ChatGPT for Google Ads Copy: The Honest Playbook

Most guides about using ChatGPT to write Google Ads copy skip the one thing that wrecks the workflow: ChatGPT cannot count characters reliably. Tell it “30 characters max” and it’ll hand you headlines that are 34, 38, sometimes 41. One agency documented their attempt and found character limits aren’t always followed – they couldn’t get a single branded headline under 30 characters. Every tutorial out there gives you a “15 headlines, 4 descriptions” prompt and walks away. They don’t tell you half the output won’t paste into Google Ads Editor without edits.

So we’re going to reverse the usual approach. Instead of “write a prompt → generate → paste,” the real workflow is “generate too much → filter mechanically → let ChatGPT rewrite only the violators.” Here’s how that looks.

The specs ChatGPT needs to actually respect

Before any prompting, you need to know the constraints – not because ChatGPT will follow them (it won’t), but because you’ll be the one catching the violations. Responsive search ads let you enter multiple headlines and descriptions; Google tests combinations over time and learns which perform best. The hard numbers come from Google’s official RSA documentation:

Two things the table above doesn’t show. Every character in a double-width language like Korean, Japanese, or Chinese counts as 2 characters per Google’s official spec – so a 15-character Japanese headline already maxes the limit. And emojis? ChatGPT will add them if you don’t tell it not to. Google’s advertising policies prohibit emojis in headlines, descriptions, and display URLs – ads containing them are automatically rejected.

A prompt structure built around ChatGPT’s actual weaknesses

The standard “act as an expert Google Ads copywriter, write 15 headlines and 4 descriptions” prompt is everywhere. It also produces mediocre, repetitive output. Position2’s audit of ChatGPT ad copy found the range of descriptors is narrow – digital, business, marketing, strategies, success, growth, results, expertise. That word soup is exactly what Google’s Ad Strength algorithm penalizes.

Try this instead. It’s longer, but each block fixes a documented failure mode:

You are writing Google RSA copy for: [product/service]
Audience: [specific person, pains, current alternative they use]
Primary keyword: [the exact search term]
Offer/proof: [number, guarantee, social proof - be specific]
Forbidden words: "unlock," "empower," "seamless," "solutions," "experience"
(force variety - these are ChatGPT's defaults)

Generate 25 headlines, not 15. I'll cut the weak ones.
Each headline must:
- Be 28 characters or fewer (leave buffer - your count is unreliable)
- Use a distinct angle: benefit, proof, urgency, objection, keyword-match, question, price, comparison
- Stand alone with no dependency on another headline
- Contain no emojis, no exclamation stacking, no ALL CAPS words

After generating, list each headline with its exact character count
in brackets, like: "Cut Onboarding Time in Half [27]"
Then flag any over 28 so I can see them.

The buffer is the key move – ask for 28 characters and even when ChatGPT miscounts by 2-3, you’re still under 30. Forcing it to print the count next to each headline turns the model into its own QA layer. It still gets counts wrong, but you’ll see the wrong counts and can paste headlines into a character counter to verify.

Pro tip: Never trust ChatGPT’s self-reported character counts. Paste every headline into Google Ads Editor or a dedicated counter before going live. The model treats “character” as roughly synonymous with “token,” and tokens aren’t characters.

The pinning trap nobody warns you about

Once you have your 25 ChatGPT headlines and pick the best 10, you’ll be tempted to pin your brand name to Headline 1. Resist. Or at least understand the cost. Pinning just one headline already decreases the amount of testing Google runs on the RSA by 75%, per Honcho Search’s RSA breakdown.

Think about what that means. You used ChatGPT specifically to generate variety so Google’s algorithm could test combinations. Pin one headline and you’ve thrown away three quarters of that benefit. If you have text that must appear in every ad, Google requires it to be pinned to Headline 1, Headline 2, or Description 1 – but only do it when legal, brand, or compliance reasons actually require it. “I like this headline best” is not a reason.

