If you sell physical products, the honest answer on Google Flow is this: it will not shoot your product, and it will shoot almost everything around it. Backgrounds, weather, streets, a monsoon window, a mountain road at dusk, a wedding hall you can't book. That is where the money is. Every team we've seen get value out of Google Flow for ecommerce is using it for the 80% of a video ad that isn't the product, then cutting in real footage of the thing they actually ship. The ones treating it as a replacement for a product shoot are burning a subscription and shipping ads that get returned.
What Google Flow actually gives an ecommerce team
Flow is Google's video generation tool, built on its Veo and Imagen models and announced at I/O in 2025. You write a prompt, or feed it a still frame, or feed it reference images of a character or an object, and it produces a short clip with sound. Clips run roughly eight seconds. If you want longer, you chain them in Scene Builder, extending the last frame into the next shot. There are camera controls, so you can ask for an orbit or a slow push instead of hoping the model picks one.
The reference-image feature is the part that matters commercially. Consistency across shots has been the thing that made generative video useless for advertising: the model's model would change hairstyle between cut one and cut two. Feeding it fixed references gets you closer to a usable sequence. Closer, not there.
Google's generated output carries SynthID watermarking. Assume anything you publish is identifiable as AI-generated, and plan your disclosure accordingly rather than hoping nobody checks.
Ten ways to put AI video to work in an ecommerce funnel
1. Hook variants for paid social
The first two to three seconds of a Meta or YouTube ad does most of the work. Generating fifteen different openers for the same body footage is cheap in a way a reshoot never is. We build one master cut with real product footage, then swap only the opening beat: a hand reaching for a door, a city at 6am, a suitcase closing. Same creative, fifteen entry points, and the ad account tells you which one earned attention.
2. Background plates and set extension
You have a clean packshot on white and no budget for a location. Generate the location, composite the packshot in. This works far better than generating the product, because the product on screen is your product, pixel for pixel.
3. Environment b-roll instead of stock
Stock libraries are full of footage that ten of your competitors already used. Generating a specific establishing shot to your brief costs less time than searching for one that nearly fits. Kerala backwaters at a particular hour, a Dubai high-rise interior, a Melbourne cafe. Nobody else has that exact frame.
4. Market variants of one campaign
Brands selling into India, the UAE and the US usually run one hero film everywhere because three shoots is three budgets. Now the setting can change per market while the product cut stays identical. A skincare ad set in a Bengaluru apartment and one set in a Sharjah villa, same bottle, same voiceover script, different rooms.
5. Pre-visualisation before a real shoot
This is the use we'd defend hardest and the one nobody markets. Generate the shot list before the crew arrives. Everyone — founder, DOP, agency — argues about the frame the day before instead of at 11am on shoot day with a model on the clock. A studio day in Bengaluru with DOP, lighting, model and post commonly gets quoted somewhere around ₹1–1.5 lakh. Losing ninety minutes of it to a creative argument costs you real money.
6. Festive spots on a compressed timeline
The Diwali brief lands three weeks out, half your inventory plan changes, and the shoot window is gone. Generated ambience — diyas, lanterns, a rangoli in a doorway, Eid table settings — plus your existing product footage gets a campaign out. It's not the film you'd have made in August. It ships.
7. Homepage and collection motion
A four-second looping hero on a landing page, or a subtle motion banner behind a collection header. Keep it short, keep it quiet, and keep the file small — more on that below.
8. Lifecycle email and WhatsApp
Abandoned-cart and winback flows respond well to a two-second animated frame where a static image sat before. Export as a small looping clip or GIF. Test it against the static version; sometimes the static wins because it loads instantly on a bad 4G connection in a lift.
9. Pitch and trade films
Selling into a marketplace, a retail buyer or an investor deck. Nobody in that room believes the b-roll is real anyway. Generated aspiration is fine here in a way it is not in a consumer ad.
