Google Merchant Center Image Requirements for AI-Generated Product Photos

Merchant Center needs each product's main image to be a clean photo of the exact item, at least 500 x 500 pixels from 31 January 2027 (1500 x 1500 recommended), under 16MB and 64 megapixels, with no text, watermarks or borders. If generative AI made the image, keep its IPTC digital source type metadata, usually TrainedAlgorithmicMedia.

The second half is where AI product photos fail, because the tag is invisible. Nobody sees a missing tag in the picture; they see it later as a warning in Diagnostics, or as products that quietly stop appearing. The checklist below is built so you find that problem before Google does.

Follow me on Instagram@sagnikteaches

The upload checklist, in the order to run it

Run these checks on the file that goes into the feed, not the file your design tool saved. The two often differ after a store platform, a content delivery network or an image-compression plugin has touched the image. Each line says what the rule is and how to verify it.

Connect on LinkedInSagnik Bhattacharya

Format and size

1. File type. Supported formats are JPEG, WebP, PNG, GIF, BMP and TIFF, and the extension must match the real format. A PNG renamed .jpg produces an "Unsupported image type" error. Check by opening the file in an editor, not by trusting the filename.
2. Pixel size. Google's page announces a minimum of 500 x 500 pixels for all products from 31 January 2027 and recommends around 1500 x 1500 or more. Aim for the recommendation now, so you do not have to redo the catalogue later. Look up the current minimum on the page the day you upload, because the transition period matters.
3. Weight and resolution ceiling. No image above 64 megapixels and no file above 16MB. An AI tool that upscales aggressively can breach this on large originals.
4. Link format. The image URL is 1 to 2,000 characters, ASCII only, with symbols and spaces encoded. A file name with an accented letter or a space is a common cause of a broken link.
5. Stable address. Use a URL that does not change on every feed submission. Timestamps and session IDs in the link make Google refetch images needlessly.
6. Crawl access. Make sure robots.txt allows both Googlebot and Googlebot-image, otherwise Google cannot fetch the file. Allow up to 24 to 48 hours for a robots change to take effect.

Content of the main image

1. No overlays. Google lists what to avoid: calls to action such as "Shop now", price or shipping text, warranty claims, promotional words like "Best", condition labels like "New", barcodes and QR codes, stock-photo watermarks, copyright symbols, and added brand or store logos (unless the logo is physically part of the product).
2. The real item. No placeholders, "coming soon" graphics, company logos instead of a photo, one generic image reused across different products, or illustrations that do not depict the physical product. A few categories, such as hardware, vehicle parts and computer software, are allowed generic images.
3. No border, no solid-colour square, no extras. Do not include products that are not sold together, and no single-colour image, except in a short list of paint and ink categories.
4. Framing. Google's best practice is for the product to fill no less than 75% and no more than 90% of the frame.
5. One variant per image. Submit a distinct image for each variant, showing only the colour or style being sold. Do not reuse one hero shot across every colourway.

The AI check

1. Was generative AI used to create any part of this image? If yes, the metadata rule applies.
2. Does the exported file still contain the IPTC digital source type field? Test it (method below).
3. Is the image used in image link, additional image link or lifestyle image link? The rule covers all three.

What the AI-image rule says, in plain words

Google's generated-content policy says that images created with generative AI must contain metadata showing they were AI-generated, for example the IPTC DigitalSourceType value TrainedAlgorithmicMedia. It also tells you not to remove embedded tags from images made with generative tools, naming Google's own Product Studio as an example. Three IPTC values are listed as ones to preserve:

Subscribe on YouTube@sagnikteaches

• TrainedAlgorithmicMedia: made using a model derived from sampled content. This is what most image generators produce.
• CompositeSynthetic: a composite that includes synthetic elements, such as a real product photo placed on a generated background.
• AlgorithmicMedia: created purely by an algorithm with no training data, such as an image produced by a mathematical formula.