Why your ChatGPT-generated RSA might never serve

Here’s a failure mode missing from every tutorial: if your headlines or descriptions are too similar, Google won’t even show your search ad. ChatGPT’s default behavior is exactly this – generating 15 minor variations on the same sentence. “Save Time on Bookkeeping,” “Cut Hours From Bookkeeping,” “Faster Bookkeeping for SMBs,” “Bookkeeping Made Quicker.” Identical semantic content, different word order. Google’s diversity check flags it.

Force angle diversity in the prompt (the “distinct angle” line above does this). After generation, sort your headlines into buckets: benefit, proof, urgency, objection-handling, price, question, brand, keyword-match. More than two in the same bucket? Cut the weaker ones. You want coverage across as many angles as possible – that’s also what Google’s Ad Strength scoring rewards, and per Google’s reported figures, advertisers who improve Ad Strength from Poor to Excellent see 15% more clicks and conversions on average.

How long until you actually know if it worked

This is where most “ChatGPT vs. human copywriter” comparisons online fall apart. People run an ad for a week, see 40 clicks, declare a winner. That’s noise, not signal. According to Search Engine Land’s RSA reporting guidance, a single asset needs at least 100 clicks to begin evaluating its click-through rate and at least 100 conversions to properly evaluate its conversion rate or cost per acquisition.

Multiply that across 10 headlines and 4 descriptions and the data requirement is substantial – think months, not days, for most small-budget campaigns. You’ll get directional signal in 2-4 weeks and statistical signal only after 1-2 months. Plan accordingly. Don’t rewrite the entire ad after three days because one ChatGPT headline isn’t getting impressions yet.

When not to use ChatGPT for this

ChatGPT is good for variation, brainstorming, and breaking writer’s block. Not useful in three situations:

• Highly regulated verticals. Finance, health, legal – the model invents claims (“FDA-approved,” “guaranteed returns”) that will get your account suspended. Human review isn’t optional, it’s the entire workflow.
• Brand voice that’s already strong. If your brand has a distinct, established voice, ChatGPT will sand it flat into the same descriptors Position2 found in their test. You’ll get clicks, but they’ll feel like everyone else’s clicks.
• Local/hyperspecific offers. ChatGPT doesn’t know your zip code, your competitor’s prices on Tuesday, or that your town’s high school just won state. The hyperlocal hook beats the AI hook on local search every time.

One more honest note: the Position2 author concluded that the ChatGPT ads didn’t suck – they were better than many copywriters they’d worked with, and could be brought up to acceptable standard in under 10 minutes. That’s the realistic bar. Not “replaces a copywriter.” Not “generates campaign-ready output.” A solid first draft that needs ten minutes of human editing.

FAQ

Does the free version of ChatGPT work for this, or do I need Plus?

Free works fine for ad copy generation. The Plus tier (reported at $20/month as of early 2025, per third-party sources – check OpenAI’s current pricing) mainly helps if you’re uploading CSVs of past campaign performance for the model to analyze, or chaining many prompts in one session without rate limits.

Can ChatGPT just write my whole RSA and I’ll paste it in?

Technically yes, practically no. Picture this: you paste 15 headlines into Google Ads Editor. Three are over 30 characters and get rejected. Four use the same three descriptors and trigger the diversity flag. One contains an emoji and gets the ad auto-rejected. You’re now editing seven of fifteen headlines by hand – which is most of them. Treat ChatGPT output as a first draft requiring mechanical verification (character counts, emoji removal, diversity check) before it touches your account.

How does ChatGPT compare to Google’s built-in AI suggestions in the Ads interface?

ChatGPT produces more variety and stranger angles – including bad ones. Google’s in-product suggestions are safer and more on-policy, but they tend to be derivative of your existing copy. Use both: Google’s suggestions for safe baseline assets, ChatGPT for the unusual angles you’d never write yourself. Human judgment picks the survivors.

Your next move: open ChatGPT, grab the prompt above, and generate 25 headlines for your current top ad group. Run every output through a character counter before anything else. Count how many violate the 30-character rule – that number is your baseline for how much you can trust the model’s self-reporting on the next campaign.