10. Category education where the product isn't on screen
How gold purity is measured. What GSM means for a bedsheet. Why cold-pressed differs from refined. Explainer content where you're teaching a concept rather than depicting your SKU is the safest generative use of all, and it feeds organic search as well as social. Pair it with written pages that can actually rank — video alone won't do that, which is a different problem and one our SEO and content work exists to solve.
Where Google Flow for ecommerce falls over
Product fidelity. If the thing you sell has a logo, a label, a print, a weave or a stone setting, the model will get it wrong, and it will get it wrong in a way that reads fine to you and reads wrong to the customer who receives the parcel. Sarees drape incorrectly. Kurta prints repeat at the wrong scale. Jewellery reflects light like plastic, and anyone in the gold trade will spot it in half a second. Text on packaging comes out as near-letters. Hands are better than they were and still not reliable when they're manipulating an object.
The commercial consequence isn't aesthetic, it's returns. Show a customer a garment that doesn't exist and you'll pay for the shipping both ways, plus the COD write-off if they refuse at the door. India's advertising self-regulation expects claims in an ad to be substantiated; "this is roughly what our product looks like" is not a defence anyone wants to run.
Dialogue in Indian languages is the other weak spot. Native audio generation handles English well enough for a background line. Hindi, Tamil, Malayalam lip-sync in our tests has been inconsistent enough that we record voiceover separately and use the generated footage silent underneath. That's a small workflow change and it saves a lot of re-generation.
Also: multi-shot consistency degrades. Reference images hold a character for two or three chained clips. By the fifth, something has drifted. Plan cuts around that limit rather than fighting it.
The cost maths, done properly
People quote the subscription price and stop there, which is like costing a shoot at the camera rental. The real number is your hit rate.
Say you want a 24-second ad. At roughly eight seconds a clip, that's three generations minimum if every one lands. They don't. Our working rate on a well-specified brief is about one keeper in five — sometimes one in three when the brief is a simple environment plate, worse than one in eight when there's a person interacting with an object. Three shots at one-in-five is fifteen generations. Add the prompt iteration before you find the phrasing that works, call it twenty-five. Then an editor spends three to four hours grading, cutting and compositing your real product footage in, because raw output is not a finished ad.
So: a subscription, plus half a day of editor time, plus the original packshot footage you still need to have shot. Against a ₹1.2 lakh studio day that yields, generously, six finished cuts — ₹20,000 a cut — the generative route is cheaper for volume and for anything set outside a studio. It is not cheaper, and not better, for the hero product shot. Shoot that once, properly, and reuse it for a year.
Don't let the video wreck your Core Web Vitals
Every time AI video gets easy, PDPs start carrying 15MB autoplay hero files and largest contentful paint goes from 2.1s to 4s on a mid-range Android over 4G. That is a real conversion loss, and it is entirely self-inflicted.
Rules we hold to: upload video to Shopify's own video hosting or a CDN with adaptive bitrate rather than dropping an MP4 into theme assets. Never let a video be the LCP element — use a compressed poster image and start playback after. Lazy-load anything below the fold. Cap looping background clips at four seconds and a couple of megabytes. If you want the score watched rather than guessed at, our speed optimisation work covers this, and SwiftStore scans and monitors PageSpeed on an ongoing basis so a new hero video doesn't quietly cost you a second of load time in week three.
A workflow that survives contact with a real brand
What works, in order: shoot the product once, on white and on one lifestyle set, in 4K, with enough coverage that you can crop to 9:16, 4:5 and 1:1 later. Build a reference-image library — your model, your packaging, your brand colours — and reuse it across every generation so the outputs sit in the same visual world. Generate environments and hooks, never the SKU. Keep every ad's product moment real. Disclose generative use where the platform or the market requires it, and keep the raw generations filed so you can prove what was made when.
The teams getting the most out of this are the ones who already had a video pipeline and are now producing four times the variants at the same cost. If you have no product footage at all, a generative tool doesn't fix that. It just gives you a faster way to make ads for a product nobody can see.
If you're planning festive creative and want a second pair of eyes on how it lands on the store itself — page weight, PDP video placement, what happens on a 4G connection in Coimbatore — book a free store audit and we'll look at the actual numbers.