Media Briefing: Ads for AI agents have a measurement problem

As publishers and advertisers determine how to put ads in front of a new kind of audience — AI agents, not humans — they’re running into a familiar problem: how do you know whether an ad actually worked?

Answering that question is driving a wave of startups building ways to serve ads to agents as they move through publishers’ pages, then measure what happens after. It’s early, granted. So those pitches from startups claiming to have cracked the code (and judging by inbox, there’s a few), probably haven’t. For now measurement is opaque, and nothing resembling a standard has caught on. Most of what’s out there is still proof of concept. 

OpenAds, for example, is going down the relatively old school route of unique referral codes, but tweaked for AI. Let’s say a person asks ChatGPT a question about the best running shoes for a half marathon. When ChatGPT scrapes a webpage to answer that prompt, OpenAds would generate and insert an ad for a brand’s running shoes with a unique 10-percent-off referral code. If ChatGPT thinks the code is relevant enough to the user looking for those shoes, it’ll “survive the retrieval” process, said Steven Liss, co-founder of AI native advertising platform OpenAds. Then when that link is surfaced in an AI-generated response, and the user clicks on the link to that product, brands will be able to track that back to ChatGPT, he added.

Oasy, a startup advertising and analytics platform to monetize AI bot traffic for web publishers, meanwhile, is currently working on landing partnerships with larger native ad platforms that already work with publishers to “start tracking something closer to a cost-per-click or cost-per-referral,” Choy Travers, co-founder of Oasy.

“Publishers, advertisers, affiliates, brands — everyone is circling the same problem and opportunity, which is within AI visibility there is no exact science. There’s no exact attribution, unlike Google Ads and Facebook Ads, where you have that click,” said Travers.

What that looks like remains to be seen, but the gist is it would be a similar mechanism to dynamic ad insertion: injecting a text-based sponsored ad into the HTML of a publisher webpage when it gets scraped by an AI crawler (detected through an integration with a publisher’s CDN like Cloudflare), Travers said. And while he said these agentic ads are currently live with publishers and advertisers, he declined to name which ones. The goal, he said, is to “influence the chance” that brands appear in the AI output to the user.

In the case of Time’s recently-served ads to AI agents, FAQ-style sponsored ads are deployed in markdown pages. Mobian then inputs those same questions into AI search engines to measure the brand’s AI visibility, favorability and accuracy scores over time. Time sells one agent ad per markdown page and measures AI bot traffic to that page. However, that’s sort of backing into the measurement, by prompting an AI chatbot to scrape a specific webpage and look for a specific brand ad, rather than tracking if — and when — that ad was surfaced to a user by the chatbot.

For what it’s worth, the IAB is currently drafting a framework around attribution for agentic advertising. The IAB released standards around measuring AI visibility and disclosing AI usage in content production this month.

The ad funnel for AI agents

In many ways, this measurement problem sits exactly where it should. Traditional digital advertising measurement follows a pretty straightforward path. A human sees an ad, clicks on it and maybe buys the product. But when an AI agent is the target audience, it has to retrieve the ad, incorporate that information into an AI-generated response to serve to the user, and ultimately influence what that person buys. All without the consumer ever seeing or clicking the original ad.

“How do you attribute that the ad was actually seen by an agent and then influenced the response?” said Brendan Norman, founder and CEO at Classify, an AI-powered contextual advertising and ad tech company. “There’s a lot of things to consider with this because it’s not as straightforward as a clickthrough or a checkout.”

A publisher can use providers like Cloudflare to track when an AI agent crawled a webpage that contained an ad — but that doesn’t mean the agent actually processed the sponsored information, or even surfaced it to a user. And even if the agent does recommend the brand or the product, it remains extremely difficult to connect that recommendation back to the original ad. Attribution gets even murkier, because an AI agent may draw on dozens of sources while researching a product or answering a question.

“That’s where everyone is stuck at the moment,” Liss said.

One way to get around this is to use GEO analytics firms as quasi-independent measurement companies, Liss said. These firms can sample synthetic prompt data or panels and measure how often a brand is surfacing before and after ads go live on publishers’ pages. But that’s not enough to scale the agentic ad marketplace, Liss said.

