Asset Studio and Gemini Omni in Google Ads 2026: Generating Creative and the Six Checks Before You Publish

Asset Studio in Google Ads stopped being a picture library in 2026 — which matters, because Performance Max and Demand Gen consume creative faster than any sanely sized design team can produce it. A single campaign wants headlines, descriptions, images in three aspect ratios, logos and video — ideally several variants of each, or the system has nothing to combine. The usual response is to produce less and quietly accept that half of Google’s inventory is closed to you.

In 2026 Google offered a different answer: move production inside the ad account. This piece covers what Asset Studio can do after this year’s updates, how to fit it into a working process, where automation ends and human accountability starts — and which creative requirements land this autumn.

Asset Studio in Google Ads: what it is and where to find it

Asset Studio is a section of the Google Ads interface (Tools → Asset Studio) that combines creating, editing and storing ad creative. Before 2026 it was mostly a library with a basic image editor: cropping to required ratios, background swaps, simple edits, with assets flowing into campaigns — primarily Performance Max and Demand Gen.

Over 2026 it turned into an actual production tool. The changes were announced at Google Marketing Live on May 20, and video generation began rolling out on August 25.

What arrived in 2026

Gemini Omni: video generation inside the account

The headline update is the integration of Gemini Omni, a multimodal model that generates video alongside images and text. Video for Demand Gen and YouTube previously had to be produced elsewhere and uploaded; now it can be assembled in the same place the rest of your assets live.

What is on offer in practice:

• storyboards and motion scenes generated from your brand guidelines and website URL — the model works from your material rather than inventing a look;
• both 16:9 horizontal and 9:16 vertical formats, so YouTube and Shorts come out of one brief;
• edits by natural-language prompt instead of rebuilding a cut from scratch;
• export straight into campaigns, skipping the download-and-reupload loop.

Natural-language creation

The second change is quieter but more consequential over time: the interface accepts a plain description of what you want instead of field-by-field configuration. You state the requirement, look at the result, refine the prompt. That shifts the unit of work — you are no longer “making a banner,” you are iterating a brief.

One-click creative testing

The third update closes the gap that left most generated creative untested. Straight after building an asset you can launch an experiment: the interface shows control and treatment arms side by side, you set the traffic split and duration, and launch with a single button. Previously that lived in a separate part of the platform, which is why it often never happened.

Three scenarios where this actually changes things

Scenario 1: fixing variant scarcity. The most common cause of weak Performance Max results is three images and one headline per asset group. The system has nothing to combine, so it serves the same thing repeatedly. In-account generation closes that gap in an hour rather than a sprint.

Scenario 2: geo localisation. Launching in five markets used to mean five rounds with a vendor. Now the base set is adapted inside the tool and the external work shrinks to a native-speaker review — which, tempting as it is, you do not skip.

Scenario 3: seasonal refresh. Ahead of a demand peak the visuals should change, but reshooting everything is expensive. Rebuilding scenes and emphasis from existing material takes hours. The broader channel logic is covered in our guide to Demand Gen in Google Ads.

The workflow: from brief to experiment

1. Assemble the inputs. Logos, palette, fonts, product photography, website URL, three core messages. The sharper the input, the less generic stock you get back.
2. Write the brief as text. Not “make it look good” — audience, format, placement, what must be in frame, what must not, and the call to action.
3. Generate a batch. Six to ten variants per hypothesis is a workable benchmark. One variant has nothing to compare against; thirty cannot be meaningfully reviewed.
4. Reject by hand. This is where the six checks below apply. Typically fewer than half survive to campaign.
5. Run the test. Control against the new variant, a fixed duration, and a hypothesis written down before launch.
6. Feed the finding back. Whatever won — a format, an opening shot type, a runtime — belongs in the next brief, or every cycle restarts from zero.

What a working generation brief looks like

Output quality is almost entirely determined by input quality. “Make a banner for an online store” returns exactly what you would expect: tidy, faceless and indistinguishable from a thousand others. A working brief has six blocks.

Block	What to specify	Common failure
Audience	Who they are, in what situation they see this, what they already know	“Anyone interested in our product”
Message	One idea, in your own words rather than a slogan	Three benefits crammed into one frame
Format and placement	Ratio, duration, where it runs, with or without sound	One universal asset “for everything”
What is in frame	Product, setting, action, mood, pacing	Leaving it blank so the model decides
What must not appear	Prohibited claims, promises, visual clichés	Skipping this block and reworking later
Desired action	What the viewer should do and why now	A CTA with no reason attached

The “what must not appear” block is the one everyone forgets, and the one that saves the most time. If your category forbids promising outcomes, guaranteeing timelines or depicting people in certain situations, that belongs in the prompt from the start — not discovered during review of the third batch.