The middle one matters most for small stores. A common workflow is to photograph a real item and have AI swap the background. IPTC's own vocabulary uses compositeSynthetic for a mix of elements where at least one is generative AI, and it also has a value Google's page does not list, compositeWithTrainedAlgorithmicMedia ("Edited using Generative AI"), for inpainting or outpainting a real photo. Which value a given editing tool writes, if any, is something to test rather than assume. Whatever value it wrote, keep it.

How to test that the tag survived

A free command-line tool called ExifTool reads these fields. After installing it, this reads the tag from a single file:

exiftool -XMP-iptcExt:DigitalSourceType product-001.jpg

To scan a whole folder and list the files that have no value, you can run:

exiftool -r -if "not $XMP-iptcExt:DigitalSourceType" -p "$Directory/$FileName" ./export

The double quotes suit the Windows command prompt. On a Mac or Linux terminal, use single quotes around the condition and the print format instead, or the shell will treat the $ signs as its own variables. When the tag is present, its value is an IPTC web address ending in the type, such as http://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia.

Illustrative result: out of 120 files in the folder, 17 print as missing the tag. Those 17 are the ones to fix before upload. The output is a short list of file names, which makes it a good job for a spreadsheet of products against files.

If you do not want to use a command line, many photo editors show IPTC fields in a File Info or Metadata panel. Open a sample and look for a digital source type entry. Whatever you use, test the file as it will be served to Google. Download a product image from your live store page by its image address, and read the tag from that download, not from your original.

Where metadata gets lost

The tag can be written correctly and still vanish between the generator and Google. These are the stages worth testing one at a time, because a pipeline can pass at step one and fail at step three.

I cannot tell you which tool does which, because behaviour changes between versions and settings. That is why the test belongs in the process rather than in a list of tools. If a stage strips the tag, either change its setting, move it earlier in the workflow, or add the tag again as the last step using a tool that writes IPTC fields. Re-adding the tag by hand is only honest for images that really were AI-generated; do not use it to claim something about a photo that was not.

A walk-through: a trade-supplies webshop with 120 products

Take an illustrative wholesale distributor of kitchen and catering supplies that has opened a small public webshop. Its supplier photos are inconsistent, so the team uses an AI tool to put each item on a plain light background and, for 30 items, to build a lifestyle scene on a counter. Of 120 products, 90 use a cleaned real photo and 30 have a generated lifestyle image.

They sort the files into three groups before touching the feed:

• Group A, 90 main images. Real photo with the background swapped by AI. Tag expected: whichever composite value the editor writes, such as CompositeSynthetic.
• Group B, 30 lifestyle images. Fully generated scenes. Tag expected: TrainedAlgorithmicMedia. These go in the lifestyle image link attribute, not the main image.
• Group C, 0 images. Nothing in the catalogue is a mock-up pretending to be a product. Anything generated that shows a product the shop does not sell goes in the bin.

They run the ExifTool scan over the export folder: 11 of the 90 group A files and 4 of the 30 group B files have no tag, because those files were saved through a compressor. They re-export the 15, rescan, and get a clean result. Then they upload one sample product to the store and download its live image by address, and the tag is there. Only then does the feed go. Total time for the whole check on 120 products: about 90 minutes, mostly the first time.

Lifestyle images: where AI helps without breaking the main-image rules

The strictest rules apply to the main image. A lifestyle image is a separate field, intended to show the product in context, so it is where generated scenes belong. Keep the main image a plain photograph of the real item and give the AI-made scene its own attribute. This also protects you if a generated scene shows the product slightly wrong, for example a handle on the wrong side. A shopper who sees a lifestyle image that misrepresents the product is a refund waiting to happen, and Google's misrepresentation policy expects listings to represent your products in a way that is accurate, realistic and truthful.

Accuracy checks for generated scenes

1. Count the parts. Does the generated image show the same number of pieces, handles, drawers or lids as the real product?
2. Check scale. Is the product the right size relative to a cup, a hand or a countertop?
3. Check text on the item. AI tools often garble printed text and logos. If the product has a brand mark, compare it to the real one at full size.
4. Check colour against a real sample. Colour shifts are a leading cause of returns.

