Advertising
Last reviewed
May 13, 2026
Sources
68 citations
Review status
Source-backed
Revision
v2 · 5,492 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 13, 2026
Sources
68 citations
Review status
Source-backed
Revision
v2 · 5,492 words
Add missing citations, update stale details, or suggest a clearer explanation.
See also: Advertising ChatGPT Plugins
Advertising is one of the largest commercial applications of artificial intelligence. AI now touches almost every step of the modern ad cycle: deciding which impression to buy in a real-time bidding auction, generating the image and copy that appears in that impression, choosing which audience segment to target, classifying the surrounding content for brand safety, and measuring the campaign's effect on sales. The industry has moved through three rough technical eras: programmatic buying with predictive bidding models, machine learning auction systems integrated into walled-garden platforms such as Google Ads and Meta Ads, and a generative phase that began around 2022 in which large language models and text-to-image tools produce the creative itself.
The shift has been fast and disorderly. Brands such as Heinz and Coca-Cola used generative AI to win press coverage as early as 2022 and 2023. By 2024 the first major commercials made almost entirely with OpenAI's Sora and other diffusion video tools were airing during prime ad festivals. The same year saw fierce public backlash against Coca-Cola's AI-generated remake of its "Holidays Are Coming" spot and growing concern about AI-generated political ads, deepfake robocalls, and AI "influencers" with millions of followers. Regulators in the United States, the European Union, and elsewhere responded with disclosure rules.
AI in advertising covers a wide range of overlapping technologies and use cases. The core categories are:
| Layer | What AI does | Representative products |
|---|---|---|
| Programmatic buying | Predicts the value of each impression and bids in real time. | The Trade Desk, Google Display & Video 360, Amazon DSP, Yahoo DSP |
| Auction-time bidding inside walled gardens | Sets prices and targeting per auction inside platform inventory. | Google Smart Bidding, Meta auction system, TikTok Smart+ |
| Performance campaigns | Single goal-based campaigns that span an entire platform. | Google Performance Max, Meta Advantage+, Pinterest Performance+, Snap Smart Campaign Solutions |
| Generative creative | Produces images, video, copy, voice, and variations. | Adobe Firefly, DALL-E, Midjourney, Stable Diffusion, Sora, Meta Advantage+ Creative |
| Virtual humans | Synthetic models, influencers, and brand personas. | Lil Miquela, imma, Shudu, Aitana López |
| Brand safety and measurement | Classifies content and detects fraud and ad context. | DoubleVerify, Integral Ad Science, Pixalate, Zefr |
| Marketing mix modeling | Estimates the sales lift of each channel. | Meta Robyn, Google Lightweight MMM, commercial MMM vendors |
These layers feed each other. A generative tool may produce 200 variations of an image. An auction-time bidder then decides which variation to show which user. A brand-safety system decides which placements are eligible at all. A mix model estimates whether the resulting spend actually moved sales.
Programmatic advertising as a recognisable industry started with display banner auctions. Right Media, founded by Michael Walrath in 2003, launched the Right Media Exchange in April 2005 and was sold to Yahoo in 2007 for about $680 million. The Right Media Exchange is widely cited as the first online ad exchange. Its CTO Brian O'Kelley left after the Yahoo acquisition and co-founded AppNexus in 2007. He is often credited with commercialising real-time bidding (RTB) and is sometimes called the "godfather of ad tech".
The years 2007 to 2011 saw a wave of new platforms: Google acquired DoubleClick in 2008 and built it into what is now Google Ad Manager and Display & Video 360. MediaMath, founded in 2007, marketed itself as the first demand-side platform. Quantcast, founded in 2006 by Konrad Feldman, brought predictive modelling to audience targeting. The Trade Desk was founded in 2009 by Jeff Green and Dave Pickles, both veterans of Microsoft adECN, and built a self-serve DSP that became dominant in the open web. Supply-side platforms such as PubMatic (founded 2006), OpenX (founded 2008), Rubicon Project (founded 2007, renamed Magnite in 2020 after the merger with Telaria), and Index Exchange (founded 2001, refocused on advertising) sat on the publisher side.
