AI Facebook ads, written and rendered for you

The gap between a draft and an ad

Ask any chat model for a Facebook ad and it will give you one in four seconds. It will be competent, it will be generic, and it will be about fifteen minutes of the two days that actually stand between you and a campaign that is live.

The rest of those two days looks like this.

• Working out which kind of person to write to, which nobody does properly because it is slow.
• Writing four more versions, because one angle is not a test.
• Finding an image. Then finding a different image, because you cannot use one image forty times.
• Getting the words onto the image without opening a design tool.
• Deciding where to put the link, and remembering that in the body it costs you reach.
• Rewriting the whole thing thirty times so no two groups get identical text.
• Finding the groups. Posting to the groups. Not getting throttled while posting to the groups.

An AI ad generator that hands you a paragraph has done step zero. The studio in Serious Poster exists because every step on that list is either a research problem, a writing problem, a rendering problem or a posting problem, and all four are now things a machine does acceptably well.

What the generator actually runs

It is three model calls in a fixed order, and knowing the order tells you where to intervene when something comes back wrong.

A generation runs detached from your browser, because twenty images take minutes and nothing good happens when a page waits that long. Close the tab. Come back to a finished set. If it fails at ad fourteen, the first thirteen are already saved and usable, and you are only charged for what was made.

Stage one: who is worth talking to

The targeting pass is the stage people skip and the one that decides everything downstream. An ad written for small business owners is an ad written for nobody. An ad written for a plumber with two vans who just lost his part-time bookkeeper is an ad that lands.

The generator returns segments at that second resolution: a portrait of one real person in the segment, what they are afraid of, what they want, what they would say to refuse you, the words they use for the problem, and an honest fit score so the weak ones are visibly weak. It also returns the phrases those people would use to search Facebook, which is what feeds the group finder later.

The full detail of that stage, including how to edit a segment it got slightly wrong, is in the guide to AI audience research.

Stage two: the copy, written for groups

Generic ad copy fails in a Facebook group for reasons that have nothing to do with quality. A group feed is not an ad slot. It is a room of people who joined to talk to each other, and an ad-shaped thing arriving in it reads as an intrusion before anybody has judged the offer.

So the copy prompt carries the rules of that room, not the rules of a landing page.

• The first two lines are the whole ad. Everything after them sits behind a See More that most people never press. The hook is written as its own field for exactly that reason.
• No link in the body. The link goes in the first comment, and the copy stage writes that comment as a separate piece of text rather than leaving you to paste a URL under your own post.
• Member voice, not brand voice. No campaign slogans, no third person, no stack of exclamation marks. It writes the way somebody in the group would write if they happened to sell the thing.
• One ask. A post asking you to comment, message, click and share gets none of those.
• Spintax baked in. Every ad arrives with {a|b|c} groups already written into the natural variation points, so one ad is hundreds of different posts. That is covered in full in the spintax guide.
• No em dashes. A small thing that is a large tell, so the prompt forbids them outright.

Each audience gets its own call, which is not an efficiency detail. Ask one model call for twenty ads and you get one ad written twenty ways. Ask it five times for four ads, each time with a different person in mind, and the angles genuinely diverge.

Stage three: the creative, rendered

The copy stage does not simply describe a mood for the image. It writes the exact words that should appear on the picture, in reading order, and a complete scene description saying where each one sits. The render stage is handed both.

That ordering is why the text on the creative is legible and on-message rather than the melted lettering image models are known for. The words were decided by the model that was writing the ad, and the image model is told to place them, not to invent them.

The AI ad creative generator guide goes through the render stage properly: what makes a good source image, how brand imagery is used, how many creatives a campaign actually needs, and why one image reused forty times is a reach problem rather than a taste problem.

Where the ads go afterwards

Every finished ad becomes a template, which is the same object you get when you write a post by hand. Nothing about it is a special AI object with its own rules. That means the entire poster applies to it immediately: media sets, group lists, randomised delays, per-group cooldowns, active hours, warmup, and the safety stop.

It also means the AI half and the posting half can be used separately with no friction. Generate ads and post them by hand. Write everything yourself and let it post. Or, most commonly, generate a batch, rewrite two of them because you know your customers better than any model does, and run the lot.

What you give it, and what that changes

The single largest quality lever is the business profile, and it is the one people fill in lazily. The model cannot write a specific ad from a vague description, and when it lacks a fact it will reach for a plausible one, which is how invented guarantees end up in your copy.

Briefs are the second lever. A brief is an additive instruction that rides along with every generation for that business, so we are pushing the winter service plan this quarter, or never mention financing, gets applied without being retyped into every run.

What AI ads are still bad at

Being straight about this is cheaper for both of us than a refund, so here is the honest list.

• It does not know your customers. It knows the category. The first batch is usually right about the shape of the market and wrong about one specific thing you know from ten years of doing this. Correct it in the profile, not in your head.
• It cannot make a bad offer work. Nothing can. Pacing protects reach, copy earns attention, and neither manufactures demand that is not there.
• It has no taste about your industry. It will occasionally write something technically fine that anybody in your trade would find slightly off. You will spot those instantly, which is why you read them before posting.
• Images are drawn, not photographed. A rendered creative is right for a concept, an offer card or an illustrated scene. It is not right for showing the actual house you just listed. Use your own photos in the media set for those, mixed with the generated ones.
• It does not read group rules. Groups that ban promotion still ban promotion, and a beautifully generated ad breaking a rule is still a removed post.

What it costs and what a unit is

Posting is free. Five hundred posts a month on a free account, unlimited on a paid one. The AI is what the plans meter, because every generated ad involves a real image render and a real reasoning call, and those cost money whether or not anyone posts the result.

There are two units, and they are counted rather than priced at you in the interface.

A free account gets a one-off credit of twenty ad creatives and three audiences, which is a real generation rather than a demo. Starter is $19 a month for forty creatives and fifteen audiences, Pro is $69 for a hundred and fifty creatives and fifty audiences, and both post without limit. Anything past a paid allowance is metered overage at a per-creative rate rather than a wall.

Same rule as posting: you are charged for what was produced. A run that stops at ad six has spent six.

Still stuck? Mail us and a person will answer.