A Week of AI Ad Creatives: A Worked Example, Day by Day

Plenty of posts about AI ad creatives promise a story of fast results. This one is different. It is a worked example, and it is made up. The bakery, the choices and the numbers are invented to show what a sensible week could look like. Nothing here is a real result, and no figures are reported.

The point is the shape of the week: what to do each day, where things tend to go wrong, and how to read the outcome honestly. The same sequence is taught on screen in Lesson 4.5: Make ad creatives in minutes.

A good first week is not about winning. It is about producing a clean comparison and writing down one thing you learned.

The setup

Say you run "Hearth Lane Bakery", a one person shop that sells weekend sourdough boxes within a few miles. You have an Instagram page, a simple website and a small ad budget you can afford to lose. You have never run a proper ad test.

Your goal for the week is modest: launch five versions of one ad, compare them fairly, and learn which kind of picture makes people click.

Day 1: write the Look and choose the offer

You start with no images at all. First, you write two short things.

The offer sentence: "Weekend sourdough boxes for families nearby, order by Thursday, collect Saturday."

The Look note: "Warm cream and deep green, friendly and rustic, close ups on a wooden table in soft daylight. No text in the image. No faces."

This takes under an hour and feels like procrastination. It is the part that stops your ten images looking like ten different bakeries.

Day 2: make the batch

You generate images using the Look note. You ask for ten different ideas across five types: close up of the loaf, the box on a doorstep, a family table, an empty bread bin as the problem scene, and a plain green background with space for words.

You end up with around fifteen. You throw out five. One loaf has a crust that looks plastic. Two show a pastry you do not sell. One has nonsense lettering on the box. One looks like a different bakery. This is normal and healthy. Cutting is the job.

You also make tall versions of the best five, since Meta recommends 4:5 vertical for single image ads on Facebook Feed and lists 9:16 for Stories and Reels (aspect ratio guidance).

Day 3: write the words and build the test

You write five headlines from five angles for your three favourite ideas. For the first test, you will only use one headline, "Fresh bread on your table by Saturday", so the headline is the same everywhere and only the image changes.

You build the test so that:

• The audience, budget, placement and link are identical across all five ads.
• Only the image differs.
• The test runs for the full period you chose.

Meta's guidance on A/B testing says to test one variable so your ad sets are identical except for it, and recommends a minimum of 7 days (source).

Before launch you check the page. Does the website say the same thing as the ad, in the first few seconds? If the ad promises Saturday collection and the page does not mention it, you fix the page first. If your website is not ready, Lesson 4.1 covers building one.

Days 4 to 9: leave it alone

This is the hardest part and it takes most of the test period. You will want to peek, tweak the budget, swap an image that looks weak after two days. Do not. A change mid test muddies what you learn.

What you can do is watch for real problems:

• Is the ad delivering at all?
• Did any ad get rejected for a policy reason?
• Does the link work?

Fix only genuine faults. If you must fix one, write down what and when, so you can account for it later.

Day 10: read the result

You open the results. Because this is an invented example, imagine the following pattern, with no numbers attached: the close up of the loaf drew the most link clicks, the family table drew the fewest, and the other three sat in between.

You check a few details before drawing a conclusion:

• Did each ad get enough views to compare? An ad shown to very few people proves little.
• Meta defines frequency as the average number of times each account saw your ad (source). If one ad was shown to the same people again and again, it may look worse than it is, or vice versa.
• Did clicks turn into orders? An image can win clicks and still lose sales.

More on which numbers to read is in measuring AI ad creatives.

Day 11: write the record and plan round two

You write one line: "In round one, close up product shots drew more clicks than lifestyle scenes. Next round: keep the close up, test five headlines."

You add to the Look note: "Prefer close ups. Avoid scenes with extra pastries." Then you plan round two, which tests headlines on the winning image.

That is the whole week. No miracle, no overnight win. You finish with a small piece of knowledge that is specific to your customers and costs you nothing more to keep.

What could have gone wrong

A real week often has hiccups. Here is what to watch for and the fix:

Why an invented example

Many "case studies" in this area are written to sell something, and numbers in them are often impossible to check. Rather than add to that pile, this post shows the method and leaves the numbers to you. When you run your own week, your results will be real and worth more than any example.

Your own week

Copy the structure. Pick one offer. Write the offer sentence and the Look note on day one. Make a batch on day two. Build the clean test on day three. Wait. Read and record.

If you want the lesson that takes you through it with the screens shown, Lesson 4.5 is here: Make ad creatives in minutes. For a reusable process behind it, see a repeatable framework for AI ad creatives.

Questions people ask

Is the bakery in this post real?

No. It is invented for illustration, and the post reports no real results.

How long should a first test run?

Meta recommends at least 7 days for A/B tests. If your customers take longer to buy, allow longer.

What if I only have time for two ads?

Two is better than one. Change only one thing between them and you still learn something.

Should I change the budget during the test?

Avoid it. Changes can muddy the comparison. Keep budget, audience and link the same until the test ends.