Machine learning lived inside these systems from the start. Bidders had to estimate click-through and conversion rates in milliseconds, and the only feasible approach was a logistic-regression or gradient-boosted model trained on prior auctions. That same machinery later supported algorithmic creative selection and audience expansion. AppNexus shipped a machine-learning optimisation engine called Xandr Invest after AT&T rebranded the company in 2018. Microsoft acquired Xandr in 2021 and consolidated it under Microsoft Advertising in 2023, and the Microsoft Invest legacy stack is being wound down in early 2026.
The second phase moved the math from independent DSPs into the platforms themselves. Google introduced Smart Bidding under that brand in 2016 as a family of auction-time strategies. Target CPA, Target ROAS, Maximize Conversions, and Maximize Conversion Value all sit inside Smart Bidding, with the system pricing each auction using device, time, location, audience, and other live signals. Smart Bidding became the default recommendation for most Google Ads campaigns by the late 2010s.
Google then attacked the older idea of tightly controlled keyword targeting. The company's broad match keyword type was rebuilt around large language models. Google says broad match is 73% more accurate at predicting conversion intent than in 2023, and 62% of advertisers running Smart Bidding now use broad match as their primary match type.
Meta took a similar path on the social side. Its detailed targeting controls were progressively replaced by an automated audience system that picked the buyer for each impression using a deep learning ranker.
The third phase started when text-to-image and text-to-video models became good enough for commercial creative. Heinz's "AI Ketchup" campaign in July 2022 used DALL-E 2 to draw "ketchup" and discovered that almost every output looked like a Heinz bottle. The Cosmopolitan magazine cover generated by DALL-E 2 with Karen X. Cheng appeared in June 2022. By 2023 Coca-Cola had launched the "Create Real Magic" platform built on GPT-4 and DALL-E in partnership with OpenAI and Bain & Company, and aired its "Masterpiece" television spot, which combined live action and Stable Diffusion frame transitions. By 2024 Toys "R" Us premiered the first major brand film made with Sora, and Coca-Cola produced an AI-generated remake of its "Holidays Are Coming" Christmas ad. The generative era also drove agency reorganisations, NVIDIA partnerships, and the first wave of cost cuts at companies that began running AI-only campaigns.
Programmatic advertising is the automated buying and selling of digital ad inventory. It currently makes up the majority of digital display spending worldwide. The main pieces are:
AI lives at every layer. Bidders use predictive models trained on conversion data to estimate the value of each user view. SSPs use machine learning to choose the floor price and to predict which buyer is most likely to win. Ad servers use models to choose which creative to deliver. Fraud detection and brand-safety filtering, described below, also run on AI.
The Trade Desk has built much of its product on a system it calls Koa, which surfaces machine-learning recommendations to traders. Quantcast uses its measurement panel to model audience interests directly. Amazon DSP uses Amazon's own retail purchase data as a training signal for prediction.
The largest advertising platforms have all converged on a single product format: a goal-based campaign that uses AI to handle bidding, targeting, placement, and creative selection across the platform's full surface area. The naming differs by vendor.