How not to fool yourself with one-click testing

Ease of launch is also the risk: a test takes ten seconds to start and conclusions get drawn on insufficient data. Three rules keep it honest:

• One variable. If the new variant differs in image, headline and video at once, you learn “better or worse” but never why.
• A decision rule written in advance. Which metric decides and what difference counts as a result — agreed before launch, not selected afterwards from whatever moved.
• Sufficient duration. A creative test stopped on day three usually measures novelty, not quality. The mechanics of clean experiments are covered in our piece on experiments in Google Ads.

Generation removes the constraint on volume, not the constraint on meaning. Twenty versions of a weak message lose to three versions of a strong one — the model will not work out for you why people buy.

Six checks before publishing

A generated asset is a draft, and you own it — in front of the platform, the regulator and the client. The minimum review set:

1. Facts and figures. Price, delivery time, discount size, ingredients, warranty. The model confidently supplies plausible values, and plausible is not the same as yours.
2. Brand. Logo undistorted, colours from the palette, correct typeface, brand name spelled properly. Generated text inside an image is its own risk zone.
3. Legal. Mandatory disclaimers, category restrictions, local regulatory requirements. No model knows your compliance obligations.
4. On-screen text. Proof it letter by letter, especially in non-English languages, and confirm it is legible on a phone rather than collapsing into a grey smear.
5. Landing page match. The promise in the creative has to exist on the page behind the click. A mismatch costs you conversion rate and quality signals at once.
6. Cultural context. Gestures, scenarios, interiors and casting that read as normal in one market can read as odd in another. Only someone from that market catches this.

Where generation reliably underperforms

An honest list of limitations is more useful than a feature list — it saves review time.

• Small on-screen text. The most common defect: letters warp, characters repeat, non-English scripts sprout characters that do not exist. The practical workaround is to keep text in a layer you control rather than asking the model to paint it into the image.
• Exact product geometry. If your product has a recognisable silhouette, a logo on the body or complex texture, the generated version will be “similar” rather than correct. In those categories your photo is the base and generation handles the surroundings.
• Numbers and terms. Prices, percentages and timelines get filled in confidently and wrongly. Verify every figure manually, even one that looks familiar.
• Hands, faces and fine motion in video. The classic weak spot: forgivable in a still, glaring in motion, and corrosive to brand trust.
• Market specificity. Interiors, clothing, casting and everyday detail default to an averaged Western picture. In local markets viewers notice, and response drops.

The conclusion is simple: generation is excellent at volume and variation and poor at precision. The best operating mode is not “make me an ad” but “multiply and adapt what I have already validated.”

How assets affect ad metrics

Asset variety directly determines how many combinations the system can assemble, and indirectly influences Ad Strength. Do not confuse cause with effect: Ad Strength rates the variety and relevance of your material, not future performance. Chasing an “Excellent” label for its own sake is pointless; what is genuinely worth changing based on the asset report is covered in Ad Strength and RSA asset reporting.

Practical rule: cover every required format including square and vertical, keep at least three or four variants of each type, and retire assets with consistently low performance ratings rather than the ones you personally like least. Search ad assets get their own treatment in our guide to ad extensions (assets) in Google Ads.

Changing creative requirements

Alongside the new capabilities, Google is tightening mandatory elements. Both items below belong in your production plan.

• Text disclaimers in responsive search ads. Introduced in August 2026, this format guarantees that important terms, conditions or regulatory disclosures appear in the description area. For regulated categories it is a way to place a disclaimer where it is actually seen rather than in small print on the landing page.
• Business name and logo in YouTube responsive ads. From October 12, 2026 they become mandatory when creating or updating those ads, unless default branding is set at the advertiser level. Practical takeaway: set advertiser-level default branding now, or every ad edit stalls on the same day in October.

What stays with people

The division of labour that works: the model produces variants, the human makes decisions. What remains human is everything generation cannot reach — knowing what you are actually paid for, access to real customer objections, accountability for facts, and judgement about which argument lands in your market.

There is a practical argument against a fully automated pipeline too. If your creative is assembled from the same public signals as your competitors’ creative, it will look like your competitors’ creative. The differentiator is not the tool but the input: your photography, your phrasing, your data about what actually stops a buyer. Turning that data into targeting is covered in our piece on custom segments and intent audiences.

