The demo works. That is the honest place to start, because most writing on this subject either sells you a miracle or sneers at one, and neither helps you decide what to do on Monday.
Video used to be the format small businesses could not afford. A produced minute ran into the thousands, sometimes far past that, and the cost was mostly people and time rather than equipment. That is the number generative video has actually broken. A clip now costs cents and arrives in under a minute, and the businesses that were priced out of motion entirely are not priced out anymore.
So the question is no longer whether you can make a video ad. You can. The question is which shots you can hand to a model and which ones still need a camera, and almost nobody is answering that, because the people demonstrating the tools are demonstrating the shots that worked.
That is worth sitting with for a second. Every AI video demo you have watched is a selected take. You are seeing the one that came out right, not the eleven before it, and not the specific frame where the thing fell apart. The failures are not random either. They cluster in one place, and it happens to be the place your ad cares about most.
Your Product Is the One Thing It Can't Keep Straight
Here is the moment that should shape how you use these tools.
A creator at OpenArt built a complete AI-generated testimonial ad for a skincare serum and benchmarked every current model by hand along the way, which makes it one of the more useful things published on this subject, mostly because it documents its own failures while selling you the software. The character came out convincing. The skin texture held up under zoom. The lip sync landed. Then the presenter turned the bottle in her hand, turned it back, and in his words, "the whole text changed."
Not smudged. Changed. The label on the product being advertised became a different label, mid-shot, in an ad whose entire job was to make you want that product.
His summary of testing every animation model available to him: "One issue all of them have is that the text does not remain fully readable during the process. This is just the current limitations of these models."
That is the mechanism, and it is worth understanding rather than memorizing, because it explains every other failure you will hit. These models do not store your product and then move it around. They redraw the scene, frame after frame, each one conditioned on the last. Anything with fine structure that must stay identical across all of them is being re-guessed constantly. Your logo. Your label copy. The shape of your bottle. The stitching on your bag. The model has no concept that those things are supposed to be fixed.
Faces got solved first because there was enormous incentive and enormous training data. Your packaging is not in the training data. Neither is your storefront, your equipment, or the specific thing you sell.
This is the same line we drew for AI product photography, where the failure showed up as a wrong label in a single frame. Motion does not soften that problem. It multiplies it by the frame rate.
Eight Seconds Is the Real Unit
The second thing the demos hide is length.
Most generators produce short clips in a single pass, on the order of five to fifteen seconds depending on the model, and reach longer runtimes by chaining those clips together. That chaining works, up to a point. Practitioner testing and the platforms' own guidance converge on roughly the same place: quality holds for the first half minute or so, and past that you start seeing characters drift in appearance, lighting shift between segments, and motion get less natural.
Treat that as the shape of the medium rather than a limitation waiting to be patched.
You are not generating a video. You are generating shots, and then cutting them together. Which means the skill that decides whether your AI ad is any good is not prompting. It is knowing what sequence of shots tells your story, and that is the same skill it has always been. If you have not decided what the ad is doing before you open the tool, the tool will hand you thirty beautiful seconds of nothing. Our breakdown of video ad structure is the piece to read first, and it has not been made obsolete by any of this. It got more relevant.
What the Workflow Is Doing When It Cuts Away
Watch how experienced people actually assemble these ads and you notice something they say out loud without flagging it as a problem.
In that same walkthrough, B-roll gets introduced twice, and both times the stated reason is the same: it is "great for covering imperfections in AI generated talking footage," and later, when assembling the final cut, "this is also a great way to handle any lip sync imperfections. You can just simply cover them with B-roll."
That is not a gotcha. It is genuinely how the craft works, and it is how the craft has always worked. Editors have been cutting away from flubbed lines since film was physical. But it should reset your expectation of what you are buying.
The demo shows you a prompt and a finished clip. The actual pipeline is: write the script, generate the character, generate the hero frame, animate it, generate the supporting shots, then edit the supporting shots over the parts where the animation failed. That last step is a person with judgment and an hour. Budget for it, or you will ship the version with the visible seams.
The Only Sorting Question That Matters
Everything above collapses into one question you can ask about any shot in your ad.
Does this shot have to be true?
Some shots are claims. The product in a hand. The label. The result on a real customer. Your premises, your team, your equipment. If a customer could hold the real thing up next to the frame and find a difference, that shot is making a promise, and a generated version of it is making a promise you did not keep.
Other shots only have to be plausible. The kitchen counter the bottle sits on. The morning light. The street outside. The abstract texture behind your text. The seven hook variations you want to test before you know which one earns attention. Nobody is auditing your background against reality, because your background is not the offer.
