ChatGPT Ad Copy Prompts for Google, Meta and LinkedIn

The most common AI ad-copy mistake isn’t bad writing - it’s writing for a platform that doesn’t exist. One generic “write ad copy” prompt produces headlines that are five characters too long for Google, primary text that Meta truncates mid-sentence, and LinkedIn intros whose point dies behind the “…see more” fold. (The copy then feeds automated delivery on both platforms - Google’s RSA assembly and Meta’s Advantage+ campaigns - which is exactly why the variety and compliance rules below matter.)

Ads are the one format where the container is non-negotiable: the platform enforces its limits whether your prompt knows them or not. So I wrote three prompts, not one - each built around its platform’s official specs - plus a compliance guardrail block that bolts onto any of them.

The specs that do the work

These limits come from the platforms’ own documentation, not blog folklore:

Platform	Field	Limit
Google RSA	Headline	30 characters, up to 15 headlines (Google Ads Help)
Google RSA	Description	90 characters, up to 4 (Google Ads Help)
Meta (Feed)	Primary text	~125 characters shown before truncation (Meta ads guide)
Meta (Feed)	Headline	~27 characters displayed (Meta ads guide)
LinkedIn	Headline	70 characters (LinkedIn ad specs)
LinkedIn	Intro text	~150 characters before truncation (LinkedIn ad specs)

The pattern I’d internalize: on every platform, the displayed limit is smaller than the allowed limit. Write to what’s displayed.

Prompt 1: Google Responsive Search Ads

Google assembles RSAs from your headline and description pool, mixing and matching per auction - so the real job isn’t writing one ad, it’s writing 15 headlines that work in any combination. (How that assembly works, and when pinning is worth its cost, is covered in my responsive search ads guide; what each combination costs you is the bidder’s decision, and choosing the right Smart Bidding strategy is its own guide.)

You are a search-ads copywriter. Product: [what it is, for whom].
Keyword theme this ad group targets: [keywords].
Offer/proof available: [price, guarantee, review count, or none].
Landing page promises: [paste its H1 and subhead].

Write 15 RSA headlines (max 30 characters each - count them) and
4 descriptions (max 90 characters each).

Headline mix required:
- 4 containing the keyword theme naturally
- 3 benefit-led, 2 proof-led (use ONLY the proof I gave you),
  2 offer-led, 2 call-to-action, 2 plain-descriptive
Rules: every headline must make sense next to any other headline;
no headline may repeat another's first word; no exclamation marks;
nothing the landing page doesn't support.
Format: numbered list with character counts.

Why it works: the mix quota mirrors how Google assembles combinations - without it, models write 15 variations of the same benefit headline, and the “makes sense next to any other” rule is the difference between a pool and a pile. Demanding character counts makes the model self-check its own worst habit: it will happily assert a 34-character line is 30 unless forced to count.

Check it anyway. Character counting is exactly the kind of task models get confidently wrong - paste the output into a character counter (or the Google Ads editor itself) before trusting it. Expect to fix two or three.

Prompt 2: Meta (Facebook/Instagram) feed ads

Meta copy fails in the first line or not at all - roughly 125 characters of primary text show before “See more,” and the displayed headline under the creative is ~27 characters.

You are a paid-social copywriter. Product: [what + for whom].
Audience temperature: [cold / warm retargeting / cart abandoners].
The creative shows: [describe the image/video].
Proof available: [real numbers, reviews - or none].

Write 5 ads. Each has:
- Primary text: first sentence max 125 characters and it must
  carry the entire hook; up to 2 further short sentences after.
- Headline: max 27 characters.
For cold audiences: lead with the problem or a pattern-interrupt.
For warm: lead with the product and the reason to return now.
Rules: the text must not describe the creative ("as you can
see…"); no fake urgency; no invented statistics or testimonials.
Format: label each ad's angle.

