The response since we launched Agent Ads with TIME has been overwhelmingly positive.
Brands are reaching out. Publishers are reaching out. Most people we talk to are excited about brand-verifiable information provided in a sourced and cited way, and about a new channel that treats agents as the audience they have become.
Andrea Brimmer, CMO of Ally, told Business Insider she was quick to experiment with advertising to AI agents because "the signals are impossible to ignore." Consumers visiting Ally's website from AI platforms were 3.5 times more likely to open an account than those from traditional search, and Ally's AI-driven traffic is up 9 times year over year. "When the way people discover brands is shifting fundamentally that fast, I'd rather be in the room helping shape what comes next than standing outside it," she said. That is why brands are moving quickly. The agents are already in the buying journey.
What an Agent Ad delivers is a set of brand-verifiable facts, advertiser approved and source-cited, with the sponsorship disclosed in the first line. We could call these brand fact impressions, and maybe we will. Ads makes it clear that this is inventory, which it is. Brand facts makes it clear that what is being provided is different from traditional ads for humans, and that is critically important.
This is early. It is the first. It will evolve. When the first ad ran on the internet there were plenty of naysayers. When sponsored results showed up in search, people reacted similarly. In many cases sponsored results became preferable, because you know the information is provided by the brand. A hotel advertising a specific location with current availability is better than a non-authoritative result. Same for a restaurant confirming it is open or carries a particular item, a movie listing what times it plays, and countless other cases. These are brand facts. Getting them directly from brands is meaningful. It does not mean every question has to end in a brand fact. It means having that information available is incredibly valuable, and as a result incredibly important.
I want to address some of the frequently asked questions and some of the claims we have heard.
"Is this spamming the models?" No, it is literally the opposite. An Agent Ad is one unit on one page, placed where the content is contextually appropriate. The page is what the agent is choosing to read while answering its query or the user's. The first line of every ad is a disclosure that it is sponsored. Every fact carries a source citation. There are no hidden instructions of any kind. Anyone can open the file and read exactly what an agent reads. Spam depends on volume and concealment. This is a single labeled unit of brand facts with sources, fully inspectable. There is nothing deceptive about a disclosed, labeled, cited unit that anyone can open and read.
"Is this cloaking?" No. Cloaking means showing a crawler something different from what a person sees in order to deceive. Agent Pages give agents the publisher's content in a format built for how agents consume it, and we serve slim HTML as well, so it is not markdown or nothing. The ad is labeled as sponsored in its first line. Nothing is concealed from anyone. Manipulation involves concealment. This is the opposite.
"Do agents prefer markdown over HTML?" This is less about markdown versus HTML than people think. Mobian already serves both markdown and slim HTML to agents, and slim HTML is a version of HTML optimized for agents. Framing it as markdown versus HTML makes it about the language, when the experience is the whole point. Mobian is powering an optimized experience for agents on TIME. It is fewer tokens for the agent to consume, more efficient, and we think more effective. Historically bots and agents did not have markdown as an option, so HTML was always the standard, and we continue to support HTML. That may or may not change, but it is beside the point. The real question is whether publishers optimize experiences for the agent audience, and that is what this is.
"I tested this myself with curl or my browser and saw the ad." That is meant to happen. Anyone can request the agent version and read exactly what an agent reads, and that inspectability is for testing and verification purposes. Those requests do not count as impressions. Impressions are counted when the request is verified, which has a number of different mechanisms available such as rDNS confirmation of the agent operator, depending on the CDN and agent operator.
"Won't models block this, and what happens if they do?" A model can choose to use the information or not. That has always been true and it should stay true. A model can say it will block labeled, sourced brand facts. But it would be to their detriment if they do. You can choose to not have information about what times a movie plays, or whether a hotel has availability, or whether a restaurant is open. The models that ignore accurate, current, sourced brand facts will get facts wrong and lack real-time information, and we will see it in their results.
