The AI Product Photo Check – Mathew's Library

01 The scene can be invented, the product can't

An outdoor store's automated ad showed one of its bicycles with two handlebars. A law firm, Reed Smith, wrote it up as a false advertising risk for the store, not the tool.

Making that kind of picture is now close to free. Google's fast image model, Nano Banana 2.1, makes a standard image for about 3 cents through its API, Google's paid access for developers. On the same day, Google published the model's known limitations in its model card:

Small text: "often blurry in 1k model". The 1K size is the 3 cent image. Consistency: "not always perfect between input images and generated output". "Occasional confusion" between left and right.

It is built to make a believable picture, not a correct one. A label that is 95 percent right is the one that ships.

Rule book What it says, as written

Google's shopping rules (Merchant Center) Your main product image must clearly display the exact item being sold. No mockups that don't show the physical product. AI images must carry a tag in the file saying so, and you should not remove it.

The US Federal Trade Commission (1983 test) Is it likely to mislead a reasonable customer about something that matters to the purchase? Nothing in the test asks what made the picture.

The EU AI Act (since August 2, 2026) The draft guidance asks whether people could take realistic AI content as genuine.

This is the rule as written, not legal advice.

The scene can be invented. The product can't.

02 What to shoot for real, what AI can build

Shoot it for real Let the AI build it

Your main listing image The background and the season

Anything with fine print The lifestyle scene

A pattern, a fit or a texture someone is paying for Square, tall and wide versions

The one true reference photo Ten variations for an ad test

What it costs. Industry price guides put a basic white background shot at $25 to $75 an image, and a shot with a model at $150 to $400. A hundred AI versions of one shot come to about $3.40 in image output.

The one file to protect. The model needs a real photo of your product to put it in a picture. That one true photo is the most valuable file you own.

03 How to make the AI version

Nano Banana 2.1 became generally available to developers on October 6. It is also in the Gemini app, AI Mode in Search, and Google's Flow and Stitch apps, and Google lists Google Ads among the places it is rolling out.

Attach your true photo as the reference image. It can take up to 14 reference images at once and keep up to 10 objects looking the same. Describe the scene in plain words, then keep editing. It edits the same image turn after turn. Pick the size. If your product has small text, skip the cheapest 1K size. 2K is about 5 cents through the API. Put your logo on in an editor. Use the exact artwork, never a redrawn one. When The Rundown asked the model to recolor a real logo, the first try came back the wrong shape.

A first prompt to adapt

Use the attached photo as the product. Keep its label, shape, size, color and pattern exactly as they are.
Put it on [SURFACE OR SCENE]. [LIGHT, e.g. morning light].

Next turns:
Now a wide banner.
Now a tall one for stories.

Two asterisks on the price. The 3 cents is only the image coming out. What you send in costs three times what it did on the old model. Run images as a batch and the output price halves again.

04 Three checks before any image goes out

Do these on every image, at full size, before it goes anywhere.

Read every word on the product, like a proofreader. Small text is where the cheap size goes soft. Raw and pure look the same in a feed.

Count the parts and check the shape and size. Caps, handles, buttons, handlebars. In one test, a horse came out closer to a pony.

Hold the color and the pattern next to the real thing. Viewers of earlier versions said a dress pattern changed.

Check left and right. Google's model card lists occasional confusion between them.

Keep the AI tag in the file when you export. Google's shopping rules say not to remove it. Every image from this model also carries Google's invisible SynthID watermark.

Write your product's must-be-exact details down once, and check every image against them.

Must-be-exact sheet

PRODUCT: [NAME]

Every word on it: [COPY THE LABEL WORD FOR WORD]
Parts to count: [e.g. 1 cap, 2 handles, 3 buttons]
Shape and size: [e.g. tall jar, fits in one hand]
Colors: [NAME EACH COLOR]
Pattern or texture: [DESCRIBE IT]
Left or right details: [e.g. handle on the right]

Real reference photo: [FILE NAME]
Checked by: [NAME] Date: [DATE]

Sources

Reed Smith: AI and advertising Google DeepMind: Nano Banana 2.1 model card Google: Gemini API release notes Google: Gemini Developer API pricing Google: Image generation in the Gemini API Google Merchant Center: Image link Google Merchant Center: AI-generated content FTC Policy Statement on Deception (1983) Baker McKenzie: New EU guidance on AI transparency The Rundown: Nano Banana 2.1 tutorial The Decoder: Nano Banana 2.1 test Squareshot: Per-image pricing explained

Mathew's LibraryLibrary card

Free access

Keep reading free with your library card

One email opens this checklist and all 3 resources in Mathew's library. No account, no password.