Industries / Finance and business services / Advertising and marketing
Search and social ads
Search and social ads are short ads that platforms sell by automated auction, one placement at a time, next to search results, in social media feeds and now under the answers of AI chatbots.
7
Milestones achieved, from 1998 to 2026.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1998
Search ads sold per click
GoTo.com, a search engine built at the Idealab startup incubator, let advertisers bid on keywords and charged them only when someone clicked their ad, a model now called pay-per-click. Ads were listed in order of bid, and each click was billed at the advertiser's own bid. Economists Benjamin Edelman, Michael Ostrovsky and Michael Schwarz wrote in 2007 that its ease of use and low entry costs "quickly led to the success" of the platform, but that bids could be changed so often that prices were unstable. Their paper dates the model to 1997; a history of the company says Bill Gross unveiled it at the TED conference in February 1998.
LessonUnder this first-price rule, the advertiser who could react to rivals' bids fastest gained "a substantial advantage", which led to bidding robots and volatile prices.
Credit
Bill Gross and GoTo.com (later Overture, which became part of Yahoo)
2002
Ads ranked by clicks too
Google had sold keyword text ads since October 2000, priced at $10 to $15 per thousand times shown. In February 2002 its AdWords Select charged per click and ranked ads by "a combination of ad performance (click-thru rate) and how much an advertiser agrees to pay per click", so the highest bid alone no longer took the top place. Google also charged each advertiser only enough to stay above the next one, a rule Edelman, Ostrovsky and Schwarz named the generalized second-price auction; they showed that under it, bidding one's true value is generally not the best strategy. Overture later adopted the same auction.
LessonThe paper says the first-price auction caused "volatile prices and allocative inefficiencies" and that Google "addressed these problems" with the AdWords Select rules.
googlepress.blogspot.comgooglepress.blogspot.combenedelman.org
Credit
Google
2007
Purchases shared without consent
Facebook's Beacon program, launched in November 2007, posted what members did on partner websites, such as renting a film from Blockbuster.com, to their friends. The US Court of Appeals for the Ninth Circuit later wrote that Beacon "never required members' affirmative consent". After complaints, Facebook added an opt-out and then shut the program. To settle a class action it agreed to end Beacon permanently and pay $9.5 million, about $6.5 million of it to set up a privacy foundation, and the appeals court upheld the settlement in 2012.
LessonUsing what people do on other websites in ads without asking first cost Facebook the program and a settlement.
Credit
Facebook (a failure)
2021
Phones ask before tracking
With iOS 14.5, released on 26 April 2021, Apple required apps to get a user's permission before tracking them "across apps or websites owned by other companies for advertising", a rule it calls App Tracking Transparency. On Meta's February 2022 earnings call, chief operating officer Sheryl Sandberg named two problems it caused, "targeting and measuring performance", and finance chief David Wehner put the cost to Meta's 2022 revenue "on the order of $10 billion", a figure that is Meta's own estimate.
LessonThe company that makes the phone can change what every advertiser on it is able to track.
Credit
Apple
2021
One campaign for every channel
On 2 November 2021 Google opened Performance Max to all advertisers. The advertiser supplies a goal, a budget, creative pieces and hints about the audience, and Google's software sets the bids and chooses where the ads run across YouTube, Display, Search, Discover, Gmail and Maps. Google moved its Smart Shopping and Local campaigns onto Performance Max in 2022. By Google's own figures, advertisers using it saw "an average increase of 13% total incremental conversions at a similar cost per action".
Credit
Google
2025
Scam ads as revenue
Reuters reported in November 2025, from internal documents, that Meta had projected about 10% of its 2024 revenue, about $16 billion, would come from ads for scams and banned goods, and that its users were shown an estimated 15 billion "higher risk" scam ads a day. A February 2025 document said the team vetting suspect advertisers could not take actions costing Meta more than 0.15% of its revenue. Meta's spokesman said the 10.1% estimate was "rough and overly-inclusive" and that the 0.15% figure was not a hard limit.
LessonThe documents show that removing scam ads was weighed against the revenue it would cost.
Credit
Reuters
2026
Ads under chatbot answers
OpenAI began testing ads in ChatGPT in the US on 9 February 2026, for logged-in adults on its Free and Go plans, and by 21 October 2026 offered them in 89 countries; it says more than 100,000 advertisers have run campaigns. An independent audit by Emma Lurie, Ro Encarnación, Sorelle A. Friedler and Danaé Metaxa used 91 test accounts and collected 3,602 ads from 191 advertisers between February and May 2026. It found the ads labeled "Sponsored" and kept apart from the answer text, and found that lower-income accounts were more likely to be shown ads.