"Are Agent Ads good for the models?" Yes. Brand facts are good for the models. Our latest internal data shows that models get brand facts wrong around 34% of the time, so this is a fundamental challenge that must be solved. Agent Ads are built to solve it. They are clearly disclosed, with the sponsor named in the first line and a citation behind every fact, so a model knows exactly what it is reading and who provided it before it reads a single claim. Most of the web offers no provenance at all, so facts that are sourced and attributed are high quality data. Brand facts are also often more real-time than any other source. Offers, rates, availability, and timing change constantly, and a brand can update its facts and reach agents in real time. A model forming an answer gets current, sourced, brand-verified information at the moment it is consuming the page, and it can weigh that information and use it or not. Our early Causal Ad Impact results show that it works. Models that previously answered a brand question incorrectly answer it correctly with exposure.
"Are Agent Ads good for publishers?" Yes. The models depend on publishers for quality, credible content, and the business models need to evolve. Agent Ads are a step in that direction. The models get more accurate information, brands get the ability to update their facts and reach agents in real time at the moment of consumption, and publishers get a way to monetize their inventory, which has been challenging with some of the model companies.
"Why sponsored at all, rather than just brand facts on a brand's own site?" We named these Agent Ads as the idea is to provide information at a moment it makes sense. In traditional advertising we say right person, right message, right time. This is the right AI, the right message, and the right time. The moment that matters in this case is the moment a model chooses to research, from the sources it chooses to research from. The models already have enormous amounts of information about most brands. Much of it conflicts, because their training data certainly contains conflicts. In a Mobian study of more than 750 brands, roughly 17 percent of the sources AI engines cited conflicted with the brand's own facts and/or positioning. What a model chooses to present is the key, and the moment that decides it is the agent read moment. An agent is consuming a page while forming an answer, and that is exactly when and where the brand's facts appear. Being on the pages agents read is not enough on its own. Being there at the moment of consumption is what matters.
Brand facts are the foundation. The speed of getting them to the models in the right moments is where value multiplies. Brand facts change constantly. Offers, rates, availability, timing. What has been missing is the cycle. Brands haven't had a way to update brand facts in real time and get them in front of models at the moment those facts are relevant.
And if you ask the models, they want to move towards truth. That is important and a good thing. But they need real-time brand facts at the right moment. We built one system that takes a brand's current facts, including current offers, and places them in the pages agents are reading, across models, at the moment of contextual consumption. A brand can change a fact and it reaches agents in real time.
For a human, the atomic unit of an ad is the brand and, very closely related, the story. For an agent, it is credible, trusted facts. This is not influence the way we think about influencing a human with emotion. It is presenting factual, verifiable information from the source. The model can use it or not.
"Is it working?" Yes. Both of our measurement approaches are showing statistically significant results. In Causal Ad Impact testing, models that answer a brand question incorrectly answer it correctly with exposure to the brand facts unit, measured against the same model and the same question. In daily Model Impact tracking, which asks the campaign questions with no page and no prompting, the campaign question with the most room to move has shown a statistically significant visibility increase against the pre-campaign baseline at our 95% confidence standard. It is early and the results will keep evolving, but the early answer is yes.
"Does an Agent Ad replace an existing impression?" No. The agent was already reading the page, just without potentially valuable brand facts. It is the same page, in the format the agent reads, with the brand's cited facts included and clearly labeled as sponsored.
"How is an Agent Ad built?" Agent Ads are verifiable, cited, and sourced from brand facts. Mobian's tooling autonomously drafts each Agent Ad, informed by our measurement of how the brand wants to show up and how it actually shows up across AI today. The brand edits where appropriate and approves. Nothing runs without their sign off. The system is end to end, from brand intelligence through creative generation, serving, and outcomes measurement.
"Does Mobian provide the facts or does the brand?" The facts are all brand-verifiable facts. We create a draft for the brand, which the brand edits and approves. They are the brand's facts, according to the brand.
"Is sponsored content clearly identified?" Yes. Disclosure appears in the first line of every Agent Ad, naming the sponsor and identifying them as the source of the material. Each fact carries a source citation.
"Are Agent Ads contextually targeted?" Yes. Agent Ads are contextually targeted based on brand and publisher preferences including targeting by content, genre, persona, and more.
"What reporting is available?" Granular reporting is available by agent, agent fetch mode, section, page type, individual content, campaign, and more.
"How do you measure outcomes?" Every campaign gets automatic measurement of outcomes. We track how AI models answer the campaign's questions, daily and hundreds of times a day, across models including the largest ones and many others. Early indicators show that providing factually accurate brand information to a model corrects information it previously got wrong. Today, we measure outcomes two ways, Causal Ad Impact and Model Impact.
