AI Product Videos for E-commerce: A Catalog-Scale Workflow

You have 400 SKUs, one good photo of each, and zero product videos. Amazon says a first shoppable video on a detail page was followed by an average 24% sales increase in its January to June 2026 data. Google Merchant Center now takes video in the product feed. Filming 400 products is a quarter of work and a budget you do not have. Generating them one by one in a web tool is a month of clicking.

The workable version sits in between: treat the catalog as data, turn each product photo into a short image-to-video clip with a model chosen per SKU, run the whole thing from a spreadsheet or an agent, and export files that pass each marketplace's spec on the first upload. This guide covers that pipeline end to end, with the arithmetic.

Why video per SKU is worth doing now

Two changes in 2026 moved product video from a nice-to-have on hero products to something you can justify across a catalog.

Amazon published the lift. Its seller blog reports the 24% average and breaks it down by category: toys at 32.7%, apparel at 30.5% and home at 26.7%. A listing can carry up to four shoppable videos. These are Amazon's internal averages, so measure your own.
Google added a video attribute to the feed. The video_link attribute accepts up to 10 videos per product. According to Search Engine Roundtable, submissions opened on April 14, 2026, with serving plus policy and quality validation starting June 30, 2026.

Both channels reward coverage. One video on your top five products leaves the other 395 listings without one.

What each marketplace accepts

Start from the specs, because they decide duration, aspect ratio, resolution and file size before any creative choice. Here is what the sources say, and how solid each one is.

Shopify

The Shopify Help Center accepts MOV, MP4 and WEBM, up to 1 GB and 10 minutes, with width and height between 100 and 4096 pixels. Recommended aspect ratios are 16:9, 9:16, 4:3, 3:4 and 1:1, at 60, 30 or 25 fps.

Amazon

Amazon's own page lists .mov or .mp4, up to 1080p and 5 GB, and eligibility for professional selling accounts with at least three months of selling history. As summarized from that page, the stated length reads 1 to 12 minutes, which sits oddly with short product clips; a third-party guide from GOAT Consulting describes different rules, including a 1280x720 minimum, 16:9 preferred and Brand Registry as a requirement. Check Seller Central for your account before you render 400 clips to a length Amazon may reject. Review takes about 3 days, up to 7 business days at holidays.

Google Merchant Center

The video_link spec is the strictest on shape: 6 to 240 seconds, 500 MB maximum, 720p (1280x720) minimum, and an aspect ratio of 9:16, 16:9 or 1:1. Sound is optional. A non-YouTube URL must point directly at a raw video file, be publicly reachable and not expire. You must also hold the licensing rights for marketing use.

Walmart, TikTok Shop and Etsy

Walmart: the Marketplace Learn rich media guide calls for a direct .mp4 URL with no redirects, 100 MB or less, up to eight videos per item, and a submission that includes the item ID, GTIN, a closed-caption file name and the video language. Confirm these in Seller Center.
TikTok Shop: GoDataFeed's disapproval rule says one product video per listing, at 5 MB or smaller, when sent through a feed channel. Other upload paths quote larger limits, so the limit depends on how you submit.
Etsy: a Wyzowl summary puts listing videos at 5 to 15 seconds, 100 MB, MP4 or MOV, played silently, one per listing. This is secondary, so verify against Etsy Help.

The decision rule that falls out of it

Duration 6 to 10 seconds. Six clears Google's floor, and anything up to 15 fits the Etsy range as reported. Amazon is the exception to check.
Resolution 720p or higher for anything going to Google or Amazon. A cheaper 480p render is fine for a Shopify gallery and nowhere else.
Two aspect ratios, at most. 1:1 or 9:16 for Google, TikTok Shop and mobile storefronts, 16:9 where Amazon's preference applies. Render natively in each ratio you need; cropping a 16:9 clip to 9:16 usually cuts the product.
Design for silence. Google makes sound optional and Etsy plays silently as reported, so the clip has to read with the sound off. Audio is an add-on for social placements.
Encode per destination. One master file, then a small export for TikTok Shop's feed limit and a sub-100 MB one for Walmart.

From product photo to clip: the image-to-video step

Image-to-video is the right primitive for a catalog because the product is already defined by the photo. The model's job is to add motion around something that exists, which keeps the SKU recognizable. Text-to-video invents the product from a description and will get the label, the cap or the stitching wrong.

Prepare the source image

Use your highest-resolution photo, with the full product in frame and some margin for camera moves.
If the only photo is a packshot on white, generate a styled still first (on a counter, on a desk, in hand) with an image model, then animate that still.
Keep people out of the frame unless you need them. The disclosure rule covered below applies to photorealistic AI-generated people, and a product-only clip avoids it.

