What problem does it solve?
A paid media campaign needs many versions of the same idea: different sizes for different placements, different hero images for different audience segments, different hooks to test against each other, refreshed often enough that people do not tune the ad out. Producing that many variants by hand, with a photo or video shoot behind each one, is slow and expensive. Google frames this from the platform side: asset variety is a key ingredient of a successful campaign, and advertisers have said that creating and scaling assets can be one of the hardest parts of building and optimizing one.
Lysol describes the same bottleneck from the brand side: building high quality, impactful creative remains an expensive, time consuming process, and the team draws a clear line between experimenting with generative AI tools and actually shipping AI built creative.
How does it work?
- Brief the model. A marketer describes the product, the audience and the message, and can ground the brief in market and competitor research the model helps assemble.
- Generate variants at scale. The model produces many concepts and asset variations: image backgrounds, expanded or resized images, video cuts, headlines and descriptions, far more than a team could brief and shoot by hand.
- Score before spending media budget. Some platforms score generated assets against predicted performance, or against a small real budget test, before a variant reaches full campaign spend.
- A human picks and approves. A marketer reviews the generated set, keeps or edits what works, and rejects what does not; nothing generated is required to go live unreviewed.
- Test and learn in production. Winning variants run at scale, the campaign platform keeps testing new ones against them, and results feed back into the next brief.
- Audience
- Employee facing
- Autonomy
- Copilot
- Adoption
- Early adopters
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Cost reduction | Too few to pool | 4% to 80% | 2 | 1 organization, 1 vendor |
| Productivity gain | Too few to pool | 5x to 46x | 2 | 2 vendor |
Value drivers: Speed and cycle time, Employee productivity, Revenue growth.
Indicative value
A consumer brand running USD 20 million a year in paid social and search media
USD 600,000 to USD 2 million
Creative production cost avoided per year
How this is calculated
Formula: annualMediaSpend * creativeProductionShare * costReduction. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Annual paid media spend annualMediaSpend, USD per year | 20,000,000 | 20,000,000 | The reference brand. |
| Share of media spend that a comparable creative production budget represents creativeProductionShare, fraction of media spend | 0.1 | 0.2 | Editorial assumption, replace with your own creative production budget as a share of media spend. |
| Reduction in cost per creative asset from generative production costReduction, fraction of creative production cost | 0.3 | 0.5 | Conservative against the benchmark on this page (Lysol reports an 80% reduction in cost per asset for its generative AI pilot), because that figure covers one 8 week pilot on one product line, not sustained production economics across a whole media budget. |
What it leaves out: Gross production cost avoided only. It leaves out the cost of the generative tools and the human review time they still need, and it makes no claim about whether AI generated creative performs better, the same, or worse than what it replaces; Lysol reports only that its best AI asset performed nearly identical to its best traditional asset in short term sales lift studies, not that it beat it.
Who already uses it?
8 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
bol
Netherlands · Retail and ecommerce · 2026
Online retailer bol built a "digital twin" of a real fashion model with generative AI: a large photo library of the model in different poses and lighting is used to generate new campaign images of that same person in different outfits and settings, without a further physical photoshoot. After two and a half years of internal testing, shadow mode comparisons and a small customer pilot, bol's first AI generated fashion campaigns went live in early March 2026 for both women's and men's ranges. Models give explicit, scoped consent for the use of their digital twin.
No outcome disclosed.
Lysol (Reckitt)
United States · Manufacturing · 2026
Lysol, a Reckitt brand, piloted a fully generative AI creative development process for its Pivot Laundry Sanitizer product, in partnership with BCG and Google. Over an 8 week sprint, the team used Gemini to generate and score hundreds of creative concepts, then used Veo and Imagen for synthetic production of broadcast ready 15 and 30 second video spots, with consenting, compensated actors used as the basis for AI generated likenesses, replacing a physical shoot.
- Cost reduction: 80%, 8 week pilot, one product line
"We achieved an 80% reduction in cost per asset compared with our traditional process."
Claimed by: organization
Alibaba Group
China · Retail and ecommerce · 2018
Alibaba built Luban, an AI graphic design system that generates banner and promotional images for Tmall and Taobao merchants, using a deep learning design algorithm on top of a knowledge base built from millions of analyzed marketing pictures. A user enters the desired style and size and Luban handles material analysis, retouching, color blending and layout, generating several design options in real time. During the 2018 Double 11 shopping festival Luban designed about 6 million banners for merchants on the platform, and Alibaba Cloud also offered the same design capability to customers in other industries.