A realistic failure: a generated scene of a stainless-steel tray shows four compartments; the real tray has three. Nobody notices until a customer returns it. The checklist item "count the parts" exists because of that kind of error. For more on judging the quality of generated backgrounds, see AI backgrounds for product photos: steps and quality checks.

What to do with supplier photos

If your store resells supplier images, the same Merchant Center rules apply: no supplier watermark, no logo overlay, no promotional text baked in. Using the same photo as hundreds of other resellers also makes your listings look like copies. The tutorial on replacing dropshipping supplier images safely covers when to reshoot and when generated replacements are acceptable. Whatever you choose, label what is generated, and keep a record of which files were made how.

Automatic image improvements

Merchant Center has a setting to turn on automatic improvements, which lets Google fix minor image policy issues such as promotional overlays to avoid disapprovals. It can be a useful safety net, but it is not a reason to relax the checklist. It also means Google may alter the image shown, and you should look at how your products appear in the listings periodically. Google also notes that it may test subtle cropping on images to frame the item better.

Using AI itself to audit the image list

A chat tool cannot open your files and read their hidden fields, so it cannot tell you whether the tag is present. What it can do is turn the output of your checks into a to-do list. Export the ExifTool scan, the pixel dimensions and the file sizes into a spreadsheet, paste the rows into the tool and ask for exceptions only.

Here is a table of product images: SKU, width, height, file size in MB,
format, file extension, AI tag found (yes/no), image purpose (main/lifestyle).
List every row that breaks any of these rules and say which rule:
- width or height under 1500 (warning, recommended size)
- width or height under 500 (fails the minimum)
- file size over 16MB
- extension does not match format
- AI tag missing on a row whose purpose says "AI background" or "generated"
Return a table: SKU, rule broken, fix.

Illustrative output: 6 rows under 1500 pixels, 1 row at 480 x 480 failing the minimum, 2 files with a .jpg extension that are really PNGs, and 15 rows with the tag missing. The tool is doing arithmetic and sorting here, which it does reliably when the table is complete. It is not looking at the pictures, so a human still compares generated scenes with the real products.

The export routine that keeps the tag

Once you know where a tag is lost, fix the workflow once instead of every batch. A routine that works for a small team:

1. Generate or edit in the AI tool and export at full size.
2. Run the cleanup or cutout tool, then re-read the tag. If it is gone, note the tool as a stripper.
3. Resize to 1500 x 1500 or above with a method that keeps metadata, and re-read the tag.
4. Upload to the store, download the served image by address, and re-read the tag.
5. Only then add the image address to the feed.

Writing this as a five-line checklist stuck above the desk is worth more than any tool. When a new team member joins or a plugin is installed, the routine tells them what to retest.

What this costs in time

For a 120-product catalogue, the first full pass takes about an hour and a half, as in the webshop walk-through. Each later batch of ten new products takes ten to fifteen minutes, because only the new files need testing. Compare that with finding out through Diagnostics that 40 products stopped showing the week before a sale, and re-exporting under pressure. The check is cheap; the cost sits in not having a routine.

If something is disapproved

Google's image page says a non-compliant image will not appear on Google surfaces and a notification shows in the Diagnostics section of your account; product-level issues are worked through from the Needs attention tab. Fix the cause in the source file, resubmit it through your feed, and request a review where the page offers one. Google's review guidance says its review timeline applies at both account and item level, and that a review button may stay unavailable until merchant verification is done and the data source holds at least one product. If you are seeing account-wide problems rather than a handful of products, read the tutorial on a suspended Merchant Center account, because a systematic image mismatch can be one part of a larger trust problem.

A mismatch that looks like an image problem

Not every image disapproval is about pixels. If the image shows a different variant from the landing page, or the page shows a product the feed image does not match, Google treats it as inconsistent information. A seller who photographed the blue version once and used it for all five colours has this problem, however good the photo. Match each variant's image to its own page.

A short record that makes audits painless

Keep a simple sheet with one row per product: SKU, file name, how the image was made (photograph, AI background, fully generated), expected IPTC value, date checked, and the tool used last. After a platform change or a new plugin, re-test a sample of ten rows rather than the whole catalogue, and escalate only if any fail.