Write motion prompts that protect the product

The failure mode in catalog work is drift: the model redraws the logo, changes the color of a cap or adds a feature the product does not have. Amazon lists a mismatch between video and listing among its rejection triggers, according to GOAT Consulting, so drift is a compliance problem as well as a quality one. A motion prompt template that limits it:

Camera: one move only. "Slow push-in", "slow 30-degree orbit left", "static camera".
Subject: "the product stays unchanged, label and colors exactly as in the image".
Environment motion: small and secondary. "Steam rising", "fabric moving slightly in a breeze", "light sweeping across the surface".
Exclusions: "no text, no hands, no people, no new objects".

Before and after, for a ceramic mug:

Before: "Cinematic video of our amazing mug, dynamic, trending, beautiful lighting." The model picks its own camera path, invents latte art, and the glaze color shifts between frames.
After: "Slow push-in on the mug on a wooden table, morning light from the left, thin steam rising. The mug, its glaze color and its logo stay exactly as in the image. No text, no hands, no people." One move, one secondary motion, explicit locks.

The same discipline is what keeps generated footage from reading as synthetic; how to make AI ads that don't look AI-generated goes further on lighting and motion choices.

What a SKU video costs, with the arithmetic

Image-to-video is billed per second of output, and resolution and audio change the rate; fal's image-to-video catalog shows that pattern across providers. The table below uses public Aitachyon prices at the time of writing, for a 6-second silent clip, the shortest length Google accepts.

Hailuo 02 at $0.085/s: $0.51 per clip.
Kling v3 silent at $0.16/s: $0.96. With native audio at $0.32/s: $1.92.
Wan 2.7 at $0.19/s: $1.14.
Veo 3.1 Fast from $0.19/s: from $1.14.
Seedance 2.5 from $0.20/s at 480p: $1.20. At 720p, $0.44/s: $2.64.

The styled still, if you make one, adds $0.057 on Seedream 4.0, $0.057 to $0.086 on FLUX.2 [pro] or $0.13 on Nano Banana, all at the time of writing. A voiceover line for a social cut is priced at $0.043 to $0.086 per 450 characters on ElevenLabs, which covers a two-sentence product line.

Three batch budgets

50 SKUs, budget tier. 50 Seedream stills at $0.057 is $2.85. 50 Hailuo 02 clips at $0.51 is $25.50. Total $28.35, or about $0.57 per SKU.
50 SKUs, 720p for Google and Amazon. Seedance 2.5 at 720p for 6 seconds is $2.64 per clip, $132 for 50. Add $2.85 in stills for $134.85, about $2.70 per SKU.
500 SKUs, mixed. 450 long-tail SKUs on Hailuo 02 ($229.50) and your 50 best sellers on Kling v3 with audio for social ($96). Add 500 stills ($28.50). Total $354.

Budget for retakes. Some clips will drift and need a second render, and a render you reject is still a render you paid for. Only a render that fails outright is refunded. If you plan on half the SKUs needing one retake, multiply the video line by 1.5.

How that compares

Cheaper per-second rates exist. Atlas Cloud's product video guide lists Wan 2.6 Flash at $0.018/s ($0.144 per 8-second clip), Seedance v1.5 Pro at $0.047/s and Kling 3.0 Standard at $0.071/s, and estimates a 500-product catalog at about $72 to $188. Those are older or lighter model versions quoted on a vendor blog, so compare output quality on your own products before you compare rates. On the premium end, fal lists Veo 3.1 at $0.20/s silent and $0.40/s with audio, and Kling 3.0 Pro at $0.112/s silent. A model-by-model view is in AI video generation API pricing in 2026, and every current rate is on the models page.

Picking a model per SKU

A catalog is uneven, and the model choice should follow it. Use a three-tier rule and write the tier into the spreadsheet so the batch picks the model on its own.

Tier C, long tail: the cheapest model that holds the product shape. Silent, 6 seconds, one aspect ratio.
Tier B, steady sellers: a mid-priced model at 720p or higher so the file qualifies for Google and Amazon, two aspect ratios.
Tier A, top 5 to 10%: the best model you can afford, native audio where it helps a social placement, and several variants to test.

Run a calibration batch before committing. Take five SKUs that are hard for your category (reflective packaging, small text on labels, transparent bottles, fabric texture) and render each on three models. It costs a few dollars. For a direct comparison of two common picks, see Kling v3 vs Seedance 2.5.

Batching from a spreadsheet or an agent

Once each SKU is a row, generation becomes a loop. The spreadsheet is the source of truth, and the script or agent only reads rows and writes results back.

The columns

sku and gtin (Walmart submissions need the GTIN).
source_image_url, the photo to animate.
tier (A, B, C), which maps to a model.
still_prompt, optional, for a styled still.
motion_prompt, built from the four-part template above.
duration_s and aspect, from the decision rule.
channels, for example "amazon,google,shopify".
status, ref and cost_usd, written back by the run.
qa, filled in by a person: pass, retake or drop.

