Write Meta feed ads for [ANGLE]: 3 primary texts (short, medium, story), each hooking within 125 characters, plus 5 headlines (max 27 characters).

13. Meta funnel trio. Use it when every audience sees one ad.

Write 3 Meta ads for [PRODUCT]: cold (lead with the problem), site visitors (lead with proof), cart abandoners (lead with [GUARANTEE]). Primary text under 125 characters, headline under 27.

14. Meta carousel. Use it when several benefits need proof.

Write a 5-card Meta carousel for [PRODUCT]: card 1 hooks, cards 2 to 4 each prove one benefit with Brand Block proof, card 5 states the offer. Per card: 5-word headline, visual, proof.

15. LinkedIn ads. Use it when targeting B2B buyers.

Write 4 LinkedIn single image ads for [JOB TITLES] at [COMPANY SIZE] on cost, risk, speed, and peer proof. Intro text under 150 characters, outcome first; headline under 70.

16. LinkedIn founder voice. Use it when selling a point of view.

Turn this insight into 3 LinkedIn sponsored intros: [IDEA]. First 150 characters carry the point, first person, no hashtags, each ending with a specific question for [JOB TITLE].

17. TikTok text and hook. Use it when launching a TikTok video.

Write 8 TikTok ad texts for [PRODUCT]: max 100 characters, no emojis or hashtags, 8 different angles. Pair each with a spoken opening line under 10 words for the first 2 seconds.

18. TikTok script. Use it when briefing an editor or creator.

Write a 20-second TikTok script for [PRODUCT]: hook (0 to 2 seconds), problem (2 to 6), demo of [PROOF] (6 to 16), CTA (16 to 20). Table: time, on-screen text (6 words max), voiceover, shot.

19. ChatGPT Ads pairs. Use it when filling a ChatGPT ad group.

For [CUSTOMER NEED] and these buyer questions [PASTE 4], write 3 title and description pairs per question: titles 16 to 24 characters (max 50), descriptions 32 to 48 (max 100), each a different angle phrased as a helpful answer.

20. ChatGPT Ads plan. Use it when planning a ChatGPT campaign.

Plan ChatGPT ad groups for [PRODUCT LINE], one customer need each. Per group: name, 10 context hints (topics where the ad fits), the answering landing page URL, 2 titles under 24 characters. JSON-style list.

21. Reddit headlines. Use it when promoting in a niche subreddit.

Write 5 promoted post headlines for r/[SUBREDDIT] about [PRODUCT]. Key point in the first 80 characters, under 100 total, written like a helpful member, one admitting a real limitation. No hype.

22. Reddit post body. Use it when the community pushes back.

Write a promoted post body under 500 characters answering the top objection in r/[SUBREDDIT]: [OBJECTION]. Plain language, one specific Brand Block number, a soft CTA, zero marketing adjectives.

The 50 Ad Prompts: Hooks and Angles (23 to 30)

You get genuinely different angles by asking for motives before copy: list distinct reasons to buy, keep the most different ones, then write one hook per reason. Asking for "20 headlines" in one shot returns 20 versions of the same promise, which teaches nothing.

A Science Advances experiment found AI-assisted stories were more similar to each other than stories written alone, while Wharton researchers found chain-of-thought prompting raised GPT-4 idea diversity close to human-group levels. AI ad angles vs. variations explains the testing logic.

23. Ten-motive angles. Use it when every ad sounds alike.

Think step by step. List 10 reasons [AUDIENCE] buy [PRODUCT], one per motive: money, time, status, fear, identity, convenience, health, belonging, control, curiosity. Keep the 6 most distinct. Table: angle, motive, proof, 8-word hook.

24. Synonym detector. Use it when auditing a batch.

Cluster these headlines by underlying promise: [PASTE]. Report the number of distinct promises, flag synonyms, and write one headline for each of 3 missing promises.

25. Hook structures. Use it when an angle wins.

For [ANGLE], write 10 hooks, one per structure: question, number, myth, mistake, confession, comparison, objection, instruction, outcome, curiosity gap. Label each. 10 words max, no clickbait.

26. Contrarian beliefs. Use it when your category sounds identical.

Name 5 beliefs [CATEGORY] buyers hold that my Brand Block proof contradicts. Table: belief, counter-evidence, contrarian hook under 12 words, backfire risk.

27. Specificity upgrade. Use it when copy feels vague.

Rewrite each line with a specific number, timeframe, or named detail from my Brand Block only. Where no fact exists, write [NEEDS PROOF]. Lines: [PASTE]

28. Tone pairs. Use it when testing tone alone.

Write 6 hook pairs for [PRODUCT], one emotional and one rational per promise. Predict which wins for [AUDIENCE] on [CHANNEL] and why, one sentence each.

29. Objection flips. Use it when retargeting.

Turn each Brand Block objection into 2 hooks that voice the objection, then flip it with proof. 14 words max each.