Why it works: splitting by audience temperature is what most prompt lists skip - cold and retargeting audiences respond to opposite openings, and one prompt covering both produces mush. The “don’t describe the creative” rule exists because models, told what the image shows, narrate it - burning the only 125 characters that matter.

Prompt 3: LinkedIn sponsored content

LinkedIn is the expensive platform, which changes the copy job: you’re not maximizing clicks, you’re qualifying them.

You are writing LinkedIn sponsored content. Product: [what it is].
Target: [job title(s) at company type]. The problem it solves for
that title specifically: [problem].
Proof available: [real customer/number/report - or none].

Write 4 ads. Each has:
- Intro text: first 150 characters must state who this is for and
  the problem - the qualified reader should recognize themselves
  before the fold. Up to 400 more characters after, but the ad
  must survive on the first 150 alone.
- Headline: max 70 characters, concrete, no wordplay.
Rules: professional but human - no corporate throat-clearing
("In today's business landscape"); name the role, not "leaders";
nothing that would embarrass the reader if a colleague saw them
click it.
Format: label which ads lead with problem vs. proof.

Why it works: “recognize themselves before the fold” reframes the truncation limit as a targeting tool - on a platform with high CPCs, copy that repels the wrong clicker is doing profitable work. The embarrassment rule sounds odd until you remember LinkedIn is browsed at work, in public, with colleagues nearby.

The compliance guardrail block

Append this to any prompt above. It exists because ad platforms reject - or worse, quietly restrict - accounts over claims the model invented to be persuasive:

COMPLIANCE RULES (apply to every line):
- No claims I did not explicitly provide: no invented statistics,
  awards, review counts, savings percentages or "results."
- No superlatives ("best", "#1", "guaranteed") unless I provided
  substantiation.
- No implied urgency or scarcity unless I stated a real deadline
  or real stock limit.
- Nothing targeting or calling out personal attributes (health,
  finances, age, etc.) - write to the problem, not the person.
- If a rule conflicts with persuasiveness, the rule wins. Flag any
  line where you were tempted to break one.

Why it works: models optimize for the goal you give them, and “write persuasive ads” quietly rewards invention. Making the rules explicit - and asking the model to flag temptation - turns it from a compliance risk into a first-pass compliance checker. The personal-attributes rule matters most on Meta, where implied-attribute copy (“struggling with debt?”) is a common rejection reason.

A worked example, end to end

To make the differences concrete, let me run one product through all three prompts. The company is hypothetical - an illustration of the method, not a case study - but the mechanics are exactly what you’d do with your own inputs.

Say the product is project-management software for construction teams, the ad-group keyword theme is “construction project software,” and the only proof you actually have is a 14-day free trial and 4.6 stars from 212 reviews. That last clause is the important one: the proof list you feed the prompt is a fence, and everything outside it is a claim the model may not make.

Google, prompt 1 should give you a pool like: keyword headlines (“Construction Project Software”, 30 chars exactly - models love using the full budget, which is fine), proof headlines built only from what you gave (“4.6 Stars From 212 Reviews”), offer headlines (“Free 14-Day Trial”), and plain-descriptive ones (“Track Crews, Costs & Delays”). What you’re checking in review: no headline promises an outcome you didn’t state (“Cut Delays 30%” should not exist - you never gave it a 30%), and any two headlines read sensibly together, because Google will pair them without asking you.

Meta, prompt 2, cold audience should lead with the problem, inside 125 characters: “Your foreman texts you photos. Your PM lives in spreadsheets. Your schedule lives in someone’s head.” - hook complete before truncation, product explanation after the fold. The failure mode to catch in review is the model narrating the creative or burying the hook in sentence two.

LinkedIn, prompt 3 targets the title, not the industry: “Construction PMs: the schedule slip usually isn’t the crew” fits the 150-character window and lets the right reader self-select. Note how differently the same product sells on each platform - that’s the entire argument for three prompts.