| Platform | Product | Launched | Notes |
|---|---|---|---|
| Google Ads | Performance Max | General availability November 2021 | Runs across Search, Display, YouTube, Discover, Gmail, and Maps. Replaced Smart Shopping and Local campaigns in 2022. |
| Meta | Advantage+ Shopping Campaigns (ASC) | August 15, 2022 | Automates up to 150 creative combinations and targets buyers using Meta's deep learning models. |
| Meta | Advantage+ Creative | October 2023 | Adds generative AI features for background generation, image expansion, and text variations. Full global rollout completed during 2024. |
| Microsoft Advertising | Performance Max | July 2023 | Brought the Performance Max format into Bing and the Microsoft Audience Network. Microsoft launched a follow-on AI Max for Search pilot in 2025. |
| TikTok | Smart+ | October 2024 | An AI-driven performance suite covering web, app, catalog, and lead campaigns. Builds on earlier Smart Performance Campaigns. |
| Performance+ | Public launch October 2024 (preview at Cannes Lions 2024) | Halves the number of inputs an advertiser provides, applies machine learning to targeting and budget allocation. | |
| Snap | Smart Campaign Solutions (Smart Bidding, Smart Budget, Smart Audience, Smart Ads) | 2024 onward, with successive features and AI Sponsored Snaps in 2025-2026 | Snap's answer to Performance Max and Advantage+. |
The trade-off in all of these products is the same. The advertiser gives up most of the manual control, in exchange for performance gains driven by the platform's own data and models. Pinterest reported that early Performance+ tests improved CPA by at least 10% in conversion campaigns. Meta's own A/B testing put Advantage+ Shopping Campaigns at a 12% lower cost per purchase than control campaigns. TikTok said its Smart+ web objective improved return on ad spend by 52% during closed beta tests and dropped cost per acquisition by 36% versus manual campaigns.
Generative AI has shifted from a curiosity to a default part of the ad production pipeline in less than three years. The most widely used systems include:
Video generation entered the same workflow. By 2024 OpenAI's Sora, Runway's Gen-2 and Gen-3, Google DeepMind's Veo, and Pika Labs were all being used to produce social and broadcast spots, often for shorter cuts and product demos.
| Campaign | Brand | Year | Agency or partner | AI technique |
|---|---|---|---|---|
| AI Ketchup | Heinz (Kraft Heinz) | July 2022 | Rethink (Toronto/Montreal/Vancouver) | DALL-E 2 prompts to draw "ketchup"; many results resembled Heinz bottles. Won a Clio Gold in 2023. |
| First AI magazine cover | Cosmopolitan | June 2022 | Karen X. Cheng with OpenAI | DALL-E 2 generated the cover image in roughly 20 seconds. |
| Shah Rukh Khan My Ad / Not Just A Cadbury Ad 2.0 | Cadbury (Mondelez India) | Diwali 2021 | Ogilvy India, Wavemaker, Rephrase.ai | AI voice and face synthesis let more than 2,000 small shops insert their name into a Shah Rukh Khan endorsement. |
| Create Real Magic | Coca-Cola | March 2023 | OpenAI, Bain & Company | First platform combining GPT-4 and DALL-E. Fans could remix Coca-Cola brand assets. Followed by a Real Magic Creative Academy in Atlanta. |
| Masterpiece | Coca-Cola | March 2023 | Electric Theatre Collective | Live action plus Stable Diffusion frame transitions between artworks. Won a Yellow Pencil at the D&AD Awards 2023 for Visual Effects. |
| Vogue Singapore Roots cover | Vogue Singapore | March 2023 | Varun Gupta | Used Midjourney and DALL-E to create three AI cover models, Aadhya, Melur, and Faye, inspired by early 20th-century Southeast Asian portraits. |
| The Origin of Toys "R" Us | Toys "R" Us | June 2024 (Cannes Lions premiere) | Native Foreign | First major brand film made primarily with OpenAI's Sora. About 80% to 85% AI-generated with VFX on top. |
| Holidays Are Coming (AI remake) | Coca-Cola | November 2024 | Secret Level (Los Angeles), Silverside AI (San Francisco), Wild Card (Kuala Lumpur) | Four generative models, 10,000 frames, 5,000 video segments, over 40 creatives across four continents. Produced in two months. |
| 30 AI campaigns | Klarna | 2024 | Internal teams using Midjourney, DALL-E, Firefly, and an in-house Copy Assistant | Said to have replaced significant agency spend; trimmed image production cycle from six weeks to one. |
| Beauty Genius | L'Oreal Paris | Launched October 2024, showcased at CES January 2025 | L'Oreal internal | Combines generative AI, computer vision, augmented reality, and 10 LLMs. Trained on 6,000 images and 10,000 products. |
Virtual influencers existed before generative AI, but generative tools have made them faster and cheaper to operate. The earliest wave used 3D rendering and photo compositing. The current wave uses diffusion models, animation systems, and synthetic voice tools.