- Users served: 200,000, cumulative during the 2018 Double 11 Shopping Festival
"During the 2018 Alibaba Double 11 Shopping Festival, Luban has cumulatively designed about 6 million banners for 200,000 enterprises/merchants."
Claimed by: organization
Falcon Game Studio (OneSoft)
Vietnam · Media and entertainment · 2026
Falcon Game Studio, the Vietnam based arcade and puzzle developer owned by OneSoft, wanted to expand its "Goods Sorting" game beyond its core female audience. It used Gemini to evaluate over 420 existing video assets for the visual and structural elements that drove engagement, then used Veo 3 to turn static images into new creatives aimed at men and adults aged 35 to 44. Google's case study reports the resulting campaign, the biggest the studio had run for the game, scored 1.1 times its target return on ad spend in its first month, with retention for the new male segment matching the existing female audience.
No outcome disclosed.
Funsol Technologies
Pakistan · Technology and software · 2026
Funsol Technologies, a Pakistan based gaming, utility app and custom software developer, built custom automation tools that plug directly into Gemini to scale its ad creative production. Google's case study reports the workflow generated 46 times more image assets a day and produced over 100,000 image variations within a month, up from 750 a month before, plus 10 times more video assets a day, with 79% of the resulting creative meeting the brand's own performance benchmarks.
- Productivity gain: 46x, daily image asset output
"Guided by insights from Gemini's real-time scan of market trends like search intent, the workflow generated 46X more image assets daily."
Claimed by: vendor - Productivity gain: 10x, daily video asset output
"The output lift: 10X more video assets daily."
Claimed by: vendor
SeaBank Indonesia
Indonesia · Banking · 2026
Digital banking service SeaBank Indonesia partnered with Google on a pilot to create customised video ads for 18 to 30 year old app users, using Gemini to generate character personas, Nano Banana for photorealistic imagery and Veo 3 to produce the finished videos. In an A/B test against conventional ads, Google's case study reports the AI generated, customised ads reduced cost per install by 4% and cost per conversion by 3%, alongside higher click through rates, while running in a fraction of the traditional production time.
- Cost reduction: 4%, cost per install
"That reduced cost per install by 4% and cost per conversion by 3%, while enabling SeaBank to successfully scale its digital financial inclusion effort across Indonesia."
Claimed by: vendor - Cost reduction: 3%, cost per conversion
"That reduced cost per install by 4% and cost per conversion by 3%, while enabling SeaBank to successfully scale its digital financial inclusion effort across Indonesia."
Claimed by: vendor
Event Tickets Center
United States · Retail and ecommerce · 2024
Event Tickets Center was one of the earliest beta testers of Google's asset generation feature in Performance Max. Google says the feature helped the Event Tickets Center team accelerate creative production 5x, with less time and effort.
- Productivity gain: 5x, beta period, Performance Max asset generation
"Event Tickets Center was one of the earliest beta testers for asset generation in Performance Max, which has helped the team accelerate creative production by 5x with less time and effort."
Claimed by: vendor
1-800-FLOWERS.COM, Inc.
United States · Retail and ecommerce · 2023
1-800-FLOWERS.COM, Inc. adopted Google's generative asset feature in Performance Max. Its CMO, Jason John, says the generated assets save his creative team valuable time and let it craft personalized visuals that resonate with its audience.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Brand guidelines: fonts, colors, tone of voice and claims the brand can substantiate
- A library of approved product images or footage to ground generated variants
- Historical creative performance data to brief and later score new variants against
Systems to integrate
- The paid media platforms running the campaign (their own generative asset tools)
- Digital asset management or brand guideline systems
- Rights and consent records for any person whose likeness or voice is used
Complexity: Low
The generative tools themselves are available inside the major ad platforms with little setup. The real work is process and governance: brand guidelines the model must follow, a review step before anything publishes, and rights clearance for any real person's likeness or voice used as a generation basis.
- 1
Start with one campaign and one clear brief
Pick a live campaign with a well defined product and message, as Lysol did for one product line, rather than trying to generalize the process across every brand at once.
- 2
Set a responsible AI policy before generating any likeness
Decide up front how consent and compensation work for any real person's face or voice used as the basis for AI generated video or images, and put it in writing before the pilot starts.
- 3
Generate wide, then narrow with a human
Let the model produce far more concepts and variants than you would ever brief by hand, then have the marketing team pick the strongest few to refine, rather than publishing the first pass.
- 4
Score against real performance, not opinion
Compare generated variants against your best existing asset with an actual sales or conversion lift test, not only internal preference, before scaling media spend behind them.