The run

From code, loop over rows with status empty, submit the still and the clip, poll until done, and write the returned ref and cost into the row. Atlas Cloud's own batching script runs five concurrent workers with parallel polling and JSON result logging, which is a sensible default for any provider. Write results after each job so a crash at row 312 resumes at 313. Producing hundreds of clips safely covers retries, rate limits and spend caps in more depth.

From an agent, the same loop is a prompt. With a hosted MCP server connected, you can ask Claude Code to "read products.csv, render every row with status empty using the model for its tier, write back the ref and cost, and stop if the batch passes $60". The agent does the loop, and the spreadsheet stays the record. A working setup for automating video production with agents walks through the pattern.

Keep the spend visible

Use a dedicated API key per batch or per client, so each run's spend is isolated and the key can be revoked if a loop misbehaves.
Estimate before you run (rows times duration times rate, per tier) and compare the itemised total afterwards. A gap means retakes or a wrong duration.
Store the ref, not only the file. A stable ref lets a later step, such as the export job, fetch the exact file that passed QA.

QA and compliance before upload

Run each clip through this checklist before it leaves the spreadsheet.

The pre-upload checklist

Product match. Compare the last frame to the listing photo. Logo, color, shape, component count. Any drift is a retake.
No price, discount or urgency. GOAT Consulting lists price, discount and time-limited offer references among Amazon rejection triggers, along with reviews or star ratings, competitor or other-marketplace mentions, and fake UI elements or watermarks. Models sometimes hallucinate text on packaging, so look for it.
People check. If a photorealistic AI-generated person appears, Amazon now requires the keyword contains-synthetic-performer in the dc:subject XMP field, added with an IPTC-compatible metadata editor before upload. Per ChannelX, the rule covers product images, lifestyle images, product videos and A+ Content in all Amazon stores, and Amazon may show shoppers an indicator. Content with no people, real people or non-photorealistic people is exempt.
Spec check. Duration, resolution, aspect ratio and file size against each channel in the row.
Silent read. Watch it muted. If the benefit is not visible, the clip does not work on Etsy or most feeds.
Rights. Google requires that you own the licensing rights for marketing use. Check the terms of the model provider and of any source photo you did not shoot.

Export and hosting per channel

This step is mechanical, and it is where catalog projects stall, because each marketplace wants the file somewhere different.

Google Merchant Center: host the MP4 on storage you control, at a public URL that points to the raw file and never expires, then add it as video_link in the feed. Signed or temporary links break the no-expiry requirement.
Walmart: same hosting, a URL ending in .mp4 with no redirects, plus a closed-caption file and the language field.
Amazon and Shopify: direct upload, so the constraint is file size and length. Shopify's 1 GB limit will not bind for an 8-second clip.
TikTok Shop via feed: a separate low-bitrate export to stay under the 5 MB limit reported by GoDataFeed.

An export script that reads the qa column, fetches each passing file by its ref, transcodes per channel and uploads to your storage turns this into one command. For the creative side beyond compliance, see the ecommerce video ads playbook.

FAQ

Can I use AI-generated product videos on Amazon?

Nothing in the sources bans them. The video must match the listing and avoid the usual triggers such as prices and star ratings. If the clip shows a photorealistic AI-generated person, you must add the contains-synthetic-performer keyword to the file's XMP metadata before upload, per ChannelX's summary of the August 2026 rule. Product-only clips are exempt.

How much does an AI product video cost per SKU?

For a 6-second silent image-to-video clip at Aitachyon prices at the time of writing, from $0.51 on Hailuo 02 to $2.64 on Seedance 2.5 at 720p, plus $0.057 to $0.13 for an optional styled still. Add a retake allowance. Vendor estimates elsewhere run lower on lighter models.

What length should a product video be for Google Shopping?

Google's video_link attribute accepts 6 to 240 seconds at 720p or higher, in 9:16, 16:9 or 1:1, up to 500 MB. For a catalog batch, 6 to 10 seconds keeps cost down and stays inside Etsy's reported range too.

How do I generate videos for hundreds of products at once?

Put one SKU per row in a spreadsheet with the source image, tier, motion prompt, duration and aspect, then loop over the rows from a script or an agent with modest concurrency, writing the ref, cost and status back after each job.

Sources

If you want to run this pipeline without five subscriptions, Aitachyon puts Hailuo 02, Kling v3, Seedance 2.5, Veo 3.1 Fast and the image models behind one prepaid balance, callable from a web studio, a plain HTTP API or a hosted MCP server in Claude Code, with every job itemised by model and cost and failed renders refunded automatically. The API and MCP quick start gets a first batch running from a spreadsheet in an afternoon.