| Persona | Creator or studio | Launched | Notes |
|---|---|---|---|
| Lil Miquela | Trevor McFedries and Sara Decou, Brud | April 2016 | Half-Brazilian, half-Spanish 19-year-old Los Angeles persona. Brud raised $6 million in 2018. Endorsement deals with Prada, Calvin Klein, Samsung, and others. |
| Shudu | Cameron-James Wilson, The Diigitals | April 2017 | Marketed as the world's first digital supermodel. Initially built in 3D modelling tools after Wilson became frustrated with hand-painting Barbies. |
| imma | Aww Inc. (Koichi Kishimoto, Takayuki Moriya) | 2018 | Japan's first "full-fledged" virtual model. Brand deals include IKEA, Porsche Japan, Dior, Valentino, Calvin Klein, Nike, and Nomura. Aww partnered with NVIDIA in 2024 to upgrade an interactive "AI imma". |
| Aitana López | The Clueless agency (Rubén Cruz), Barcelona | July 2023 | Spain's first hyper-realistic AI influencer. Reportedly earns €3,000 to €10,000 per month. Promoted by Amazon, Razer, and Freepik. |
| Esther Olofsson | RauwCC, Rotterdam | Existed earlier; converted to AI generation in December 2022 | After the relaunch, Stable Diffusion produces her images, ChatGPT writes her captions, Movio handles lip sync, and Replica Studio generates her voice. |
| Cosmopolitan AI cover | Karen X. Cheng with Cosmopolitan and OpenAI | June 2022 | A magazine cover rather than a recurring influencer, but widely cited as the first AI cover for a mass-market publication. |
| Vogue Singapore avatars | Vogue Singapore with Varun Gupta | March 2023 | Three AI avatars on Vogue Singapore's Roots issue. |
Virtual humans introduce a set of unsettled questions for the industry. Critics argue that hyper-polished AI bodies reinforce narrow beauty norms and that AI influencers displace real creators, especially women whose income depends on visibility. Trade unions have pushed for stronger disclosure when sponsored content is computer-generated.
Brand safety and ad verification became a distinct industry segment after high-profile YouTube boycotts in 2017, when major advertisers found their ads appearing alongside extremist content. Most of the work in this segment is classification: deciding what a piece of inventory contains and whether it is suitable for a given brand.
The systems share a common pattern. A classifier first scores every piece of inventory across the relevant brand-safety taxonomy, such as the Global Alliance for Responsible Media (GARM) categories: adult content, arms, crime, death, debated social issues, hate speech, misinformation, obscenity, terrorism, illegal drugs, and so on. Suitability is then layered on top. A children's brand may avoid all GARM categories, while a financial services brand may tolerate some of them. The classifier output drives bid filtering at the DSP layer.
The rise of made-for-advertising (MFA) sites and AI-generated content has put new pressure on classifiers. IAS and DV both report that they specifically detect generative AI content and "slop sites" designed to harvest programmatic spend. Brian O'Kelley, the AppNexus co-founder, launched Scope3 in 2021. Scope3 began as a carbon-emissions company for advertising but has expanded into AI-driven brand suitability in direct competition with DV and IAS.
The death of third-party cookies has pushed marketers back toward older statistical techniques. Marketing mix modeling (MMM) uses econometric regression to attribute sales to marketing channels using aggregate data.
Google's wider response to deprecating third-party cookies was the Privacy Sandbox. The Topics API used on-device machine learning in Chrome to infer broad interest categories, such as sports or travel, without exposing detailed browsing data. In April 2025 Google announced it would keep third-party cookies in Chrome rather than removing them, and in 2025 it deprecated the Topics, Protected Audience (PAAPI), and several other Privacy Sandbox APIs. Meta runs its own privacy stack, including Aggregated Event Measurement, that relies on machine learning to model conversions when individual user data is unavailable.
