Industries / Finance and business services / Advertising and marketing
Making the ads
Making the ads is the work of writing, drawing, filming and now generating the ads themselves, and of keeping what they say within the rules.
7
Milestones achieved, from c. 1477 to 2026.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
c. 1477
A printed notice for sale
William Caxton printed a small notice in Westminster offering copies of his Ordinale ad usum Sarum, a priest's manual, and stressing that the text was correct and the book cheap. It closes with a Latin request not to remove it, which shows it was meant to be pinned to a wall or door. The Bodleian Library, which holds one of the two known copies, calls it "the earliest surviving printed advertisement in English publishing history." The other copy is at the John Rylands Library in Manchester.
Credit
William Caxton
1888
Copywriting becomes a job
The Philadelphia agency N. W. Ayer & Son hired Jarvis Wood, in the Smithsonian's words "the industry's first full-time copywriter." Ayer hired its first commercial artist in 1898, set up a copy department in 1900, and twelve years after the artist arrived became the first agency to offer clients a full-time art director, by the Smithsonian's account.
LessonOnce agencies were paid directly by advertisers (see Buying ad space), writing and drawing the ad became part of the service they sold.
Credit
N. W. Ayer & Son
1923
Ads tested by coupon
Claude Hopkins, a copywriter and later president of the Chicago agency Lord & Thomas, published Scientific Advertising. The book describes "keyed advertising": each ad carries a coupon or code, so the replies it brings can be counted, and headlines, sizes and pictures can be compared by "cost per customer or cost per dollar of sale." The Advertising Hall of Fame calls Hopkins "the first copywriter to put coupons in ads offering a free sample of the product."
LessonAn ad whose replies can be counted can be tested against another ad instead of judged by taste.
Credit
Claude C. Hopkins; Lord & Thomas
1957
A hidden-message hoax
Market researcher James Vicary said that flashing "Drink Coca-Cola" and "Hungry? Eat Popcorn" on a New Jersey cinema screen for 1/3,000 of a second, during the film Picnic, had raised Coca-Cola sales by 18.1% and popcorn sales by 57.8%. Such messages, too brief to be noticed, are called subliminal. Live Science reports that when the president of the testing firm Psychological Corp. challenged him to repeat the test, he could not, and he admitted he had made up the results.
LessonA claim that an ad works counts only when someone else can repeat the test.
Credit
James Vicary
1962
An independent ad referee
In the UK, TV ads had been controlled by law since commercial TV began in 1955. For print, posters and direct mail, the trade's Committee of Advertising Practice (CAP) set up the Advertising Standards Authority (ASA) in 1962 to apply a new code of practice. In 1988 the ASA gained the power to refer persistent offenders to the Office of Fair Trading for legal action, in 2004 it took over TV and radio ads, and in 2010 its remit grew to cover companies' claims on their own websites and social media pages. The ASA calls its published rulings "a transparent record of our policy on what is and isn't acceptable in ads."
Credit
the Advertising Association and CAP
2024
An AI ad draws protest
In November 2024 Coca-Cola released holiday films made with generative AI, software that produces images and video from text instructions, as a tribute to its 1995 commercial "Holidays Are Coming." Forbes, as relayed by NBC News, reported that three AI studios (Secret Level, Silverside AI and Wild Card) made them using four AI models. Viewers online called the result "soulless," and Coca-Cola said it had "crafted films through a collaboration of human storytellers and the power of generative AI."
LessonViewers may judge an ad by how it was made as well as by what it shows.
Credit
The Coca-Cola Company; Secret Level, Silverside AI and Wild Card
2026
Most video buyers use AI
In a report released in July 2026, the Interactive Advertising Bureau (IAB), the online ad trade body, said "nearly two-in-three buyers now use GenAI for digital video creative, up from half in 2025." Buyers here are the advertisers and agencies who pay for video ads. By the IAB's figures, about a third of their ad assets will use generative AI in 2026, up from a quarter in 2025, and 43% by 2027.
Credit
IAB, with the research firms Advertiser Perceptions and Guideline
Watch for
Fully automated ad campaigns
The Wall Street Journal reported in 2025 that Meta aimed to let a business create and target a whole ad campaign with AI by the end of 2026. Marketing Brew reported in April 2026 that Mark Zuckerberg has described advertising as one day being "as simple as" entering a credit card number and a business goal, and a Meta spokesperson said "automation delivers stronger results for most advertisers." Agency staff interviewed for the same article said clients still want control, and one said Meta's creative AI tools consistently did worse than his own third-party tools.
