Text-on-Image Ads: Turn One AI Sign Photo Into Dozens of Message Tests

What is a text-on-image ad, and why does it work now?

Picture a road sign on an empty highway. Big white letters on green. It says the exact thing your buyer was thinking this morning.

Nobody scrolls past a road sign. It looks like a photo, not an ad. That is the whole trick.

A text-on-image ad is an ad where the words are part of the picture. A sign, a sticky note, a phone alert, a tattoo. The words are in the scene, not slapped on top.

Facebook once capped text at 20 percent of an ad image. An upload tool checked every picture. A gym photo could fail just because of slogans painted on the back wall.

That rule is gone. Meta's own best-practice page now says there is no longer a limit on text in ad images.

Meanwhile, feeds filled up with pictures that have words on them. Quotes, screenshots, notes to friends. A picture with words on it stopped looking like an ad.

That matters, because people skip anything that looks like an ad. It has a name: banner blindness. Eye-tracking research shows people ignore content that even resembles an ad.

A sign in a photo slips past that filter. It reads like native content, the kind of post that fits the feed it sits in.

The ad that looks least like an ad gets read the most.

Words carry the message when the picture cannot

Some offers sell with a picture alone. A dating offer shows a happy couple. A fitness offer shows someone in great shape.

Most offers cannot. What does a tax service look like?

Or guitar lessons for adults? Or chocolate imported from Austria?

For those offers, the words do the work. The picture only has to look real and stop the thumb.

For example, one guitar teaching brand paired its words with things guitar players stop for.

A guitar. A scale chart worth studying. An old blues player, with a short line of words beside him.

There is a second reason we like the format. A sign ad is the cheapest creative you can make. So it is also the cheapest way to test a message.

Cheap tests mean more tests. More tests mean faster answers. Faster answers mean you stop paying to guess.

Most creative advice is about prettier pictures. Pretty pictures are slow, they cost real money, and each one tests a single idea.

A gorgeous image with the wrong words is still the wrong ad.

Pick a sign that belongs in your buyer's world

The sign is the center of the picture. Pick one your buyer would actually see on a normal day.

The better you know your buyers, the better your ideas get. A trucker sees a trailer. A desk worker sees a sticky note on the monitor.

Someone with shoulder pain might see the words written on the sore shoulder.

Other surfaces that work: road signs, billboards, a sandwich board on a busy street, a phone alert, a desktop pop-up, a t-shirt, a bumper sticker, a greeting card. A metal sign on the product itself works too.

Some surfaces look fake fast. Skywriting comes out blocky.

Cake lettering and handwriting can look pasted on. Sticky notes often come out tiny.

And a sign where no sign belongs looks corny. Would this sign exist in real life? If not, skip it.

Want more ideas? Search the Meta Ad Library for your niche. Note the sign ads that keep running for months.

Those are probably paying for themselves.

A real scene gets read. A fake one gets skipped.

How do you make an AI sign image? One base prompt and a few fixes

Open your preferred AI image tool and ask for six things, in this order.

First: photorealistic. The sign type. The exact words, in quotes.

Then: the sign as the main focus, easy to read. Zoomed in on the sign. The shape your ad placement needs, at high resolution.

Typing it takes seconds. The picture takes a minute or two.

Image tools write letters far better than they used to. They still slip. The most common slip is extra words you never asked for, sometimes a web address for a site that does not exist.

Fix it with one short line: remove the extra text and use only my words.

Fix framing the same way. Zoom in. Make it look like a real alert on a real phone.

Same mistake three times? Stop arguing. Open a new chat and paste the first prompt again.

A fresh chat starts from zero. The old chat keeps digging the same hole.

Found one you love? Download it right away. Later, upload that file to a new chat and ask for the same picture with new words.

Meta sometimes changes its image shapes and the parts of the screen it covers. The box below holds Meta's current rules, with the date we checked.

Lock the picture, then change only the words

This is where one picture becomes thirty ads. Once the image looks right, keep it.

Then tell the tool: keep everything the same and change the words to this. Paste the next line. Repeat.

Same scene, same light, same angle. Only the message changes. So when one ad beats another, the words did it.

