What to check on an AI-generated ad before it goes live

Check these nine things, roughly in this order: every word of visible text, whether the product is actually yours, hands and counts, the safe zone at real canvas size, whether the frame was composed or cropped, brand palette and typography, any claim or number the image makes, legibility at thumbnail size, and whether two creatives in the batch are secretly the same creative. It takes about ninety seconds per ad and catches most of what gets pulled after launch.

The failure mode nobody plans for with generated creative is not that it looks bad. Bad is easy — you see it and discard it. The problem is the ad that looks completely fine in a review tool at 400 pixels wide and is wrong in a way you only notice after it has spent money.

What follows is a review pass, not a style guide. Every item is something that has visibly gone wrong on real campaigns.

1. Read every word out loud

Text is the single most common defect in generated imagery, and it is the one your eye is worst at catching, because reading is pattern-matching and a plausible word shape reads as the word you expected.

So read it aloud, character by character, including:

• the headline
• any badge, sticker or burst
• the small print
• anything printed on the product itself — packaging copy, a label, a screen

Image models build the whole picture at once and aim for plausible, and a plausible-looking letter is not the right letter. Long headlines, small type, busy backgrounds and non-Latin scripts fail most often. If you are running a language you do not read, this check is not optional and it is not something you can eyeball — get a speaker of that language to look.

2. Is that actually your product?

This is the check that separates a generated ad from a photographed one, and the one most likely to embarrass you.

Look at the object, not the composition:

• Count the parts. Buttons, straps, wheels, compartments, pockets. Models add and remove them freely.
• Check the closure. A zip that became a button, a pump that became a spray, a screw cap that became a flip top.
• Read the label. Not "is there a label" — read it.
• Check the logo shape. Not the colour, the actual letterforms and mark.
• Check proportions against a real photo. Side by side, not from memory.

If your ad shows a product that does not exist, the problem is not aesthetic.

3. Hands, counts and symmetry

The classic artefacts, still worth thirty seconds:

• fingers and thumbs — count them
• teeth, and the alignment of both eyes
• jewellery and watches that merge into skin
• repeated objects that should match and do not — a pair of shoes, two chairs, four wheels
• reflections and shadows that disagree with the light source
• text or objects that continue through something they should be behind

4. Check the safe zone at the real canvas size

For anything vertical, the platform's interface sits on top of your creative. Meta states its safe zones as percentages, not pixels — the top 14%, the bottom 35% and 6% of each side on a 9:16 frame — which means a pixel figure from a guide is only right at one canvas size.

Look at the ad at the size it will actually run, with the reserved bands drawn on it. What you are looking for:

• a headline sitting in the bottom third, where the caption and CTA button go
• a logo in the top band, behind the profile row
• anything important within 6% of a side edge

The bottom band is bigger than almost everyone expects. That is the one that catches people.

5. Composed for the frame, or cropped into it?

A 9:16 ad is not a 4:5 ad on a taller canvas. If a creative was made at one ratio and resized into another, it shows: a subject drifting off-centre, a product cropped at the edge, an empty band of background where the composition ran out, text that was comfortable in the original and is now crowded.

Cropping one master into every placement is the habit this check exists to catch. The tell is usually negative space that is doing nothing.

6. Brand palette, typography and interfaces that do not exist

Three separate failures that all present as "it looks a bit off":

• Colour drift. Compare a flat swatch of your brand colour against the ad, side by side. Close-but-wrong is worse than obviously wrong, because nobody catches it until a customer does.
• Invented typefaces. A model asked for "a modern sans" will produce something that is not your typeface and is not any typeface.
• Fabricated product UI. If the ad shows a screenshot, a dashboard or an app screen, check that every element exists in your actual product. Shipping an ad containing an interface you do not have is a specific, recurring, entirely avoidable embarrassment.

7. Every claim and every number

Generated copy invents specifics with total confidence. Before launch, confirm against a real source:

• percentages and statistics
• review counts and star ratings
• prices, discounts and "was" prices
• guarantees, warranties and returns windows
• superlatives — "the best", "the only", "number one"
• health, financial and income claims, which are separately regulated nearly everywhere
• awards, certifications and press mentions

A number in an ad is a claim you are making. If nobody can point at where it came from, take it out.

8. Look at it small

Open the ad at roughly thumbnail size, or just step back from the screen. Then ask:

• can you tell what the product is?
• is the headline readable?
• is the CTA distinguishable from the background?
• does the whole thing still read as one image, or turn to mush?

Most people will see your ad smaller and for less time than you are looking at it now.

9. Are two of these the same ad?

The batch-level check, and the one people skip because it feels like a strategy question rather than a QA one.

Describe each creative in one sentence, without naming its format or colours. If two sentences come out the same, you have one creative twice. That matters beyond tidiness: Meta's own guidance is about diversified ads — a wide range of themes, messages and visuals — and five executions of one idea satisfy any volume target while diversifying nothing.

Two habits worth more than the checklist

Measure your usable rate. Of everything generated for a campaign, what fraction actually shipped? That single fraction tells you more about whether a tool is working for you than any output-per-hour number, and it is the figure teams who run a lot of generated creative end up tracking.

Do this before the review meeting, not in it. These checks are fast and mechanical. Making three people find the same missing thumb in a call is the expensive version.

---

None of this is an argument against generated creative. It is an argument against shipping it unread — which is a different thing, and the one that actually costs money.

Terrads composes each placement for its own frame and carries your real brand assets into the render, which removes some of these checks. It does not remove the habit of looking.