Credit
OpenAI; audit by researchers at the University of Pennsylvania and Haverford College
Watch for
Ads do not change answers
OpenAI says "Ads do not influence the answers ChatGPT gives you" and that ads are "clearly labeled and separate from ChatGPT's answers".
Achieved whenan independent controlled test that runs the same prompts with and without ads finds no difference in which brands the answers name.
Claimed by
OpenAI, February 2026, repeated in October 2026
Fewer scam ad reports
Meta says that "over the past 18 months, we have reduced user reports of scam ads globally by 58 percent" and that it removed more than 134 million pieces of scam ad content in 2025 up to November.
Achieved whenan outside measure of how many ads are scams, rather than how many users report them, shows a similar fall.
Claimed by
Meta spokesman Andy Stone, November 2025
Open
01
Test whether ads sway chatbots
A controlled test of whether ads change what ChatGPT recommends. The same shopping prompts are run many times on two groups of Free accounts: one that sees ads and one that has opted out of ads, an option OpenAI offered at launch in exchange for fewer free messages a day. For each advertiser that appears under an answer, the test counts how often the answer text itself names that advertiser in each group.
The briefCredit · unclaimed02
Sample ads to count scams
A monthly estimate of what share of the ads shown in one EU country are scams, made from a random sample of each large platform's public ad repository rather than from keyword searches, with the labeling method and its error rate published, together with how many days scam ads ran before removal.
The briefCredit · unclaimed
Open 01 · the brief
Test whether ads sway chatbots
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A controlled test of whether ads change what ChatGPT recommends. The same shopping prompts are run many times on two groups of Free accounts: one that sees ads and one that has opted out of ads, an option OpenAI offered at launch in exchange for fewer free messages a day. For each advertiser that appears under an answer, the test counts how often the answer text itself names that advertiser in each group.
Why it is still open
An October 2026 review of ChatGPT ad trackers found that three trackers saw the advertiser cited in the answer above its own ad only about 3% of the time, and noted that "nobody has run a controlled test of the same prompts with and without an ad". The Lurie et al. audit measured which ads were shown, not whether the answers changed. A further search in November 2026 found only observational counts of ads, not a with-and-without test.
Built with what exists
Free ChatGPT accounts, with OpenAI's ad opt-out (or Plus accounts, which carry no ads, as a second comparison group); the public prompt list and ad archive from the Lurie et al. audit; volunteers who paste prompts by hand, as SparkToro's 600 volunteers did in 2026, or a browser automation tool such as Playwright or Browser Use (see the Note "Deleting 80% to start over") if OpenAI's terms allow it; a small, low-cost AI model to pull brand names out of answers (see the Note "The jobs agents repeat"); Python or R for the statistics.
First step
Read OpenAI's terms on automated use and check that the Free-tier ad opt-out still exists, then open the accounts this month, since the audit found that ads began about 14 days after an account was created.
Done looks like
A test registered in advance on at least 50 prompts that drew ads in the audit, with at least 60 runs per prompt in each group, published with the share of answers that name each advertiser in each group, the difference between groups and its 95% confidence interval, and all prompts, answers and code.
Open 02 · the brief
Sample ads to count scams
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A monthly estimate of what share of the ads shown in one EU country are scams, made from a random sample of each large platform's public ad repository rather than from keyword searches, with the labeling method and its error rate published, together with how many days scam ads ran before removal.
Why it is still open
Article 39 of the EU Digital Services Act requires very large platforms to keep a public repository of their ads, with the payer and the dates shown, searchable through "multicriteria queries" and through application programming interfaces (APIs). Published counts start from keyword searches: a Sciences Po data sprint, building on work by AI Forensics, found more than 46,000 fraudulent health ads on Meta that were shown to EU users more than 292 million times. Reuters reported from internal documents that Meta ran the keywords regulators used, to clear ads flagged as fraudulent from the library's search results; Meta said the removals came from genuine enforcement. In Poland, the ScamWatch project said in August 2026 that it had analyzed 1,760,358 ads and classed 16.5% as suspicious, but the published account does not say how the ads were collected or classified, and Meta called the report unreliable. No random-sample estimate with a published method was found.
Built with what exists
the public ad repositories of Meta, Google and TikTok; a vision-language model (one that reads images as well as text) to make a first pass at labeling each ad, for example one that returns probabilities for fixed-answer questions (see the Note "A thousand questions, one pass"); two trained human raters for a check sample; Python.
First step
Check whether each platform's repository API can list all the ads shown in one country on one day without a search word, and publish which ones can and which cannot.
Done looks like
Three consecutive months of estimates for one country, each from at least 1,000 randomly drawn ads per platform, giving the scam share with its 95% confidence interval, agreement between the model and the human raters of kappa 0.8 or better on a 300-ad check set (kappa measures agreement beyond chance, where 1 is perfect), and the median number of days a scam ad ran before removal.