The end-to-end loop, once you’ve run it a few times, settles at roughly: two minutes gathering real inputs (keywords, proof, creative description), one minute running the prompt, and ten minutes reviewing - counting characters, checking claims against the proof fence, killing lines that overlap. That review block is not optional overhead. It’s the job. The prompt only moved the work from writing to editing, which is the trade you wanted, because editing fifteen candidate headlines is faster than staring at a blank RSA form.

Reading the results (so the next batch is better)

The variant labels earn their keep after launch. When you review performance, aggregate by angle, not by individual line: if proof-led headlines win across three ad groups, that’s a finding about your market - lead with proof everywhere, and go get more proof to feed future prompts. If urgency-led lines win but conversion quality drops, you’ve learned your urgency attracts clickers, not buyers - a finding no single-ad view would surface.

Feed the findings back into the prompt itself. The base prompts above are starting points. A team’s version six months in should carry lines like “proof-led angles historically outperform for us - weight the mix accordingly” and a house ban list grown from every cliché you’ve had to delete. A prompt that accumulates your learnings is a compounding asset. A prompt reused unchanged is a plateau. (This is the same logbook logic as the subject-line testing program - paid media just pays for its lessons faster.)

How many variants to test

More isn’t better - every variant splits your budget and slows learning. My working default: 3–5 genuinely different angles per ad set, not ten shades of one angle. The prompts above generate labeled angles for exactly this reason: test angles first, and only iterate wording within an angle after it wins. Kill losers on spend thresholds you set in advance, not on day-two hunches - small samples lie.

Adapting the prompts to your situation

Three adjustments cover most real-world variations. B2B vs. B2C: the prompts above default to a considered-purchase register. For impulse-priced B2C, add “assume the reader decides in under five seconds - one concrete benefit per line, no education” to the rules block, and expect the plain-descriptive angles to matter less. Regulated industries: if you’re in finance, health, employment or housing, extend the compliance block with your category’s specific rules (e.g. “no outcome promises, no before/after framing”) - the federal baseline for claims an AI drafted is in my read of the FTC’s AI advertising rules - and have legal review the template, not every batch - a template cleared once stays cleared, while freeform generation sends every batch back to legal. Small budgets: below the spend needed to exit learning phases quickly, cut the variant count in half and test angles sequentially rather than simultaneously - three angles at starving budgets teaches you nothing about any of them. One angle at a time taught slowly still compounds.

And whatever the situation, keep the landing page in the loop. Every prompt above asks for the landing page’s actual promise as input because the most expensive ad-copy failure is a brilliant headline paying for clicks a mismatched page then wastes. If the model’s best lines don’t match the page, that’s sometimes the page’s problem - my landing page copy prompts work the same message-match discipline from the page’s side. The copy exercise doubles as a cheap positioning audit.

FAQ

Can I just let the platforms’ built-in AI write the copy? Platform AI optimizes for the platform’s goal (spend and engagement), knows nothing about your positioning, and can’t be given your compliance constraints. Use it as another variant source if you like - but judged by the same rules, against your prompt-driven copy.

What about ads inside ChatGPT itself? That’s a placement question, not a copy question - my ChatGPT ads breakdown covers the formats, the targeting and who should test the channel. The copy discipline above transfers: one concrete benefit, limits respected, compliance fenced in.

Do these prompts work for other models? Yes - the anatomy is model-agnostic (see the prompt library). Whatever model you use, keep the character-count double-check. None of them count reliably.

How does this fit a broader AI adoption plan? Ad copy is usually a mid-stage automation candidate - high volume, tight constraints, measurable outcomes. Where it fits in sequencing, and what a paid-media budget conversation looks like at the executive level, is covered in AI for CMOs and the complete guide to AI marketing.

Also in this library: ChatGPT prompts for blog posts, email subject line prompts and social media caption prompts by platform. All of them work better with a brand voice system prompt in place first, sitting inside the standing system prompt that carries your product facts and claims rules.