The AI Image QA Checklist Marketers Should Run Before Anything Ships

Search for an AI image checklist and you get two kinds. E-commerce lists about product fidelity, written for people shipping fifty SKU variants. And generic lists that say check the hands, check the eyes, check for extra fingers. Neither is written for the person at a $5M to $100M company who has to approve a LinkedIn tile, an email hero and a blog cover before Thursday and has no designer to ask.

This is that person’s checklist. It is organized by where errors hide rather than by what they look like, and it is ordered by the cost of a miss rather than by how often a miss happens. A six-fingered hand is common and mildly embarrassing. A warped logo on a co-marketing image with a partner is rare and a real problem. The order below reflects that.

The zoom rule, first

Before the layers, one habit that catches more than any single check: inspect at 200 to 400 percent zoom. The failures that cost the most (logo geometry, product proportions, label text, fine type) are sub-perceptual at feed size. A tile that looks fine at 400 pixels in a preview carries a wordmark that is 6 percent too wide and a shade too warm, and that is invisible until it is on a laptop, a slide or a printed leave-behind. Every brand element and every piece of rendered text gets zoomed before approval. It takes twenty seconds and it is the difference between a check and a glance.

Layer 1: brand

Highest cost, first check. The elements that make the image yours are the elements a model reproduces worst.

• Logo. Is it the real file, composited after generation, or did the model draw it? If the model drew it, it is wrong; the logo post explains why there is no acceptable version of a model-drawn mark. Zoom in. Check letterforms, proportions, clear space, color.
• Colors. Compare the brand colors in the image against the hex values, not against memory. Models pull palettes toward the scene; the blue on the wall is teal more often than you would think.
• Fonts. Any text in your typeface should be live text set in the real font, not model-rendered lettering that resembles it.
• Visual system. Does it look like the last ten images you published? The consistency post has the drift budget; check the image against it.

The good news is that this whole layer can be removed as a check rather than performed, by never letting the model touch it: generate the scene, then composite the logo and apply colors and fonts as locked assets. If your workflow does that, layer 1 is a confirmation, not an inspection.

Layer 2: truth

Does the image show something that is not true, or that a reasonable viewer would take as a claim?

• The product. Is that our product, at the right proportions, with the right label, the right number of ports, the right cap? A model will happily invent a feature. If the image shows the product, the product must be right, which usually means a real photo or a locked reference.
• The claim. An image can make a claim words never did. A before-and-after that did not happen. A dashboard with a number nobody achieved. A crowd at an event that had forty people. If the image implies a result, the result needs the same substantiation as a sentence would.
• The place. Generated storefronts, offices and job sites carry details (signage, landmarks, plates) that can place them somewhere real and wrong.
• The people. Does the person look like a specific real person? Models produce faces that resemble public figures and, occasionally, your customers. If a face is recognizable as someone, it is a rights question (layer 5) and a truth question at once.

Layer 3: text

Anything the model rendered as words.

• Every letter, every word, every number. Spell-check by eye, at zoom. Text rendering has improved enormously, and the errors that remain are the confident ones: a real-looking word with one wrong letter, a price with a transposed digit, a phone number that does not exist.
• Background text: signs, labels, screens, papers on a desk. Models fill these with plausible gibberish, and gibberish on a wall behind your founder reads as a tell to anyone primed to spot AI images.
• The rule that removes most of this layer: any text that matters is overlaid as live text after generation, and the prompt asks for no text in the scene. The text-in-images post covers when to let the model render type and when not to.

Layer 4: anatomy and physics

The classic checks, cheapest to run, lowest cost when missed, still worth a pass because they are the ones viewers screenshot.

• Hands, fingers, teeth, ears, the join between limbs and bodies.
• Eyes: direction, symmetry, reflections that match the scene.
• Reflections and shadows: does the mirror show what it should, does the shadow fall with the light.
• Objects: a chair with five legs, a keyboard with two spacebars, a van with the door in the wrong place, a coffee cup that merges into the table.
• Edges: hair, glass, fabric against background, where diffusion artifacts cluster.

