BRENNEN

Industries / Finance and business services / Advertising and marketing

Buying ad space

Buying ad space is choosing where ads appear and paying for the space or time, whether a newspaper page, a TV slot, a poster site or a single showing of an ad on a web page.

6

Milestones achieved, from 1876 to 2020.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1876

    Agents paid by advertisers

    N. W. Ayer & Son introduced the "open contract" with Diggee & Conard, growers and agricultural suppliers. Before it, the Smithsonian's finding aid says, most agencies worked as "space-jobbers," independent wholesalers of advertising space. Under the open contract the advertiser paid the agency a fixed commission on the space bought and could see the rates the newspapers actually charged. By 1884 nearly three-quarters of Ayer's billings were on open contract. The Advertising Hall of Fame dates the plan to 1875 and says the commission eventually settled at 15 percent.

    LessonIn the Smithsonian's account, a known commission put the agent "firmly on the side of the advertiser.".

    sirismm.si.eduadvertisinghalloffame.org

    Credit

    Francis Wayland Ayer, N. W. Ayer & Son

  2. 1994

    Web space sold by banner

    On 27 October 1994 HotWired, the online offshoot of Wired magazine, launched with banner ads from 14 companies, including MCI, Volvo, Club Med and Zima. Writing about itself, Wired says "legend has it" the first was AT&T's "Have you ever clicked your mouse right here? You will." Wired notes that for the first time a marketer could know how many people saw an ad and how many interacted with it.

    wired.com

    Credit

    HotWired; the ad agency Messner Vetere Berger McNamee Schmetterer, which made the ads for its clients

  3. 2012

    Common rules for ad auctions

    In real-time bidding, each single showing of an ad, called an impression, is put up for auction while a web page loads; the 3ve fraud report describes these auctions running "billions of times a day, in the milliseconds before a page loads." OpenRTB is the shared technical standard for those auctions. Technology firms from the buying and selling sides formed the project in November 2010, released version 2.0 in October 2012, and handed it to IAB Tech Lab, the online ad trade body's standards group, which still maintains it.

    iabtechlab.comservices.google.com

    Credit

    the OpenRTB Consortium; IAB Tech Lab

  4. 2017

    Safer placement by default

    After "a number of cases" in which brands' ads appeared beside content that did not match their values, Google's chief business officer, Philipp Schindler, wrote: "We know that this is unacceptable." Google changed the default so that ads ran only beside content meeting a higher brand-safety standard (brand safety is keeping ads away from content that could harm the advertiser's name), gave advertisers account-wide tools to exclude sites and channels, and removed ads from content that attacks people for their race, religion or gender.

    LessonBuying an audience across millions of videos also buys whatever the ad appears beside.

    blog.google

    Credit

    Google

  5. 2018

    A fraud ring shut down

    Google and the fraud-detection firm White Ops reported on 3ve, a scheme that at its peak controlled more than 1 million internet addresses, had up to 700,000 infected desktop computers active at any time, produced more than 10,000 counterfeit website domains and generated more than 3 billion bid requests a day. The firms referred their findings to law enforcement, and the US Department of Justice announced criminal charges. Their report urged buyers to use ads.txt, which IAB Tech Lab had released in June 2017: a public file in which a website lists the companies allowed to sell its ad space, making counterfeit space harder to sell. The report counted more than 500,000 domains that had published one.

    LessonAutomated auctions let fraudsters sell fake space under real sites' names until buyers can check who is allowed to sell it.

    services.google.comiabtechlab.com

    Credit

    Google and White Ops, with an industry working group; IAB Tech Lab for ads.txt

  6. 2020

    Where the money goes

    ISBA, the trade body for UK advertisers, had the accounting firm PwC trace programmatic ad spending (ads bought by automated systems) from 15 advertisers through 12 agencies, 11 ad technology firms and 12 publishers between January and March 2020. Of 267 million impressions, only 31 million (12%) could be matched from end to end. On those, publishers received about half of what advertisers spent, and 15%, which the study called the "unknown delta," could not be traced to any company.

    LessonWhen the records of each company in the chain cannot be matched, nobody can say where a large share of the money went.

    isba.org.uk

    Credit

    ISBA and PwC, with the Association of Online Publishers (AOP)

Watch for

  1. AI agents buy ad space

    AdCP (Ad Context Protocol), an open set of rules for AI agents that plan, buy and sell media, says "the first-ever agent-to-agent media buy was executed" on 16 October 2025 with LG Ads inventory, and that more than 163 companies have joined. In November 2025 IAB Tech Lab put its Agentic RTB Framework v1.0 out for public comment; MediaPost reports the group says it cuts the round trip of a bid request, now about 600 to 800 milliseconds, by as much as 80%.

