You recently introduced Apostra as an “ad gateway”, something that filters and combines information to reduce costs and token usage. Is that purely technical, or do you also select which opportunities reach the buyer’s agent? Will you act as a neutral pipe, or curate offers based on their relevance to the brief?
Brian O’Kelley: We saw in the first programmatic era what happens if you don’t have standards. If you say everything is equal, you get a lot of fraud and a lot of crap. So I think we have to have an opinion on quality and on who we work for. I think we work for the advertiser. Our job is to make sure the advertiser gets their outcomes.
Could we have ratings like Uber or eBay? Maybe. We’ve talked about having historical stats on how often sellers meet their commitments. If you say you’ll run a campaign for 10,000 euros and hit a particular CPC, we know whether you do or not. So we could do things like that. We’re not doing that on day one.
That said, I think staying out of the transaction is really important. We don’t take a percentage of media. We don’t participate in media transactions. Everything is completely transparent. I think that’s really important.
Given that responsibility to the buyer, do you vet new seller agents before accepting them, to make sure they aren’t fraudulent?
Yes. Like everything else in adtech, you have to vet the company, the technology, the people involved and the media. Those are all different layers. You could have an agent built by a very credible company sitting on top of crappy media.
We’re very thoughtful and experienced in how to do this, but we’re not a brand safety company. We don’t intend to do the actual measurement ourselves. We’ll partner and work with the players in the ecosystem.
You offer consolidated billing but don’t guarantee buyer payments. If a publisher starts working with an unfamiliar buyer through Apostra, who assesses the buyer’s creditworthiness and bears the risk of non-payment?
We support both models: the seller can bill the buyer directly, or we can consolidate. But even when we consolidate, we’re not participating in the transaction. The buyer still has to pay those ten publishers directly.
Let’s say a buyer connects to Apostra and says, “I want to work with these ten publishers.” The first thing we do is send each publisher a message using MCP: “Do you want to work with this buyer, and do you want to give them credit?” That’s part of the protocol. Each seller explicitly extends credit to the buyer.
We will support credit cards. We may integrate with third-party companies that can help as well. If a buyer doesn’t have great credit or isn’t well known, we could have a third party step in and do that clearing for them.
Publishers could also say, “I don’t want to take credit risk. I’m going to contract that out to a third party.” If you think about Wall Street, there are entire companies that are just clearing houses managing things like payments. I think we’re overdue for that in advertising. It will be good for everybody to have that unbundled from the system. We don’t need to be your bank. We can connect you to Stripe or a third-party partner.
You also give buyers access to Meta and TikTok. As far as I know, they don’t have seller agents. How do those connections work? Are you using their APIs or MCP servers, even though your vision is based on agent-to-agent interaction?
We built an AdCP adapter for every major social platform. We took their APIs and converted them into AdCP, so you can fully control each platform through AdCP. These are proprietary adapters we built as part of the Apostra platform and got certified with the social platforms.
We’re a certified developer on TikTok, Meta, Google, Amazon, Pinterest, Reddit, Spotify, ChatGPT and LinkedIn. Our thesis is that, at some point, they’ll build their own. For now, I think we have the only full MCP suite for social.
What does an AdCP adapter add compared with connecting Apostra to Meta through an MCP server? You’ve argued that APIs change frequently and can be difficult to maintain. Is the adapter essentially a stopgap until social platforms support agent-to-agent interaction natively?
Let’s say you want to build a buyer agent and control hundreds of things: TF1, M6, Le Monde, Le Figaro, TikTok, Yahoo, Meta, Google. It would be nice to have one protocol for all of them. That’s what AdCP gives us.
The problem is that not all of them speak AdCP. Apostra becomes the gateway layer: we translate from AdCP to APIs, to a DSP or to an ad server. We’ve basically built a universal translation layer between everything and AdCP.
If Meta supported AdCP, you might still go through the gateway because of the convenience of consolidating everything in one place. But you shouldn’t need to know whether what’s behind us is AdCP, APIs or people.
All you should know is that, if you want to run a campaign on Meta, it works. You get reporting and event notifications. You get an efficient way for an agent to manage Meta, and it would take almost no effort to make it manage every other platform.
If you look at the consolidation of ad spend on the big platforms, one reason is that it’s hard to integrate with Meta. Once you do, you have a whole team managing it. Can you afford the same effort and human capital for ten platforms? Almost nobody can.
You mentioned DSPs and SSPs. AdCP is often presented as an alternative to the existing infrastructure, including by you. Will the media buying industry still require DSPs or SSPs or do you think they will gradually disappear?
I think the idea of a demand-side platform is confusing today. When the internet was the web, all you had to do was buy the web. But now the lines are blurring, and you can’t buy every digital ad through a DSP.
You can’t buy social. You can’t buy ChatGPT. DSPs don’t even cover the whole web. You want to buy native ads on The New York Times? Those aren’t in a DSP either.
Are we supposed to think of the DSP as only covering RTB surfaces? Fine, but then, as a buyer, I need a DSP for some things, APIs for others and people for others. It feels like we’re moving toward a different layer: either universal, where you can buy everything, or very niche and focused on one thing. DSPs are caught in the middle, trying to figure out which they are.
People think I’m anti-DSP. I’m just saying that looking at a DSP as a universal way to buy the web isn’t accurate. If that isn’t true, what’s the point?