If the image passed layers 1 to 3 and fails here, regenerate from the continuity sheet, not from the failed image.

Layer 5: rights and disclosure

The layer most checklists skip and the one that has changed most in the last year.

• Likeness. Any face that resembles a real person, any voice or body that could be read as a specific individual. Rules on synthetic likeness in advertising have tightened: New York’s requirement to disclose synthetic performers in ads took effect in June 2026 and other states are following. Treat any image of a person as a disclosure question until you have checked the rules where you publish.
• Other people’s marks. A partner’s logo, a competitor’s product, a recognizable third-party brand in the scene. Distorted or not, it is a trademark question; the logo post covers the Getty v Stability rulings that made that explicit.
• Platform labeling. Some marketplaces and ad platforms now require AI-generated imagery to be labeled, and the requirements change. Check the platform’s current policy before publishing, not the one you remember.
• Provenance record. Keep a note of how each image was made: the model, the references, whether a real photo was used. When a question comes later, the answer should be a lookup, not a reconstruction. The commercial-use post covers licensing terms by model.

Layer 6: placement

The image is correct. Does it survive where it is going?

• The crop. LinkedIn, email clients, X, and Open Graph previews each crop differently. Is the subject inside every safe zone? Is the logo in a corner that survives the crop?
• The size. At the channel’s display size, is the text legible, is the logo readable, does the detail that justified the image still show? An image built for a blog hero often dies as a 600-pixel email header.
• The context. Next to the headline and the call to action, does the image say the same thing? An upbeat image on a churn-prevention email is a mismatch a checklist catches and a glance does not.
• The template. If the image is going into a slot in a branded layout, does it need a logo at all, or does the template already carry one? A second logo is clutter.

The checklist, on one screen

Where this lives in a workflow

A checklist that lives in a doc gets skipped on Thursday at 6pm. It has to sit at a gate the image cannot pass without.

In Marqeable, layer 1 is removed by construction: the model never draws the logo, the real file is composited with placement rules, and colors and fonts are applied as locked assets from your brand look. Layer 3 is reduced the same way, with important text overlaid live. Layers 2, 4, 5 and 6 are what the human review gate is for: every image sits at a ready gate with the original and the composited version side by side, and nothing ships until a person clears it. The checklist is the gate’s rubric. The gate is what makes it happen every time.

Frequently asked questions

What should I check before publishing an AI-generated image?

Six layers in order of cost: brand (logo, colors, fonts), truth (product, claims, place, people), text, anatomy and physics, rights and disclosure, placement. Zoom to 200 to 400 percent on every brand element and every piece of text.

Why check at zoom?

Because logo geometry, product detail and label text fail sub-perceptually at feed size and show at laptop, slide or print size.

Do I need to disclose that an image is AI-generated?

It depends on the platform and what the image shows. Some marketplaces require labels on AI product imagery, and some jurisdictions require disclosure when a synthetic performer stands in for a person. Treat any image of a person as a disclosure question and check current policy where you publish.

Can the checklist be automated?

Partly. The brand layer disappears when the logo is composited and colors and fonts are locked after generation. Text can be caught by a review pass. Truth, rights and placement still need a person at a gate.

The bottom line

The expensive failures in AI marketing images are quiet ones: a logo that is almost right, a product with a feature it does not have, a face that belongs to someone, an image that is perfect and gets cropped wrong. Check by layer, order by cost, zoom on everything that carries the brand or a word, and put the checklist at a gate the image cannot skip. Then remove the most expensive layer altogether by never letting the model draw the parts that have to be exact.

Marqeable runs your campaigns, answers every visitor, text, and email in seconds, and turns them into booked jobs and meetings - even at 9pm on a Saturday. We’re in private beta with a small early cohort. Get early access