    Achieved whenan outside auditor confirms that agent-placed campaigns delivered and billed as booked at routine scale, and someone outside IAB Tech Lab measures the speed gain.

    adcontextprotocol.orgmediapost.com

    Claimed by

    the AdCP project, October 2025; IAB Tech Lab, November 2025

  2. Court-ordered auction changes

    In April 2025 Judge Leonie Brinkema of the US District Court for the Eastern District of Virginia found that Google had illegally monopolized publisher ad servers and ad exchanges for open-web display ads. Her remedies opinion, unsealed on 16 September 2026, found a forced sale of Google's AdX exchange "neither realistic nor needed." It orders six years of conduct rules instead: AdX must bid through rival auction software such as Prebid, publishers can export their data, Google must explain auction decisions ad by ad, and a monitor will check compliance. Appeals are expected.

    Achieved whenthe rules take effect, the monitor reports that AdX bids reach rival publishers' auctions as ordered, and any appeal is decided.

    natlawreview.com

    Claimed by

    the court's remedies opinion, September 2026

Open

  1. 01

    Open test for MFA sites

    A public, labeled set of websites and an open classifier that judges whether each is a made-for-advertising (MFA) site, one built mainly to show ads to visitors bought cheaply elsewhere, with its accuracy published.

    The brief
    Credit · unclaimed
  2. 02

    Trick tests for buying agents

    A simulated ad marketplace where AI buying agents get a brief (budget, audience, sites to avoid) and deal with seller agents, some of which lie: they pass off MFA sites as premium news, inflate audience counts, or hide instructions to the buyer in their product descriptions. The test records how much of each budget ends up with the dishonest sellers and how often the brief's limits are broken.

    The brief
    Credit · unclaimed

Open 01 · the brief

Open test for MFA sites

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public, labeled set of websites and an open classifier that judges whether each is a made-for-advertising (MFA) site, one built mainly to show ads to visitors bought cheaply elsewhere, with its accuracy published.

Why it is still open

In a 2023 study of 35.5 billion impressions bought by 21 member companies, the Association of National Advertisers found MFA sites took 21% of impressions and 15% of spending. The lists buyers use to avoid them are sold by verification firms. Pixalate's, for example, is a daily client feed that rates sites on ad density, ad refresh rate and the share of traffic bought from ads and social media, and its documentation gives no accuracy figure. Web and GitHub searches found no public list scored against human judges; the nearest, a small tool posted in January 2026, labels URLs with a language model and reports no accuracy.

ana.netpixalate.comgithub.com

Built with what exists

Playwright or Browser Use (see the Note "Deleting 80% to start over") to load each page and record its ads, refreshes and page weight; a vision-language model to measure how much of the screen the ads take; a model that returns probabilities for yes-or-no checklists, such as Gemini Decision 1 (see the Note "A thousand questions, one pass"); Python. It uses only what a visitor's browser can see, since traffic sources are not public.

  1. First step

    Write a one-page labeling guide from the ANA's description and label 100 sites by hand.

  2. Done looks like

    1,000 sites labeled by two trained people against the ANA's description, with their agreement published, and a classifier that reaches at least 90% precision and 80% recall on 300 sites held back from training, released with its code and labels.

Open 02 · the brief

Trick tests for buying agents

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A simulated ad marketplace where AI buying agents get a brief (budget, audience, sites to avoid) and deal with seller agents, some of which lie: they pass off MFA sites as premium news, inflate audience counts, or hide instructions to the buyer in their product descriptions. The test records how much of each budget ends up with the dishonest sellers and how often the brief's limits are broken.

Why it is still open

Microsoft's open-source Magentic Marketplace (October 2025) simulates buyer and seller agents and can test resistance to manipulation, but its examples are restaurants and contractors. AdCP's compliance tests for buyer agents cover finding products from a brief, negotiating terms and watching delivery; none of the test files is named for a lying seller or injected instructions. Searches found no such test for ad buying.

github.comarxiv.orggithub.com

Built with what exists

Magentic Marketplace (MIT license); AdCP's schemas and its Python and TypeScript libraries (Apache 2.0 license); AI models from several makers as buyers, including low-cost ones such as Muse Ember (see the Note "The jobs agents repeat"). No real money or real ad accounts are involved.

  1. First step

    Copy Magentic Marketplace, replace its restaurants with 50 made-up ad products written in AdCP's product format, and script five kinds of lying seller.

  2. Done looks like

    At least five AI models, each run on 100 briefs in a market where one seller in five lies, with the share of budget lost to lying sellers and the share of briefs broken published per model, and the test released as open code.