30. Angle scorecard. Use it when choosing what to test.

Score these angles 1 to 5 on audience relevance, distinctness from [COMPETITORS], proof strength, and visual potential: [PASTE]. Rank them and pick 3 to test, each with a one-line hypothesis.

The 50 Ad Prompts: Creative Briefs and Visuals (31 to 38)

Yes, ChatGPT makes ad images: OpenAI says its model renders text accurately, transforms uploaded images, and accepts aspect ratios and hex codes. These prompts brief the visual, generate it from your real product, script UGC-style video from real reviews, and QA every image before upload.

ChatGPT images carry C2PA metadata, which Meta uses to detect third-party AI and apply an AI info label.

31. Creative brief. Use it when design work starts.

Turn my Brand Block and [ANGLE] into a one-page brief: objective, audience, single-minded proposition, proof, mandatory elements, tone, formats [FORMATS], what to avoid, success metric. Under 250 words.

32. Static concepts. Use it when briefing a designer or image tool.

Give 6 static concepts for [ANGLE]. Each: on-image headline (6 words max), visual, layout (logo, headline, product positions), CTA, best ratio (1:1, 4:5, 9:16, or 1.91:1).

33. Image prompt. Use it when generating visuals in ChatGPT.

Create a [RATIO] ad image of [PRODUCT] in [SCENE] using brand colors [HEX CODES]. Keep the top 20% clear and render exactly "[HEADLINE]" there in [FONT STYLE]. No other text, no invented logos.

34. Product scene. Use it when a studio shoot is too slow.

Using the attached product photo, create a photorealistic image of [PRODUCT] on [SURFACE] in [SETTING], [LIGHTING]. Keep label, shape, and colors identical. Square, no people.

35. UGC script. Use it when briefing a customer or creator.

Write a 30-second UGC-style script using only these real review phrases: [PASTE]. Beats: hook, problem, first try, result with one proof point, CTA. Flag every line the speaker must confirm is true for them.

36. Creator brief. Use it when hiring creators.

Write a brief for a [CREATOR TYPE] making a [PLATFORM] ad for [PRODUCT]: 3 hook options, 3 must-say points, banned claims [LIST], paid-partnership disclosure, deliverables and specs.

37. Storyboard. Use it when a script is approved.

Turn this script into a 6-shot storyboard table: shot, duration, visual, on-screen text (6 words max), audio. Product visible by shot 2, captions throughout. Script: [PASTE]

38. Image QA. Use it when an AI visual is ready.

Check the attached ad image: logo matches [LOGO DESCRIPTION], spelling is correct, headline reads at phone size, product matches the real item, nothing important near the edges. Pass or fail each, with the fix.

The 50 Ad Prompts: Landing Pages (39 to 44)

Landing page prompts carry the ad's exact promise onto the page, because a click that lands on a mismatched headline is paid traffic thrown away. Run prompt 39 for every new angle, then use the rest to tighten message match, cut form friction, and tailor heroes by persona.

39. Matched page. Use it when launching a new angle.

Write a landing page for this ad: [PASTE AD]. Headline repeats the ad's promise, subhead adds proof, then 3 benefit sections, an objection FAQ, one CTA. Under 400 words.

40. Hero variants. Use it when testing above-the-fold copy.

Write 5 hero variants for [PAGE], one per angle: [ANGLES]. Each: headline (10 words max), subhead (20 words max), outcome-led button (4 words max).

41. Objection FAQ. Use it when visitors bounce.

Write 8 FAQ entries for [PAGE] answering my Brand Block objections plus [EXTRA QUESTIONS]. Answers under 50 words, citing Brand Block proof where it exists.

42. Ad-to-page audit. Use it when clicks rise but sales stall.

Compare this ad [PASTE AD] with this page [PASTE PAGE]. List every mismatch in promise, offer, price, or wording, score message match 1 to 10, and rewrite the headline and first paragraph.

43. Form fix. Use it when lead forms underperform.

My form asks for [FIELDS]; the button says [CTA]. Name fields to cut, rewrite the button 5 outcome-led ways, and write one line of microcopy answering the top objection.

44. Persona heroes. Use it when prompt 6 is done.

Rewrite this hero for each of my 4 personas, keeping the offer identical and changing only headline, subhead, and proof. Table: persona, headline, subhead, proof. Hero: [PASTE]

The 50 Ad Prompts: Testing and Analysis (45 to 50)

Testing prompts close the loop: they design a clean test, read results by angle instead of by ad, expand winners without cloning them, and diagnose fatigue. Paste real exports, make ChatGPT state what the data does not prove, and never let it crown a winner on thin spend.

45. Test plan. Use it when new creative is ready.

Design a creative test on [CHANNEL] at [DAILY BUDGET]. Pick 1 control and 2 challengers from [ANGLES]. State the hypothesis, the single changed variable, the primary metric, and minimum spend per ad before calling a winner.

