Creating ads with AI: from the brief to live variants
You open ChatGPT, type “write me ads for my coffee shop,” and a few seconds later you have a list of headlines. Almost none of them fit in Google Ads, several promise things you don’t sell, and they all sound like a brand that isn’t yours. The model was missing almost everything you know about your business, and nobody took a careful look at what it returned.
This guide covers both sides: how to ask for the copy so it comes out usable, and what to do with it afterward, all the way to turning it into a workflow that runs every week without copying and pasting.
How to ask AI for ad copy: campaign, offer, and format
A language model doesn’t know who you sell to, what your margins are, or what your lawyer told you not to say last year. Whatever you don’t tell it, it fills in with the most likely answer, and the most likely answer is generic. That’s why the prompt that works looks more like a brief than a question.
That brief has three layers. The campaign layer says what the ad is for and at what point in the buying journey the customer will see it, because you don’t write the same way for someone searching “buy moka pot” as for someone scrolling Instagram. The offer layer says what you sell, at what price, with what shipping, and what sets you apart, with facts, not adjectives. The format layer says which platform it runs on, how many pieces you want, and what the limits are.
A Google Ads template that stays within character limits
Google’s responsive search ads accept up to 15 headlines of 30 characters and 4 descriptions of 90, plus two 15-character path fields. Google combines the pieces on its own, so every headline has to work alone and next to any other.
- You are a search ad copywriter for [business], which sells [product] to [typical customer] in [area].
- Campaign: [goal] · ad group keyword: [keyword] · landing page: [URL].
- Offer: [price, terms, shipping, warranty]. Real differentiators: [list].
- Off-limits: [promises you can’t keep, words you don’t use, competitors].
- Write 15 headlines of 30 characters max and 4 descriptions of 90 characters max.
- At least 3 headlines include the keyword. None depends on another to make sense.
- Return it as a table with the text and its character count.
The count the model gives you back is no use as a check, because models read text in chunks and are bad at counting letters. It just nudges the model toward short copy. The real check happens in the spreadsheet with `=LEN(A2)`, and anything over 30 goes back to the model or gets trimmed by hand.
On Meta the limits are looser: the primary text gets cut off after a few lines, and where it’s cut depends on the placement, so there ask for a first sentence that makes sense on its own and check the preview for each placement.
Your brand voice: your own ads as examples and a negative list
“Friendly but professional” tells the model nothing, because that’s how almost every brand describes itself. Showing it works better: paste five or six of your own ads that performed well into the prompt, or copy from your website that represents you, and ask it to write in that style.
The other half is the negative list. Every business has words it doesn’t use, because they sound like a different industry, because a customer complained, or because they belong to a competitor. A secondhand furniture store might not want “low cost,” an accounting firm might never say “no paperwork.” That list goes into the prompt and also into the review, because the model will respect it almost always, and almost always isn’t enough. Keep the prompt, the examples, and the list in a shared document, because it’s the piece you’ll reuse the most.
General assistants, ad generators, and the AI built into Google Ads and Meta
General assistants, ChatGPT, Claude, or Gemini, accept a long prompt with examples, let you iterate on a specific variant, and work just as well for Google as for Meta. The paid plans for ChatGPT and Claude run about $20 a month, although for writing a small business’s ads the free plan is enough to get started. Their limit is that they know nothing about your ad account: you supply the context.
Dedicated generators, like HubSpot’s campaign assistant or Canva’s ad generator, come with the format already handled. You pick a platform, describe the product, and get back pieces already cut to size, often with an image. You skip the template but lose control over the tone, and it’s worth checking what the free plan includes before building anything on top of it, because it changes often.
Then there are the features built into the platforms themselves. In Google Ads, text customization writes headlines and descriptions based on your landing page, your ads, and your keywords. Since September 1, 2026, Google has moved search campaigns that had automatically created assets turned on over to AI Max, and in those campaigns text customization is on by default. You can turn it off, and if you want control over every line that runs under your name, check that setting before anything else in this guide. On Meta, Advantage+ creative generates variations of the primary text and headline, as well as backgrounds and images.
If you already pay for ChatGPT or Claude, we’d start there with the template, because that’s where you keep the most control over tone, and we’d only try a dedicated generator if you want something set up in an afternoon.
What to check before publishing: promises, facts, brand, and policies
A model that doesn’t know whether you ship within 24 hours may write “24-hour shipping” because that’s what ads in your industry say. It’s the most expensive mistake: you won’t spot it skimming, and it commits you to something with the customer. Every ad goes through four questions before it runs.
- Promises. Can you deliver the timelines, guarantees, and results it mentions? “Results from the first week” is a promise, even if it sounds like a stock phrase.
- Facts. Does every price, discount, or percentage exist, and is it on the landing page when the user clicks?
