A thirty-second product ad used to mean a shoot, a crew, and a line item big enough to make a small business owner wince. As of October 2026, the honest version of that sentence has an exception: if the product is genuinely simple to show (a physical object, a clear before/after, a short demo) you can get a usable rough cut using two tools you probably already pay for or can try free, ChatGPT and Google's Flow. This is a step-by-step workflow for doing that properly, not a "AI will replace your video team" pitch. It won't replace a real shoot for anything that needs human performance, brand-specific talent, or footage of your actual physical location. It's genuinely useful for a product demo, a social ad test, or a concept you want to see moving before committing real budget to it.

What each tool is actually for in this workflow

Google Flow is Google's AI filmmaking tool, built on the Veo video models (currently Veo 3.1) and the Nano Banana Pro image model, and it lives at flow.google.com after the earlier Google Labs address started redirecting there. It generates short video clips from text prompts, reference images, or a starting and ending frame, and gives you a lightweight in-browser editor (Scenebuilder) to stitch clips together. It does not write scripts, plan shots, or think through a creative brief for you, it generates footage from instructions you give it.

ChatGPT is the planning layer: the brief, the script, the shot list, and the actual Flow prompts, written as structured text. It has no native video generation of its own in this workflow, it's doing what a human creative director or copywriter would do before a shoot, just faster and for a fraction of the cost. Dividing the labor this way, planning in one tool, generation in the other, plays to what each one is actually built for instead of asking either to do the other's job badly.

Step 1: Build the brief in ChatGPT

Start with a brief, not a prompt. Give ChatGPT the actual constraints: what the product is, who it's for, the one feeling or action you want the ad to produce, the length (thirty seconds, in this case), and the platform it's going on, since a vertical Instagram Reel and a horizontal YouTube pre-roll want different pacing and framing. Ask it to push back on a vague goal before writing anything; "sell more of this" isn't a brief, "get a 25-34 year old scrolling Instagram to tap through to the product page" is closer to one.

A useful prompt structure: "I'm making a 30-second ad for [product, one sentence]. The audience is [who]. The goal is [specific action]. The tone should be [adjective, adjective], not [adjective you want to avoid]. Write me three different creative concepts for this ad, each in two or three sentences, before we pick one and go deeper." Getting three directions before committing to one catches a weak concept early, when it costs nothing to throw away, instead of after you've burned Flow credits generating clips for it.

Step 2: Turn the concept into a script and shot list

Once a concept is picked, ask ChatGPT to break it into a shot-by-shot script: each shot's duration (keep individual shots to roughly 4-8 seconds, since that matches Veo 3.1's native clip lengths and avoids forcing an awkward cut), what's visually happening, any on-screen text or voiceover line, and the transition into the next shot. For a thirty-second ad, that's typically four to six shots once you account for an opening hook, one or two product or benefit shots, and a closing call to action.

This is also where you should ask ChatGPT to flag anything Flow will likely struggle with. Photorealistic human faces doing precise, repeated actions across multiple clips, exact on-screen text rendering, and strict brand-color accuracy are all areas where generative video is still inconsistent, and knowing that before you generate ten clips saves real time. If a shot genuinely needs a specific actor's face held consistent for brand recognition, that's a sign the shot belongs in front of a real camera, not in Flow.

It's worth deciding at this stage, not later, whether any shot needs a voiceover or on-screen speaking line. Flow's native audio generation handles ambient sound and background noise reasonably well, but a product narrator reading precise copy is a different and harder ask than environmental audio. If a line needs to land word for word, plan to record that line separately (your own voice, a text-to-speech tool, or a human voiceover artist) and sync it in the editing pass rather than hoping Flow's generated audio gets the exact wording right.

Step 3: Write the actual Flow prompts

This is the step people skip, and it's the one that determines whether Flow gives you something usable or ten throwaway clips. Ask ChatGPT to convert each line of the shot list into a structured Flow prompt, not a vague sentence. A workable template: subject and action (what's in frame, what it's doing), camera (static, slow push in, handheld, overhead), lighting and mood (soft natural light, moody and dim, bright and clean), and setting. "A ceramic coffee mug on a wooden table, steam rising slowly, static close-up shot, warm morning light through a window, cozy kitchen setting" will get you something far more usable than "coffee mug ad."