Generate the plausible. Film the true.
The useful part of that rule is that it does not move when the models improve. It is not a claim about what the technology can do this quarter. It is a claim about which parts of your ad carry a promise, and that does not change.
It also points somewhere cheaper than either option. Most businesses already own the true shots and have forgotten. There is a folder from a shoot you paid for, and a few hundred clips on your phone, and the fastest route to an honest ad is usually editing what you already have rather than generating a synthetic version of footage you own.
Four Things That Will Stop the Ad You Just Made
The operational details that never appear in a thumbnail, and any one of them can cost you a launch.
The watermark. Google's consumer video plans stamp a visible badge on your output below the top subscription tier. You can generate a perfectly good ad and not be allowed to run it. Check before you build a campaign around a tool, not after.
The watermark you cannot see. Separately, Google embeds SynthID in generated output, an imperceptible marker that its own documentation says is "designed to stand up to modifications like cropping, adding filters, changing frame rates, or lossy compression." Cropping out the visible badge does not remove it. Assume anything you generate is identifiable as generated, permanently, and plan on that basis rather than on getting away with it.
Platform disclosure. Meta labels AI content in ads on its own now, without waiting for you to declare it, and the strict self-disclosure requirements bite hardest on social issue, electoral and political advertising. Google and YouTube run their own separate regimes, and detection increasingly rides on embedded content credentials rather than the honor system. Compliance on one platform is not compliance on the next.
Rights and the version rot. Read what you actually own on your plan tier, because commercial-use terms differ between free and paid access on most of these tools. And treat every model-specific fact you read, including the ones in this article, as perishable. Capabilities that were true in the spring were wrong by the fall. The sorting rule survives. The feature list does not.
The Bar Moved While You Were Reading the Demos
There is a cost to the generated look that has nothing to do with whether you used the tools well, and it is the part most marketing coverage misses because it only talks to marketers.
In late August, a consumer investigation channel with more than a million and a half subscribers published a video that has since been watched over two million times. The premise was simple: buy the products behind the AI-generated ads flooding his feed and show what arrives. A garden hose nozzle claiming to replace a pressure washer, which does roughly what a five dollar nozzle does. An ear cleaning device that tickles. Two orders totaling around two hundred and fifty dollars that never shipped at all. He ends on this: "AI is ruining advertising to the point where I don't trust a single thing that I see online anymore."
Be careful with what that does and does not prove. He went looking for scams and found them, so this is not a measurement of how many AI ads are fraudulent. It is something more awkward. It is evidence about the audience you are advertising to.
The generated aesthetic got adopted early and hard by dropshipping operations, for the obvious reason that if the product does not exist you cannot photograph it. Consumers have been trained on that association at enormous scale. When your honest ad for your real business carries the same visual signature, some part of your audience files it next to the hose nozzle, and you never see that in your metrics. It shows up as a slightly worse hook rate you attribute to the creative.
We wrote about a narrower version of this when AI presenters started showing up in small business ads. The trust problem has gotten broader since. It is no longer only about a synthetic person. It is about the whole look.
None of that is an argument for avoiding these tools. It is an argument that the output has to be good enough not to read as generated, which is a much higher bar than the demos set, and which is exactly the bar a real shot of a real product clears for free.
Brief It Like a Shoot
The most useful framing came from an agency team who had put AI video through enough production cycles to stop being excited about it. Their conclusion was that the mistake is treating generation as a button rather than as a production function. As they put it, a strategist or an editor "would never go out and shoot the content themselves. You'd hand that off to a production team with a brief."
Nothing about generation changes that. You still have to decide what the ad argues, who it argues to, what the offer is, and what has to appear on screen for the argument to land. A model that renders whatever you type is not a shortcut past that thinking. It is a faster way to find out you never did it.
Which is the same finding as ever, arriving in new clothes. A bad brief produced a bad ad slowly and expensively before. Now it produces forty bad ads by lunch.
What to Actually Do
Split your storyboard in two before you open any tool. Mark every shot that makes a promise and plan to film it, on a phone, in an afternoon, which is genuinely enough. Mark every shot that only sets a scene and generate it. Keep clips short and cut between them rather than fighting for one long take. Budget an hour of real editing to cover the seams, because there will be seams. Then look at the finished thing with fresh eyes and ask the only question that counts, which is whether it looks like an ad from a business that exists.
That last question is the whole discipline. The tools got cheap. Judgment did not, and it is still the part that decides whether any of this works. If you would rather hand the judgment to someone who does this daily, that is what we do.