Digital out-of-home (DOOH) advertising covers billboards, transit screens, retail panels, and other physical displays. Programmatic DOOH lets buyers purchase those screens in real time, much like a banner ad. The largest specialist in this category is Vistar Media. Vistar runs a DSP, SSP, ad server, player, device-management platform, and traditional OOH planning software, and lets buyers connect to DOOH screens worldwide. Its Dynamic Creative product changes images and messaging based on contextual data such as time of day, weather, or local traffic.
Generative AI is creeping into DOOH along two paths. The first is creative production: brands generate dozens of region-specific or weather-specific variants from a single master image. The second is interactive ad units, such as Sponsored AI Lenses on Snap and AI Sponsored Snaps that let users converse with a brand's agent inside a placement. Snap launched Sponsored AI Lenses powered by its own generative AI tech and previewed AI Sponsored Snaps that integrate conversational agents into ad units in 2025-2026.
Microsoft is pushing a related concept it calls the "agentic web," with ad surfaces inside Copilot. Microsoft's Offer Highlights format surfaces selling points such as free shipping inside an AI conversation. Best Buy was an early Offer Highlights partner.
The widening use of generative AI in political and commercial advertising has triggered an active period of rule-making.
In the commercial space, the Federal Trade Commission has signaled that AI-generated endorsements or fake reviews violate existing law on deceptive practices. Several agency holding companies, including WPP and Publicis, have published internal AI usage policies covering disclosure, training data, and IP rights.
In June 2024 Toys "R" Us Studios premiered "The Origin of Toys 'R' Us," a short brand film, at the Cannes Lions International Festival of Creativity. The film tells the story of company founder Charles Lazarus and the giraffe mascot Geoffrey. The agency Native Foreign, whose chief creative had alpha access to Sora, produced the spot in a few weeks. By Native Foreign's account roughly 80% to 85% of the imagery was AI-generated, with traditional VFX layered on top.
The ad was the first major brand film made primarily with Sora, and the trade press treated it as a milestone. Reaction in the wider public was harsher. Critics on Fast Company, Reddit, and Twitter said the film had inconsistent character continuity, an uncanny lighting feel, and an unsettling tone for what was meant to be a nostalgic origin story. The ad also reignited a long-running concern about job displacement in commercial production. Independent commercial directors, VFX artists, and union groups argued that high-profile uses of Sora would put pressure on traditional crews.
Toys "R" Us's public response framed the ad as a creative experiment rather than a finished broadcast spot. The film never aired as a network commercial in the United States.
Coca-Cola's 2024 holiday campaign was a remake of the company's 1995 "Holidays Are Coming" spot, which features red trucks driving through snowy towns delivering bottles. The new version was built by three AI studios: Secret Level in Los Angeles, Silverside AI in San Francisco, and The Wild Card in Kuala Lumpur. The studios used four different generative video models, rendered 10,000 frames, produced 5,000 video segments, and worked with over 40 creatives across four continents. Coca-Cola said that what would have been a 12-month project shrank to two months.
The ad aired in November 2024. The reaction was sharp and largely negative. Reddit and X users called the spot "soulless" and "devoid of any actual creativity." Animator Alex Hirsch, creator of the Disney series Gravity Falls, posted a parody criticism. Marketing critics pointed to visual artefacts: trucks with the wrong number of wheels, characters whose faces shift between frames, and unnaturally smooth surfaces.
Coca-Cola defended the campaign. The company said in a statement that it "remains dedicated to creating the highest level of work at the intersection of human creativity and technology" and pointed to the project as part of its broader push to integrate AI into its creative toolkit. Internal metrics presented to trade press argued that the spot generated more impressions than expected. System1 Group, a creative effectiveness firm, reported in its own analysis that the ad still drove emotional positivity on tracking measures.
Coca-Cola repeated the experiment with a refreshed AI-generated "Holidays Are Coming" spot in late 2025, which faced a similar backlash.
Both controversies have become reference points for the wider debate about what generative AI does to ad craft. Defenders argue that the cost and speed advantages are too large to ignore. Critics argue that the medium is not yet good enough to carry the emotional weight of category-defining brand work and that audiences detect the difference.