Causal Ad Impact comes from a controlled exposure experiment. It is the same AI model and the same question, with the brand facts unit present in one cell and removed in the other. Answers are scored against the brand's stated answer and reported in both directions. As an example, take a brand whose customers can do something the models widely say they cannot. Without exposure, the models repeat the error. With exposure, the same models answer correctly. Controlling for the model, that lift is the impact of exposure. It does not claim to change the underlying model. It shows that when a model has the right exposure, it gets the facts right.
Model Impact is daily tracking of the same questions asked on their own, with no brand facts unit or page provided. A baseline of Visibility, Favorability, and Accuracy is established on an agreed set of prompts before launch, and after launch we track movement against that baseline every day. Changes that exceed natural day-to-day variation at 95% confidence are reported as statistically significant, and the full daily data is reported either way. We are seeing statistically significant results on both. For Causal Ad Impact, significance is determined with an exact McNemar test on the paired trials, corrected for multiple comparisons. For Model Impact, movement is tested with Welch's t-test against the pre-launch baseline, with Benjamini-Hochberg correction across a pre-registered family of tests. This is the design we are using today, and we expect it to continue to evolve rapidly with the market and with what advertisers ask us to measure. Related, a 2025 study of how AI agents interact with online advertising found that agents neither ignore nor systematically avoid ads, and that they favor structured data and keywords over visual formats.
"If an agent sees an ad once, is that enough? Do more AI impressions matter?" It appears as though more exposure matters, and models seem to forget facts without repetition. This makes technical sense. On the training side, a model's ability to answer a fact-based question scales with how many documents containing that fact it saw during training (Kandpal et al., ICML 2023). Allen-Zhu and Li (2024) quantified it. A fact needs roughly 1,000 exposures during training to be stored at full capacity, and at around 100 exposures storage capacity drops by half. Chang et al. (NeurIPS 2024) showed why one read is not enough. Models progressively forget facts after seeing them, following a power law, so a fact seen once fades unless it recurs. How much repetition a given fact needs will of course depend on context, including how distinctive the fact is, how much conflicting information exists around it, the model's own weights and training mix, and whether the model is answering from training or retrieving live. What the research supports is the direction. Repetition builds retention, and a fact left unrepeated fades.
"How important is source credibility?" On the retrieval side, where live AI answers pull from the web in real time, credibility signals dominate. Across 10,000 queries, the KDD 2024 GEO study (Aggarwal et al.) found that content with credibility markers such as statistics, source citations, and quotations improved visibility in AI answers by up to 40%, while classical SEO tactics like keyword stuffing had little to no effect. Algaba et al. (2025) found that LLMs systematically favor already highly cited sources when generating references, so existing prominence compounds into future visibility.
"Does context matter?" The research on co-occurrence could imply that relevance strengthens the effect. Co-occurrence means two things appearing together in the same content, such as a brand name and a topic showing up in the same passage. Kang and Choi (EMNLP 2023) found that a model's ability to recall a fact tracks how often the fact's terms appear together in training data, and that models struggle with facts whose terms rarely co-occur. By that logic, a brand's facts appearing alongside relevant subject matter is what builds the association between the brand and that subject. On the retrieval side, answer engines select sources by relevance to the query, so facts on topically relevant pages are more likely to be pulled into answers on that topic. No study yet isolates the same fact on a relevant vs an irrelevant page.
"Is a longer or shorter ad more effective?" There is an argument that longer is better, backed by research, as long as longer means more distinct facts. The research supports density of information, not length as such. On the training side, Allen-Zhu and Li found that stating the same knowledge in diverse wordings does not cost a model's storage capacity and makes the knowledge more extractable, while repeating identical sentences wastes capacity on memorizing sentence structure. The co-occurrence research points the same way. Each additional distinct fact adds another pairing between the brand and a claim or topic. On the retrieval side, the GEO study found that pages dense with statistics, source citations, and quotable material won measurably more citation share in AI answers. We expect ad formats to evolve over time with different use cases.
There will be ebbs and flows along the way. Keep the longer horizon in mind. Brand facts matter to users, and therefore they will matter to models. Over time, the brands, publishers, and models with the most accurate and most real-time facts will win on consumer trust. Please reach out if you want to learn more.