Two Flow-specific features matter a lot for a multi-shot ad staying visually consistent. Ingredients to Video lets you feed in reference images (a product photo, a specific character or object) and have Flow use them as a consistent visual anchor across multiple generated clips, rather than the product looking subtly different in every shot. Frames to Video lets you set a specific starting frame and ending frame and has Flow generate the motion that bridges them, which is the tool for a smooth transition between two shots you already know the start and end state of, like a product rotating into its final resting position.

Generate more variations than you'll use. Veo 3.1 clips run 4, 6, or 8 seconds depending on settings, with 1080p and 4K resolution locked to the 8-second option, and the model generates audio natively as part of the clip rather than as a separate pass, so a clip that nails the visual but gets the ambient sound wrong is a real, common outcome worth planning for, not an edge case.

Step 4: Assemble in Scenebuilder

Flow's Scenebuilder is the in-browser storyboard where generated clips get assembled into the full sequence. You can trim individual clips, extend a scene that cut off too early, and reorder clips within the timeline. A "Jump To" feature uses Gemini to generate transitions between two clips that don't naturally flow into each other, which is useful for a hard concept jump the shot list calls for (interior to exterior, product to lifestyle shot) that wouldn't make sense as a smooth Frames to Video bridge.

Build the thirty-second ad inside Scenebuilder in the order from your shot list, review the full sequence at actual speed (not clip by clip), and flag anything that reads as jarring or too fast before moving to the editing pass. A thirty-second runtime with six shots averages five seconds each, which doesn't leave much room for a shot that doesn't earn its place.

Step 5: Edit, caption, and finish

Export from Flow and bring the sequence into a standard video editor for the finishing work Flow isn't built for: adding a consistent logo bumper, overlaying brand-accurate on-screen text (since text rendered inside the generated video itself is one of the least reliable things current video models do), color-matching clips that came out slightly different in tone, and adding captions. Captions aren't optional for a social ad: a large share of video is watched muted in-feed, and burned-in captions are the difference between a scroll-past and a watched ad for that audience. If ChatGPT wrote your script with a voiceover line, that same script is your caption text, just time it to the final cut rather than reusing timestamps from the plan, since the generated clips rarely land at the exact duration you asked for.

Brand consistency across a multi-clip ad

Generative video's weakest point for advertising specifically is holding a product, character, or color palette consistent across separately generated clips, since each generation is, structurally, a fresh roll even with reference images feeding it. Three things help: lean on Ingredients to Video for every shot that includes the actual product, rather than regenerating the product from a text description each time; keep your Flow prompts' lighting and mood language consistent across shots even when the setting changes, since a dramatic shift in the implied lighting between adjacent clips reads as jarring even if nobody can articulate why; and do a final color pass in your editor regardless of how careful the prompts were, because even a well-prompted sequence benefits from unifying the color grade in post. None of this makes multi-clip generative advertising look exactly like a single continuous camera shoot, and framing it that way to a client sets up a disappointing reveal; framing it as a fast, affordable way to test a concept or produce content for a channel that doesn't require broadcast-level consistency is the honest pitch.

Commercial use and the SynthID disclosure rule

Content generated with Veo through Flow carries an invisible SynthID digital watermark embedded in the file, which identifies it as AI-generated to compatible detection tools, regardless of your subscription tier. That watermark stays on the content even if you don't notice it; it isn't something you can opt out of by paying for a higher plan. Separately, Google also applies a visible on-screen watermark to some generated content by default, controllable through a "visible watermarking" toggle in your account settings, with that visible watermark applied automatically and non-optionally for users in India, South Korea, and Vietnam. Google's own guidance frames commercial use of Veo-generated output as permitted under its terms of service as long as the content complies with its safety and acceptable-use policies (no real people depicted without consent, no content designed to deceive). That said, verify the exact terms for your plan and region on Google's own Flow help documentation before relying on anything a secondhand blog post, including this one, says about licensing specifics, since Google has updated the watermarking and usage terms more than once in 2026 and a stale screenshot of an old policy is easy to find.