- Brand. Does it sound like you, or has a word from the negative list, a competitor, or someone else’s trademark slipped in?
- Policies. Would the platform approve it?
The last one deserves a closer look, because a disapproval stops the campaign in its tracks. Google doesn’t allow presenting an unlikely result as likely, advertising offers that aren’t available or aren’t easy to find on the landing page, or hiding the total cost the user will pay. Its editorial rules reject repeated punctuation, like “deal!!!”, and capital letters used to grab attention, which is exactly what a model produces when you ask for something punchier.

As for AI itself, Google requires a visible disclosure of synthetic content in election ads that depict people or events in a non-authentic way, a requirement that doesn’t extend to a regular commercial ad. Meta adds an “AI info” label to ads whose image or video was created or significantly edited with its generative tools, places it next to the sponsored label when a generated photorealistic person appears, and has announced it will also detect images from third-party tools using standards like C2PA.
Automating the workflow, from the brief sheet to ads uploaded as paused
For one campaign a quarter, all of the above is done by hand. When the pace picks up (weekly promotions, several product lines), the bottleneck stops being the writing and becomes copying, pasting, counting characters, and uploading.
The workflow has four pieces. A sheet where someone fills in the brief, one row per ad group. A call to the model that reads the row and writes the variants to another tab. A human review of that tab. And an upload of whatever was approved, always as paused, so that turning it on stays a person’s decision.
How to build it in n8n or Make and upload through each platform’s API
In n8n, the workflow starts with a Google Sheets trigger on rows marked “brief ready.” The next node calls the model with the prompt assembled from the columns (OpenAI, Anthropic, and Gemini each have their own node), and a code node measures each headline and flags the ones that run over before writing everything to the review tab. Keep in mind that the model’s API is billed by usage, separately from your chat subscription.
The upload comes with fine print. n8n’s Google Ads node only reads campaigns; it doesn’t create ads, so you have to use the HTTP Request node against the Google Ads API. In Make, the Google Ads app includes a module for making direct API calls that does the same thing. Either way you need a developer token, which at first only works with test accounts, and to touch your real account you have to request a higher access level from Google and wait for their review. Factor in that step before committing to dates.
Meta is more straightforward. Its Marketing API lets you create the ad as paused, and from n8n you call it with the Facebook Graph API node or with HTTP Request. The Google Ads API also supports paused ads, and its documentation recommends creating the campaign as paused until it’s ready.
n8n has a free community edition if you install it on your own server, and its cheapest cloud plan costs €20 a month billed annually, with 2,500 executions. Make has a free plan with 1,000 credits a month and a $9 plan with 10,000. For a workflow that runs a few times a week, the entry plan is more than enough.
Where human approval fits: in the sheet and on the platform
In two places. The first is the sheet: the upload workflow only reads rows marked “approved” in a column the model doesn’t write to. If more people share the sheet, protect that column so only the person who decides can edit it. That’s where the review checklist gets applied, and it’s the step you shouldn’t skip, because reviewing fifteen headlines in a table takes minutes, and in the Google Ads interface it takes a lot longer.
The second is the platform. Since everything arrives paused, whoever turns it on sees the finished ad, with preview, URL, and targeting, and errors you can’t see in a cell jump out: a headline that says something odd when combined with another, a link to the wrong page. On Meta, the ad also goes through the platform’s review before returning to the status you chose, and creating it as paused lets you decide when it goes live afterward.
What it costs and how long it takes, compared with paying an agency
Picture an online store that runs one promotion a month, with two or three ad groups on Google and a couple of ads on Meta. By hand, the cycle is filling in the brief, running the template through ChatGPT or Claude, moving the output into the sheet, reviewing, and uploading. Once the prompt is dialed in, that’s a few hours per promotion, almost all of it reviewing and uploading. In tools, about $20 a month, or nothing on the free plan.

Automated, the monthly work comes down to filling in the row and reviewing the variants. In exchange there’s a setup: the workflow in n8n or Make, the token paperwork, and a few test rounds until the output is stable, which is a few days for someone technically comfortable or a one-off project if you outsource it. After that you’re left with a tools and API bill that, for a small business, runs to a few dozen euros a month.
An agency sells something else. Published rates in Spain for managing Google Ads range from under €200 a month for small accounts to several hundred for more involved campaigns, and that fee pays for strategy, bidding, targeting, and analysis, not just the copy. If what you lack is campaign judgment, a prompt won’t fix it. If what eats up your time is producing and uploading variants of something you already know how to sell, that can be done in-house. That’s what we recommend almost every time: keep in-house whatever depends on knowing the business (the brief, the negative list, and the approval) and outsource only what you don’t know how to do.
If you’d like to see this workflow built on your own brief sheet and accounts, get in touch and we’ll look at it with you.