AI Video Reframing: The 2026 Control Checklist for Performance Max

What video enhancements do to an approved master

AI video reframing is now switched on by default in Google Ads. Your approved 16:9 master can be flipped or extended into 1:1 and 9:16 without a second review, and one documented model uses generative AI to extend the frame itself. Here is what changes in your workflow, and the checks to run before any of it ships.

Google's documentation describes video enhancements as smart automation features that enhance your video in different ways, and states the default plainly: they are turned on by default and can be turned off at any time. In Performance Max the control sits in the Asset optimization section under the Video dropdown, where you check or uncheck the Enhancement box.

The enhancement that matters creates versions in different orientations. Google's wording: Google AI intelligently flips or extends your video in new aspect ratios while preserving the video's original content, and additional versions run automatically if they pass a quality review. The examples given are the original 16:9, a square 1:1 and a vertical 9:16.

One example carries a note that resets how you read the rest of the page: Model 2 uses generative AI to extend your original video in new aspect ratios. That is not a crop and not a letterbox. The pixels outside your approved composition are new, and nobody on your side has seen them before the ad serves.

Aspect ratio is a messaging decision, not a resize

Teams file aspect ratio under delivery specs. Wrong drawer. A 16:9 master is composed for a wide surface: product left of centre, pack shot right of centre, logo bottom right, qualifying line bottom left. Move it to 9:16 and roughly two thirds of the horizontal field is discarded or replaced by generated content. What survives is whatever the model judged to be the key element.

This is why AI video reframing is a creative decision with a media consequence. Google's Performance Max video guidance recommends at least one video in each orientation, horizontal, square and vertical, between 10 and 60 seconds. That recommendation exists because the placements are different surfaces, not because one frame can be reshaped to fit all of them without loss.

The upside is real. Fill the format gaps and a campaign can serve on Shorts, Discover and Gmail inventory it was locked out of, without three separate edits. It just arrives bundled with a version of the ad nobody approved.

Where reframed cuts break: supers, logos and legal lines

Start with burnt-in text. A lower-third super that sits comfortably inside 16:9 is often outside the 9:16 window entirely. Most of the failure modes that already apply to {{link}} get worse once the frame loses two thirds of its width. If a super carries the claim, the price or the call to action, the reframe has made it unreadable without changing a word of script.

Logos are the second casualty. Corner placement is a 16:9 habit. In a vertical cut the corner is cropped or re-composed, and a generated extension around a cropped logo produces the worst outcome: a partial mark on invented background.

Legal lines are the third, and the one with real exposure. Disclaimers, representative APRs, side-effect statements and offer qualifications sit at the bottom of a wide frame because that is where they intrude least. In a vertical reframe that strip goes first. Enhanced assets pass a quality review, but a quality review is not a legal review and not your brand's approval.

There is a second-order effect. Even when the reframe is clean, pacing changes: a shot that held on a pack for two seconds in 16:9 becomes a tight crop in 9:16, and the register shifts from considered to hard sell without anyone deciding that.

Most of the failure modes that already apply to AI video on-screen text get worse once the frame loses two thirds of its width.

Provenance does not survive the reframe

If your masters ship with Content Credentials, assume the manifest covers the file you uploaded, not the file that served. The C2PA specification is precise: a hard binding is one or more cryptographic hashes that uniquely identify the entire asset or a portion of it, and the values can match only that asset and no other, not even other assets derived from it or renditions produced from it.

The spec defines an asset rendition as a representation of an asset where the digital content has had a non-editorial transformation applied, such as re-encoding or scaling. A derived asset is created by actions that modify the digital content. A generative extension of the frame changes that content, so it is not a non-editorial transformation. Either way, the hard binding on your master does not travel.

What travels is a soft binding: a fingerprint, or an identifier embedded as an invisible watermark, computed from the digital content rather than its raw bits. The spec states that soft bindings are what enable derived assets and asset renditions to be identified. That gap is exactly what an {{link}} is supposed to close, and a reframe punches a hole in it.

The operational version is simple. Record every served variant: source asset ID, generated variant ID, placement, date first served, and who signed it off. Credentials on the master will not answer what actually ran.

That gap is exactly what an AI video creative audit trail is supposed to close, and a reframe punches a hole in it.

The disclosure question when a model invents pixels

Once the extension is generative, the output is no longer purely your footage. Under the AI Act transparency rules, in force from August 2026, providers of generative AI have to ensure that AI-generated content is identifiable, and certain AI-generated content should be clearly and visibly labelled, namely deep fakes and text published to inform the public on matters of public interest.

Read that against a reframe of a real location. A travel ad cut from a 16:9 master of an actual hotel terrace into a 9:16 vertical, with generated sky above and generated terrace edges below, now depicts a place that does not exist, rendered from one that does. The Commission puts the labelling duty on the content itself, and the commercial risk of getting it wrong lands on the brand, not the platform.

There is a nuance worth planning for. The obligation is written about content that is generated or manipulated, and a platform reframe is neither a routine delivery operation nor a generation your team commissioned. That ambiguity is a reason to settle your position before a campaign ships rather than after a complaint arrives.

The practical version is cheap. Decide in advance which reframed variants carry a visible label or a provenance record, document the decision, and apply it consistently across markets. A labelling policy you can produce on request is worth more than a legal opinion you cannot.

The pre-flight checklist for AI video reframing

First, inventory. Open every Performance Max campaign, go to Asset optimization, open the Video dropdown and record whether Enhancement is checked. Most accounts never touch it, so it is on. Do the same for Video and Demand Gen campaigns, where the control is the Get vertical versions of your videos checkbox.

Second, recompose for safe zones. Anything burnt in that carries meaning, supers, logos, prices, phone numbers, disclaimers, should sit inside the region that survives a 9:16 crop of a 16:9 frame. If the creative cannot be recomposed without losing the message, the answer is a native cut, not a tighter safe zone.

Third, decide per asset. A centre-framed product demo with no text and no recognisable location is a reasonable candidate for automated reframing. A regulated financial product, a health claim, a tourism asset built on a real property, or anything carrying a person's likeness is not.

Fourth, treat every generated variant as a new creative version with its own ID and preview. Run every generated cut through the same {{link}} you already use for synthetic creative. Fifth, re-bind provenance where it matters and keep your own variant record. Sixth, review monthly and pull any variant that has drifted from the approved message.

Run every generated cut through the same pre-ship trust QC gate you already use for synthetic creative.

When to ship native ratios instead of letting the model extend

The scale of the spend makes this worth deciding properly. IAB's 2026 Digital Video Ad Spend and Strategy Report projects digital video ad spend will surpass 80 billion dollars in 2026, continuing to outpace the broader ad market, and notes that generative AI adoption for video creative keeps accelerating while advertisers ask for more proof of performance.

The reach argument for {{link}} is real, which is why the answer is not to switch everything off. The rule: if the asset carries a claim, a price, a disclaimer, or a recognisable person or place, upload native 16:9, 1:1 and 9:16 masters and turn enhancement off for that asset group. If it is abstract, centre-framed and text-free, let automation fill the gap and monitor the variant report.

AI video reframing is not a scandal. It is a platform doing at delivery time what production teams used to do in an edit suite, at a fraction of the cost and with none of the briefing. Good trade for commodity assets, bad one where the frame itself carries the message. The only question is whether anyone on your side has looked at the version that ran.

The reach argument for generative video in performance media is real, which is why the answer is not to switch everything off.