Fitting this into a team process

An in-account tool changes not just speed but roles. A structure that survives past the first month:

1. A weekly creative review. Read the asset report: what is serving, what is rated low, where variants have run out. Twenty minutes, a fixed day.
2. One hypothesis per week. Not five. It gets stated, generated as a batch, reviewed, and pushed into a test. That yields ten to twelve validated findings a quarter instead of a pile of half-finished tests.
3. One named owner for review. The six checks work only when someone is personally accountable. Shared responsibility here means none, and the cost of a mistake is public: a wrong price in an ad is visible to everyone immediately.
4. A winners library. Everything that wins a test goes into one document, annotated with what made it win. That is your team’s accumulated knowledge; without it every new hire restarts the search from scratch.
5. A quarterly audit. Assets fatigue: what worked in March is worn out by September. Once a quarter, check whether half your budget is running on year-old material.

Implementation checklist

• Brand guidelines, logos and product photography uploaded, so the model works from your material.
• Advertiser-level default branding set, ahead of the October requirement.
• Every image and video format covered, with at least three to four variants per type.
• A written six-point acceptance checklist that every asset passes.
• Each new creative batch goes through an experiment rather than being silently swapped in.
• Test results recorded in one document that feeds the next brief.
• An asset breakdown in reporting — how to build it is covered in our guide to Google Ads reporting in Data Studio.

For product advertising, settle which products deserve promotion before you invest in creative for them — catalogue segmentation logic is in our piece on custom labels in the product feed. Every other technical breakdown lives in the PPC Rebels blog, and the team’s tooling is on the services page and under agency Google Ads accounts.

FAQ: Asset Studio and creative generation

Where is Asset Studio in the interface?

Under Tools → Asset Studio, alongside the asset library, editor and generation tools. Individual features may differ by account and language — the rollout is gradual.

What does Gemini Omni actually add?

In-account video generation: storyboards and scenes built from your brand assets and website URL, horizontal and vertical formats, prompt-based editing, and export of finished videos straight into campaigns.

Which campaign types does this apply to?

Primarily Performance Max and Demand Gen, plus other YouTube and Google campaigns that need video and images. Those types depend most heavily on asset variety.

Do generated creatives need to be disclosed?

Disclosure requirements vary by market and continue to evolve, so follow local law and platform policy for your category. Practical rule: if the material could be taken as documentary footage or as a statement by a real person, that is a distinct risk to review.

Does Asset Studio replace a designer?

No. It removes a volume bottleneck — the number of variants and formats. Framing the task, selecting, brand fit and accountability for facts stay with a person, and that part decides the outcome.

How many variants should one batch contain?

Six to ten per hypothesis is a workable benchmark. Fewer leaves nothing to compare; more cannot be reviewed or tested properly.

How do I know a new creative is better?

Only through an experiment with a control arm, a metric fixed in advance and adequate duration. Week-over-week “before and after” mixes creative effect with seasonality, bid changes and algorithm learning.

What should I do with old assets?

Do not delete everything at once. Retire one or two consistently low-rated assets at a time. Replacing an entire library in one go restarts learning and buys you a week of unstable results.

What are the new text disclaimers in search ads?

A format introduced in August 2026: a text disclaimer guaranteed to appear in the description area, letting you surface important conditions or required disclosures in the ad itself.

What changes for YouTube ads in October?

From October 12, 2026, business name and logo become mandatory when creating or updating YouTube responsive ads, unless default branding is set at the advertiser level. Set it in advance so ad edits do not all stall at once.

Why does text in generated images come out garbled?

A known weakness of generative models, most visible in non-English scripts. The reliable workaround is to avoid asking the model to render text inside the image and to overlay it as a layer you control.

Can I use generation if my product has a distinctive design?

Carefully. The model reproduces a similar but not identical form, and for products with a recognisable silhouette or a logo on the body, buyers notice. In those categories your photograph is the base and generation handles background, scene and format adaptation.

Is generated video suitable for the top of the funnel?

Yes — it is the most practical use case. Demand Gen and YouTube need volume and vertical formats, and the cost of a miss is lower than in bottom-funnel performance campaigns. The demand on the opening seconds is unchanged though: they decide whether anyone watches the rest.

The takeaway: the tool removes a shortage of variants, not a shortage of meaning. The winner is not whoever generated more, but whoever put something competitors do not have into the input and then verified the output with an honest test.