• Date and tool for every export, so you can trace a batch that lost its tags.
• The product's lifestyle image attribute filled only when the scene is accurate.
• A reviewer initials column: a second person looks at the generated scenes against the real product.

This is dull work that pays for itself the first time Merchant Center flags a batch. For wider feed hygiene, the Merchant Center fixes for a small store tutorial covers the data side, and removing backgrounds with AI deals with the cutout stage that so often strips metadata.

Image questions that stop products going live

What size must a Merchant Center product image be?

Google's image_link page announces a minimum of 500 x 500 pixels for all products from 31 January 2027, and recommends about 1500 x 1500 or more. An image may not exceed 64 megapixels or 16MB. Check the page itself for the minimum that applies on the day you upload.

Do I have to label an AI product photo if I only touched it up?

Google's rule covers images created with generative AI. Preserve any IPTC digital source type tag a tool has already written, and do not strip it. If you are unsure whether an edit counts, keep the tag the tool added and ask Merchant Center support to confirm.

Can Merchant Center disapprove an image because the metadata is missing?

Google says images that break its image rules will not appear and a notice shows in Diagnostics. The AI metadata requirement sits in the same rules, so treat a missing tag as a disapproval risk and test every export before upload rather than waiting for the notice.

Is a lifestyle image treated differently from the main image?

The AI metadata rule applies to the main image, additional images and the lifestyle image link alike. The main image has stricter content rules: clean product, no text or overlays, no borders. The lifestyle image is meant to show the product in use.

Further reads

• Does Amazon Allow AI-Generated Product Images? What Passes Review — How another marketplace treats the same kind of image.
• How to Create Product Photos With AI on a Small Budget — Making product photos with AI on a small budget.
• AI Tools for Ecommerce Photography: Backgrounds, Shadows, Consistent Sets — Consistent sets of AI backgrounds and shadows.
• Jewellery Photography With AI: Clean Backgrounds, True to Life — Keeping colour honest in generated backgrounds.
• Google Shopping Titles: Rewrite Product Titles With AI Safely — The same AI label rule applies to titles.
• AI in Fashion Photography: Lookbook Images for a Small Brand — Lookbook images for a small brand, and their limits.
• How to Use AI to Stage Furniture for Product Page Room Sets — Build room sets around your real product photos, check scale, colour and construction against the piece, and label the result honestly.
• Can You Put Jewellery on an AI Model? Disclosure Rules for Sellers — A scale check with real numbers, visual checks, platform rules for AI images and caption wording to copy for listings, ads and social posts.
• Examples of AI in Retail: Eight Jobs a Small Shop Can Copy — Eight AI jobs a one-shop retailer can copy on a spreadsheet and a free or $20 plan, each with a worked example, the tool to use and the check to make.
• Best AI Tools for a Small Ecommerce Store: 7 Worth Paying For — Seven paid tools by job, with plan prices, what Shopify already includes free, what to skip and three monthly budgets.
• Google Ads for Dropshipping With AI: What to Let It Write — What AI can safely draft for a dropshipping store's Google Ads, what only you can verify, and the budget sum that decides whether to start.
• Product Categorisation With AI: Fixing Shopify Taxonomy at Scale — Fix uncategorised Shopify products in bulk: shortlist, batch prompts, validate against the official list, audit a sample, then import only the category column.
• Shopify SEO With AI: Fix Titles and Meta Descriptions in Bulk — A CSV round trip for Shopify SEO text: export, rewrite with AI in batches, check length and duplicates, import, and judge it in Search Console.
• Missing GTIN in Google Merchant Center: What AI Can and Can't Fill — AI must never invent GTINs. Where to get real ones, when identifier_exists is the right answer, and how to check a feed.
• Perplexity Merchant Program: Should a Small Online Store Join? — What Perplexity and feed-tool guides say about the free programme, who can apply, what a product feed needs, and a sum to decide if the prep time pays.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Google Merchant Center Help (image link attribute, generated content policy, request a review), Google Search Central guidance on generative AI content, IPTC NewsCodes digital source type vocabulary.