Achieved whenMeta offers fully automated campaigns to all advertisers and an outside party compares their results with human-made campaigns on the same budgets.
Claimed by
Meta, as reported by The Wall Street Journal (2025) and Marketing Brew (April 2026)
AI screening by a regulator
The UK's ASA says its Active Ad Monitoring system scanned nearly 60 million online ads in 2025, up from 28 million in 2024. Machine-learning models and, increasingly, large language models flag ads likely to break the rules, and ASA staff decide what action to take. The ASA counts 22,383 ads amended or removed in 2025 as a result of all its work, not only the AI system. It has not published how many flagged ads turned out to break the rules, or how many breaches the system missed.
Achieved whenthe ASA or an outside reviewer publishes the share of flagged ads confirmed as breaches and an estimate of the breaches the system misses.
Claimed by
the ASA, in its briefing note on the system and its annual report for 2025 (2026)
Open
01
Match rulings to real ads
A public dataset that links each ASA ruling to a record of the ad as it actually ran, with the ruling's outcome (upheld or not upheld), the code rules it cites and the advertiser's sector. Tools that check ads before they run could then be tested on real ads, not on descriptions of them.
The briefCredit · unclaimed02
Audio description for ads
A tool that writes and voices audio description for a 30-second ad (a spoken account of what is on screen, fitted into the gaps between dialogue), and a test of it with blind and partially sighted viewers.
The briefCredit · unclaimed
Open 01 · the brief
Match rulings to real ads
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public dataset that links each ASA ruling to a record of the ad as it actually ran, with the ruling's outcome (upheld or not upheld), the code rules it cites and the advertiser's sector. Tools that check ads before they run could then be tested on real ads, not on descriptions of them.
Why it is still open
An ASA ruling describes the ad in words. The ruling opened for this file (B Sporting Ltd) shows no image of the ad, and PreFlight, a screening tool for health and supplement ads, says the ruling pages "carry no image or video." So the small projects testing AI checkers on ASA rulings in October and November 2026 rebuild each ad from the ruling's text. PreFlight tested on 79 rebuilt and made-up ads. A pilot on 30 financial-ad rulings had only 2 not upheld, and notes that the ad descriptions are the ASA's own summaries, written after it reached its decision. A GitHub and web search found no dataset that pairs rulings with the original ads.
Built with what exists
the ASA's published rulings; public ad libraries such as the Google Ads Transparency Center and the Meta Ad Library, and archived copies of web pages; a browser agent such as Browser Use (see the Note "Deleting 80% to start over") to search them slowly and within each site's rules; a low-cost extraction model such as Muse Ember (see the Note "The jobs agents repeat") to pull the advertiser, date, outcome and cited rules from each ruling.
First step
Download one year of rulings, sort them by sector and outcome, and try to find the original ad for the first 50.
Done looks like
500 rulings from at least 10 sectors, including every not-upheld ruling from the same years, each linked to the ad as it ran, with two people checking each match and their rate of agreement published. Links and descriptions are published in place of copies wherever the ad's owner has not allowed copying.
Open 02 · the brief
Audio description for ads
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A tool that writes and voices audio description for a 30-second ad (a spoken account of what is on screen, fitted into the gaps between dialogue), and a test of it with blind and partially sighted viewers.
Why it is still open
A University of Surrey study for the RNIB (Royal National Institute of Blind People), posted in September 2025, compared three AI models with professional describers on more than 160 clips from three BBC factual series. It found the AI text often fluent but lacking "accuracy, cohesion, and contextual awareness," and prone to over-description that did not fit fixed broadcast slots. Participants welcomed AI description if it expanded coverage but said human review remains essential. The study covered programs, not ads, and searches found no study of AI audio description for ads.
Built with what exists
vision-language models that read video frames, such as Gemini or Claude; a speech-to-text model to time the dialogue and find the gaps, such as ElevenLabs Scribe v3 (see the Note "Fewer word errors"); a text-to-speech voice; ffmpeg to mix the track.
First step
Transcribe five ads, mark the silent gaps by hand, and draft AI descriptions to fit them.
Done looks like
For 30 ads, a panel of at least 12 blind and partially sighted viewers rates the AI description against one written by a professional describer for the same ad, and the project publishes the ratings, the count of factual errors per ad and how often a description ran past its gap.