46. Results by angle. Use it when a test ends.

From this export [PASTE AD, ANGLE, SPEND, IMPRESSIONS, CLICKS, CONVERSIONS], calculate CTR, CPC, conversion rate, and CPA per angle, rank angles, flag ads with too little spend to judge, and state what the data does not prove.

47. Winner expansion. Use it when one ad clearly wins.

My winner is [PASTE AD] at [CPA] CPA and [CTR] CTR. Explain in one sentence why it likely won, then write 5 ads that keep that reason but change hook, format, or audience. Never reuse its headline.

48. Fatigue check. Use it when a winner slips.

Weekly data for [AD]: [PASTE WEEK, FREQUENCY, CTR, CPM, CPA]. Diagnose creative fatigue, audience saturation, or auction pressure, cite the numbers, and recommend one action.

49. Search terms to headlines. Use it when mining search terms.

Group these converting search terms into themes: [PASTE]. Write 3 RSA headlines (max 30 characters) per theme in the searcher's own words.

50. Weekly report. Use it when reporting to leadership.

Summarize this week for my CEO in 5 bullets: what worked, what failed, what we learned about customers, next week's 3 tests, budget ask. Back every claim with a number. Data: [PASTE]

Go deeper with ad analysis in ChatGPT and headline testing.

How Do You Fact-Check Claims ChatGPT Invents?

Treat every number, rating, award, and comparison ChatGPT writes as unverified until it matches your proof sheet. OpenAI states that ChatGPT still hallucinates plausible but false statements, and the FTC requires a "reasonable basis" of objective evidence before an ad runs, so the fix is a rule inside the prompt plus a check after it.

Run this check on every batch:

  1. Enforce the [NEEDS PROOF] rule so gaps show up as brackets, not inventions.
  2. Ask for a claims table listing each factual claim and the Brand Block line that supports it.
  3. Strip absolute claims. TikTok rejects absolutes like "Number 1 essence on TikTok" and exaggerated results.
  4. Confirm every testimonial is real. The FTC bans testimonials from people who do not exist, including AI-generated ones.

When Do Prompts Stop Scaling and You Need a System?

Prompts stop scaling the moment you run more than one channel, more than a few angles, or more than one language. ChatGPT drafts words fast, but every draft still needs design, sizing, upload, a landing page, tracking, and analysis by hand.

A 40-person B2B SaaS company plans a fourth-quarter push with 6 angles across 5 channels in 2 languages: 60 copy sets that ChatGPT drafts in one afternoon.

StepPrompts plus manual workLapis
Copy for 60 setsOne afternoon of promptingMinutes from one brief
240 statics (60 sets in 4 ratios)40 hours at 10 minutes per fileGenerated on-brand and auto-sized
6 landing pages24 hours at 4 hours per pageRapidDomain builds one per ad
Upload and tracking in 5 ad managers10 hoursPublished in a click
Weekly reporting from 5 exports3 hours every weekWayfinder and built-in analytics

That is 77 hours, nearly two work weeks, before the first result. The ten AI ad jobs show where that time goes.

Why Is Lapis the Best Way to Turn Ad Prompts Into Campaigns?

Lapis is the best way to turn ad prompts into campaigns because it replaces the prompt loop with a system: it learns your brand from a URL, generates finished on-brand creative for every channel, builds the landing page, launches, measures, and improves. Prompts give you drafts. Lapis gives you campaigns that ship.

Lapis is one of the fastest-growing Y Combinator startups, and it is built to take over the work legacy agencies and ad buyers used to do. Its Y Combinator profile lists it as a Fall 2025 company used by 1,500+ marketing teams and 30+ enterprises, and G2 reviewers rate it 4.9 out of 5 with eight badges. Samsung, Domino's, and Dr Pepper use it.

  • Brand Intelligence replaces the Brand Block: paste a URL and Lapis extracts logo, colors, typography, product images, and voice.
  • Campaign Studio replaces prompt iteration: type "shorter headline" and the campaign updates.
  • OmniSense ships the output: sized campaigns for five networks in under 3 minutes, forecast before spend.
  • ChatSense runs ChatGPT ads end to end, as the prompt-to-ChatGPT-ads workflow shows.
  • RapidDomain builds every page, and Wayfinder ties results to confirmed sales.

That stack replaces the copywriter, the designer, and the agency retainer. Try the Free AI Ad Generator with no signup or see pricing.

Sources and Methodology

This guide draws on platform documentation, FTC guidance, and AI creativity research reviewed as of September 4, 2026.

Related Lapis Resources

Built by Lapis

The #1 AI ad generator, built into the operating system for paid growth.

Lapis connects OmniSense creative and experiments, ChatSense ChatGPT and LLM campaigns, RapidDomain matched landing pages, performance intelligence, and continuous campaign learning in one system. Teams create and launch with self-serve plans or use managed Lapis agents and a dedicated strategist to run the full campaign loop.

Lapis is rated 4.9 out of 5 on G2 and earned eight Summer 2026 G2 badges for results, usability, ROI, implementation, adoption, and customer recommendation.

Continue exploring

Frequently Asked Questions