Ad tech has struggled to move away from CPM-based or media-linked fees, partly because buyers know they only pay when they spend on media. Your model makes it harder to estimate the cost upfront. Can you share an example of what a seller pays for a live campaign, and how that compares with a typical fee of, say, 5% of media spend?
Think of it very simply: if you use Apostra and bring your own agent, it costs nothing. It’s free. We don’t charge for listing on the platform.
If you want to use our storefront agent instead of building your own, we charge based on usage. Every time a request comes from a buyer, we assemble your products, rate card and everything else, and come back with a proposal. That costs one “intelligence unit,” as we call it. Call that five euros. Every proposal costs five euros. A thousand proposals would cost 5,000 euros.
The thinking is that this is good for your business because you’re getting a lot of proposals. What percentage of revenue is that? It depends on how many you win and how much they’re worth.
We expect some publishers to be very selective about the proposals they respond to. You might say, “I’m a sports publisher. I only want to respond to sports briefs. I only want CTV proposals. I only want proposals from Publicis.”
What I like is that your fees are based on value. If you don’t get any proposals, you don’t pay. As a publisher, if I don’t have an opportunity to generate revenue, I don’t pay.
If a publisher already has its own seller agent, like Prisma Media does in France, is there any reason to use your storefront agent as well? Would it bring additional demand?
We treat them the same way. If a buyer says, “I want to buy Prisma,” whatever agent Prisma has registered will get that demand.
If they want to invest the time in building their own agent, and they think it does a great job responding to demand, fantastic. If they don’t want to build their own technology, they can use our storefront agent. Other companies are also building and hosting agents.
It’s like a lot of technology: publishers can build their own ad server or use somebody else’s. We’re not biased toward our own agent.
So there’s no preferential treatment for your storefront agent?
Right. Generally speaking, we get paid by the buyer. Our mindset has always been that big brands and agencies should be able to trust their partners to operate in their best interests.
We built the storefront agent because some sellers didn’t have the capacity to do it, and there weren’t many companies building them. Maybe in two or five years, we won’t need to do that anymore. But we also have expertise that can be helpful, having done this for a while.
It’s cheap compared with a lot of enterprise technology. I think it’s a very affordable option, and it helps make sure publishers don’t get ripped off early in the market’s development.
I’d love to get every major publisher in France on Apostra, whether they’re digital, influencers, traditional media or out of home. If they were all available, we could move a lot of money. I don’t care whether they use our agents or their own. I want everybody to be agentic because it creates so much more opportunity for a marketer to find the right medium.
Right now, if you’re stuck in programmatic, you don’t have access to everything. That’s a real problem for journalists and consumers.
You’ve had a public disagreement with IAB Tech Lab. If I built a buyer or seller agent using AAMP, would it be compatible with Apostra, or would I have to use AdCP?
It’s not compatible. I think their struggle is that they’re very committed to building on their own legacy standard. We were building for social, traditional media and digital.
Why were we able to build a Meta adapter? Because AdCP supports almost everything Meta’s API does. That’s not possible with AAMP.
We’ll see what happens with the IAB standards. If they gained a lot of traction, I could see us building an adapter to translate. But to date, we haven’t seen a single company come to us with an actual AAMP agent. Every seller, including some very big companies that have come to us recently, has built on AdCP.
I think the practical reality is that AdCP is the standard. The question for Tech Lab is how long they keep building something no one is using.
You’ve said around one million dollars has gone through Apostra. Have you already brought new media partners onto the platform and helped buyers discover opportunities they wouldn’t otherwise have considered?
Yes. We ran a campaign in South Africa, where we had no connections. We weren’t planning to launch there, but one of our big clients asked us to. We were able to go live with, for example, a local publisher that runs audio ads in taverns and grocery stores.
When people are shopping or having a drink, an ad can come on and say, “Try Guinness,” or “You should buy carrots,” or whatever it might be. They’d never done programmatic. There’s nothing to bid on. We were able to bring an entirely new kind of company to the platform, which I thought was really cool.
I think we’ll see a lot of examples like that. This isn’t about the web, or even just the surfaces we’re used to. Retail audio and in-store retail media are massive opportunities. If we can help retailers do this, we can also help more manufacturers access it.
There’s so much out there that we don’t see because it isn’t listed. It has to be sold. That’s a disservice to smaller buyers because they don’t get sold to. It’s also a disservice to non-standard media outlets, whether that’s a creator, an influencer, an athlete or a pub. They just don’t have the resources. I think that’s a really cool opportunity.
You also list AI assistants among the advertising surfaces Apostra supports. Are any of those connections live today?
We’re live on ChatGPT. We’ve also run pilots with a couple of AI ad networks that are integrating into those surfaces. We can put ads into a coding agent while you’re waiting for it to finish, or into an AI companion you’re chatting with.
I think we’ll see more of that soon. It’s a very nascent space, but it’s cool that we’re not limited to traditional digital advertising.
Is the ChatGPT connection also through an AdCP adapter?
Yes. We’ve talked to them about supporting it natively, which I think would be really cool. They haven’t yet. Part of my backlog is to reach out to these platforms and say, “As much as I love your APIs, could you do this yourselves? I don’t need to be doing it for you.”
What incentive do they have to support AdCP natively? Would it bring them more business than connecting through your adapter?
It would help, but it depends. If Apostra is the primary gateway, it won’t matter. If there are other gateways, or large buyers building agents that connect directly, then it would be good for them to support AdCP directly. The more AdCP agents there are, the better. I’m realistic: WPP is going to connect directly to ChatGPT.