It’s why the reaction from buyers to the idea of buying ads that Time has specifically designed to be read by AI agents was met with some skepticism. Before spending anything meaningful, they want to see why it works. 

As Liss explained: “For advertisers to put anything more than experimental budgets, which they do exist for this, they’re going to have to be able to measure some kind of outcome. You can try and invent a new currency like GEO lift, agent retrievals, but the best currency is currency. The way that this takes off fast is bottom-of-the-funnel, attributed [return on ad spend].”

What we’ve heard

“I don’t believe in total Google Zero, for what it’s worth. I don’t think that is happening… because I think [Google knows] they lose. They have an ad business that loses if that happens… The competition to surface is significantly more intense, but it’s not just AI. It’s just the size of the the content corpus.”

– A head of partnerships at a news publisher.

Numbers to know

More than half: The share of Time’s total revenue that the company expects its events business to make up in 2026, up from 28% in 2023.

21%: The average robots.txt AI bot rule coverage across publishers.

26%: The percentage of ChatGPT responses that contain a sponsored ad.

What we’ve covered

The Atlantic is building games by hand

• At a time when companies are using AI to automate work processes and build products faster, The Atlantic is taking an increasingly unusual approach to its latest game: building it by hand, every day.
• The Atlantic launched a new trivia game called Rabbithole this week, the brainchild of Atlantic senior editor Drew Goins, who builds the seven-question game by drawing inspiration from Atlantic articles and researching connections across the publication’s archive, and other sources.

Read more here.

OpenAI’s ads business hits Europe

• OpenAI will begin serving ads to people using ChatGPT across 31 European markets on Monday (August 24), including Germany, France, Spain, Italy, Poland, Sweden, Norway, Denmark, Ireland, and the Netherlands.
• The timing lands at the six month mark for OpenAI’s ads business. Europe was always going to be slower. Regulation there requires explicit consent for personalized targeting unless a company can prove a legitimate interest to do otherwise, which is a hard case to make. 

Read more here.

USA Fencing relies on creators to attract new fans ahead of the 2028 Olympics

• If the World Cup was any indication, creators are a crucial marketing pillar of sports leagues and major sporting events, drawing new fans to sports through their secondary broadcasts or paid partnerships. 
• USA Fencing recognizes that power, and is attempting to make fencing the hottest sport of the summer— two years ahead of the 2028 Los Angeles Olympics. This will be the first American competition where creators are welcomed into the villages, sidelines, and broadcasts. 

Read more here.

European publishers are getting hit harder by AI bot scraping, report finds

• A recent report from TollBit found that European sites get scraped more by AI bots, receive fewer referrals, and have their robots.txt rules – which websites use to communicate to crawlers what they can and cannot scrape – ignored more often.
• This disparity could be because Europe has more languages than North America, and as LLMs try to learn different languages and pull content from different sources, and because of TollBit’s dataset of publishers.

Read more here.

Walmart-backed gaming publication lays off editorial team

• In December 2024, Walmart helped start a video gaming news website called Restart.run
• The entire editorial team at Restart was laid off last week. It’s unclear whether the layoffs are due to underperformance or a lack of investment.

Read more here.

What we’re reading

Forbes fires top editor after finding out he took a $6 million payment

Forbes fired longtime top editor Randall Lane after discovering he had accepted an undisclosed $6 million payment from the founder of Shook Research, a firm that partners with Forbes on its wealth-adviser rankings, The New York Times reported.

Taboola will handle all of NBC News’ programmatic display ads

NBC News struck a new deal with Taboola to handle all of its programmatic display ad sales, consolidating its automated ads business with the vendor, Axios reported.

Apple looks to pay publishers to improve voice assistant Siri

Apple is in talks with news publishers about multiyear deals that would pay them when their content is used to provide information to its revamped AI-powered Siri, potentially creating a new usage-based licensing model for publishers, The Wall Street Journal reported.

Boston Globe editor announces reorganization and a ‘less dense’ website

The Boston Globe is reorganizing its newsroom around a more digital-first structure, giving editors greater control over how stories are packaged and promoted online while redesigning its website to be less dense, Media Nation reported.