Separate from Google's own terms, treat disclosure as a client-relationship issue too. If an ad running on a client's account was built substantially with generative video, say so plainly in whatever creative approval document you use. "AI disclosure clauses" that spell out which assets were AI-generated are becoming standard in agency contracts specifically because clients are increasingly asking after the fact, and it's a much easier conversation to have upfront than to have reactively.

What this workflow costs

Flow access requires a Google AI subscription; there's no way to use it without one. As of October 2026, Google AI Plus runs $4.99 a month with 200 Flow credits, Google AI Pro runs $19.99 a month with 1,000 credits, and Google AI Ultra splits into two tiers, $99.99 a month for 10,000 credits and $199.99 a month for 25,000 credits, with the higher tier also removing the visible watermark on Veo 3 output specifically. A standard Veo clip generation runs somewhere around 20 credits depending on the specific model and settings chosen, so a thirty-second ad assembled from six shots with a handful of regenerated takes per shot can realistically burn several hundred credits once you count the clips you throw away, which the Plus tier's 200 monthly credits won't comfortably cover for more than one serious attempt. Anyone over 18 with a Google account also gets 50 free daily credits with no subscription at all, enough to test the workflow on a single clip before committing to a paid tier. ChatGPT's planning side of this workflow works on a free account for occasional use, though the paid tiers remove rate limits that matter if you're iterating through multiple creative concepts and shot lists in one sitting.

A realistic sample workflow, start to finish

For a thirty-second ad for a reusable water bottle: brief ChatGPT with the product, audience (people who bring a water bottle to the gym), and goal (drive clicks to a product page from an Instagram Reel). Get three concepts back, pick "a single bottle's day, from gym bag to desk to nightstand, told in quick cuts." Have ChatGPT break that into five shots of roughly six seconds each: gym bag zip-open reveal, bottle on a desk with condensation beading, a hand grabbing it mid-workout, a close-up pour, and a final shot of the bottle on a nightstand with a logo end card. Convert each shot into a structured Flow prompt using Ingredients to Video with your actual product photo as the reference image so the bottle looks the same in all five. Generate two or three takes per shot, assemble the best ones in Scenebuilder in order, use Jump To for the one hard cut (gym to desk) that doesn't bridge naturally. Export, color-match in a standard editor, add burned-in captions for the two lines of on-screen text, and confirm the SynthID disclosure requirement is satisfied before it goes live on a client account.

For a broader look at where Flow sits against other AI video tools if this specific workflow isn't the right fit for your product, the roundup of AI video generator tools for marketing covers alternatives worth comparing before committing a subscription budget to any one of them.

Flow is a strong fit for cinematic product shots, but it isn't built for ad testing at scale. If you plan to run these clips as paid social ads, look at the AI video ad generator tools made for that job, and read our guide to creating Meta ads with Meta AI before you upload.

FAQ

Can I use Google Flow without a Google AI subscription?

Only in a limited way. Anyone 18 or older with a Google account gets 50 free Flow credits per day with no subscription, enough to test a single clip, but a real project needs a paid Google AI Plus, Pro, or Ultra plan for a usable monthly credit allowance.

How long can a single Google Flow clip be?

Veo 3.1 generates clips of 4, 6, or 8 seconds, with 1080p and 4K resolution only available on the 8-second option. Longer sequences are built by assembling multiple clips together in Scenebuilder, not by generating one continuous longer clip.

Does Google Flow add a watermark to generated videos?

Yes, an invisible SynthID digital watermark is embedded in every Flow-generated video regardless of plan. A visible on-screen watermark is also applied by default in most regions (and automatically, with no opt-out, in India, South Korea, and Vietnam), though Ultra subscribers can turn off the visible watermark on Veo 3 output specifically.

Can I legally use Google Flow videos in a commercial ad?

Google's terms of service permit commercial use of Veo-generated content as long as it complies with safety and acceptable-use policies. Confirm the exact current terms on Google's own Flow help pages before publishing, since watermarking and usage terms have changed more than once in 2026.

Why use ChatGPT if it can't generate video itself?

Because the planning work, the brief, script, shot breakdown, and structured prompts, determines whether Flow's output is usable or a pile of disconnected clips. Skipping straight to typing vague prompts into Flow produces noticeably worse results than writing a real shot list first.