- 5
Only then widen the process
Once a pilot shows generated creative can match or beat the traditional process on cost and performance, extend the brief, review and scoring steps to more product lines.
Guardrails
- Every generated asset reviewed by a marketer against brand guidelines before it can publish
- Written consent and compensation on file for any real person's likeness or voice used to generate creative
- Generated content labelled as AI generated wherever the platform or the law requires it
- No generated claim that goes beyond what the brand can substantiate
KPIs to instrument
- Cost per finished creative asset, generated versus the traditional process
- Time from brief to a launch ready asset
- Performance of generated variants against the best existing asset, on the same audience
- Share of generated variants a reviewer rejects or substantially edits before publishing
Human in the loop
Marketers write the brief, review every generated variant before it is added to a live campaign, and own the responsible AI policy for using anyone's likeness or voice. A brand or legal reviewer signs off before a new product line or a new use of a real person's likeness goes live.
Common failure modes
- Off brand or exaggerated output
- The model drifts from brand guidelines or overstates a product claim. Catch it with a mandatory brand and claims review before anything publishes.
- A likeness used without real consent
- An AI generated video uses a real person's face or voice without clear, compensated consent. Require signed consent on file before generation, not after.
- Comparing against a weak baseline
- A generated variant looks like a win only because the traditional asset it replaced was already underperforming. Test against your actual best asset, not an average one.
- Volume without a review bottleneck plan
- Generating far more variants than the team can review pushes weak review, or a temptation to skip it. Size the generation volume to what the review process can actually handle.
What are the risks and rules?
EU AI Act
Limited risk (transparency)
Article 50(2) puts a machine readable marking duty on the provider of the generative tool itself (Google, for example, marks Performance Max images with SynthID), with an exception only where the tool performs assistive standard editing or does not substantially alter the input.
Article 50(4) separately requires the deploying brand to disclose when it generates or manipulates image, audio or video content that is a deep fake: content that resembles a real person, place or event and would falsely appear authentic, such as the synthetic "digital twin" likenesses used in the Lysol pilot.
Content that is evidently artistic, creative, satirical or fictional is not exempt from that disclosure; the article only lets the brand make it in a manner that does not hamper the display or enjoyment of the work.
The separate exception for content under human review or editorial control, with someone holding editorial responsibility, applies only to AI generated text published to inform the public on matters of public interest, not to image or video ad creative, so it does not exempt anything on this page. Lysol's pilot is a US deployment and the source does not describe it as operating under the EU AI Act.
Controls to put in place
- Written, dated consent from any real person whose likeness or voice is used as a generation basis
- A brand and legal review step before any generated asset reaches a live campaign
- AI generated content labelled in line with the platform's and the law's disclosure rules
- Log of which assets in a campaign were AI generated, for audit and disclosure
Frequently asked questions
- Does AI generated ad creative actually perform better than creative made the normal way?
The public evidence on this page does not show a performance win, only a cost and speed one. Lysol reports that its best generative AI asset performed "nearly identical" to its best traditional asset in short term sales lift studies. Google says its asset generation feature helped Event Tickets Center speed up creative production 5x, and 1-800-FLOWERS.COM's CMO says the same feature saves the creative team time. Treat generative creative as a way to produce and test more variants for the same budget, not as a guaranteed lift.
- Who owns the risk if an AI generated ad uses someone's likeness?
Primarily the brand running the campaign, though the agency, the ad platform and the model provider can share responsibility depending on their role and contract. Lysol's pilot addressed the likeness risk by working with consenting, compensated actors as the basis for its AI generated likenesses and setting a responsible AI policy before generating anything, which is the standard any deployment should meet.
- Do AI generated ads need to be labelled as AI generated?
Yes, in the EU, where the AI Act's transparency duties have applied since 2 August 2026. Article 50(2) requires the provider of the generative tool to mark its own output as artificially generated in a machine readable way. Article 50(4) separately requires the deploying brand to disclose when an ad is a deep fake, meaning it resembles a real person, place or event and would falsely appear authentic; content that is evidently artistic, creative or fictional still needs disclosure, only in a manner that does not hamper the work. Google says it watermarks generated Performance Max images with SynthID and adds open metadata identifying them as AI generated.
How to cite this page
Blits.ai AI Use Case Library, "AI generated ad creative production and testing", last verified 29 September 2026, https://www.blits.ai/ai-use-cases/ad-creative-generation-and-testing. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 1 October 2026: Added deployments: Alibaba Group.
- 30 September 2026: Added deployments: bol.
- 29 September 2026: First published