For example, one advertiser tested about 60 headlines on a single template image for one brand, and about 30 for another. The winning lines then went into videos and other ads.

Sometimes the tool changes more than you asked. Check each new image against the first. If the scene drifted, go back to the saved original.

Prefer clean type? Generate the scene with no words, then set the words yourself in a design tool. We do this for our own images, because a typo then takes ten seconds to fix.

One catch: asking a tool to erase the words from a finished image often fails. It hands back the same picture. Ask for no words from the start.

Keep a running list of lines that won and lines that lost. That list becomes your message bank. Every new video, email and page can draw from it.

For the bigger system around it, read how to run a creative testing engine.

Write 100 lines before you open the image tool

The slow part is not the picture. It is the words. So write them first, in a plain doc, with no image tool open.

Aim for 100 rough lines. Cut them to the best 30 or so. After that, the image work is just pasting.

Where do 100 lines come from? Ingredients. Every line is built from parts you can swap.

Twelve parts cover most of it. How aware the buyer is. The outcome, and whether it comes soon or later.

The pain, and the limit they think they have.

What makes you different, and your proof. The objection, and how your product works. Who it is for, the mood, clever or clear, and your brand phrases.

What does aware mean here? Some buyers have never noticed the problem. Others are comparing you to two rivals.

More on that in how to reach the unaware majority.

Our rule: when you want to know what worked, change one ingredient per ad. Change three at once and a win teaches you nothing. More in one variable at a time.

This is just A/B testing, which means comparing two versions, with cheaper ads.

Pain lines have a trap. Meta bans lines that say or hint that the reader has a health or money problem. The box above has the rule, with Meta's own examples.

Keep the sign short: the picture stops the scroll, the ad text sells

A sign has one job. Stop the scroll. It does not have to say everything.

For example, one premium chocolate importer, at about $12 a bar, ran a handwritten note packed with details. Her ads had worked for about seven months, then faded within a week or two.

The better version cut the note to five words: chocolate that actually means something. The details moved into the ad text above the picture.

Each detail can also become its own sign ad: 170-plus flavors, organic and fair trade, made by hand in Austria.

Picture a tattoo with 40 words on it. Nobody reads it. Nobody believes it either.

Check every image on a phone before it runs. A line that reads fine on a laptop can be too small to read on a phone. Older buyers struggle more, and legibility drops fast as letters shrink.

How to waste a good sign in four easy steps:

1. Write your whole sales page on it.

2. Check it on a 27-inch monitor and nowhere else.

3. Stick it somewhere no sign would ever be.

4. Make it a fake warning sign and call it creative.

Fake warning signs get clicks, so number four is tempting. We still do not run them, because a fake warning is a lie in a safety vest.

Same with quotes. Only real customer words go on a sign. Everything else is bullshit with good lighting.

How do you know which line won? Read sales first, clicks second

Run the lines. Then read the numbers in order, from the click to the sale.

Clicks tell you the sign stopped people. A rough aim is about 1 percent of viewers clicking. It is a clue, not a rule, because a low click rate can still bring serious buyers.

Next, check how many clicks reach the page. Of the people who click, about 85 percent should reach a loaded page. If it is much lower, the problem is the page or the link, not the sign.

Then sales. A line that sells stays, whatever its click rate says. A line with great clicks and no sales goes.

Cheap clicks are not good clicks. Good clicks are the ones that buy.

So never kill a selling ad over one weak number. And never save a dead ad because one number looks pretty.

Clicks but no sales? The sign did its job and the page did not. Read how to match your page to your ad.

Not sure which number is broken? Use our guide to reading ad data, one step at a time.

Build your sign ads with AI this week

Paste the prompt below into your preferred AI. It asks what you sell and where your buyer spends the day. Then it writes the scenes, the 100 lines, the image prompt and the swap line.

It never invents a result. It uses only the facts you give it.

1. Run the prompt. Keep the best 30 lines.

2. Make one scene and check it on a phone. Then swap in the words.

3. Launch, read sales first, and add every result to your message bank.

Remember the road sign? Same sign, thirty messages. The words do the selling.

Next problem: writing the hook that becomes your best line.

Stop making more pictures for your words. Start making more words for your picture.