Subscribe now

When OpenRTB launched in 2010, most people in advertising ignored it. A few engineers understood what it meant. A handful of companies built on it early. Those companies now run the programmatic industry. Everyone else spent the next decade playing catch-up.

Something similar is happening right now. It is moving faster. The stakes are a bit higher. And this time the infrastructure is not just changing how ads get bought. It is changing what an ad even is and how it impacts the customer’s purchasing journey.

Between October 2025 and March 2026, several protocols were released that together describe how AI agents will discover, negotiate, and purchase on behalf of consumers.

Agents are software that can take a goal and execute a multi-step workflow to achieve it without a human approving each step. You tell an AI assistant to find you a waterproof jacket under $180 and buy it. The LLM powered chatbot researches options, picks one, and completes the purchase. You never opened a browser. You never saw an ad. You never navigated a checkout flow.

That scenario is not hypothetical. It is happening today inside ChatGPT.

The questions it creates for advertising is fundamental. The ad does not need to be seen anymore. It needs to win the recommendation. That is a completely different game and it requires completely different infrastructure.

THE PROTOCOLS

Before getting into how they work together, here is what each one actually is.

HOW AN ACTUAL TRANSACTION RUNS

The best way I understand how these protocols work together is to follow a single transaction from beginning to end, both what the user experiences and what is actually happening underneath it.

Someone opens ChatGPT and types: I’m hiking in Seattle next week. I need a good waterproof shell jacket under $180.

From the user’s perspective the experience is frictionless. Three seconds later GPT recommends a Patagonia Torrentshell at $179, notes it is a sponsored result, and asks if they want to buy it. They say yes. A checkout widget appears inside the chat showing the total, their saved address, and the last four digits of their payment method. One tap. Done. The jacket arrives Tuesday. The user never left the conversation.

Here is what actually ran underneath that.

The LLM’s first move was not to search for jackets. It used MCP to pull context silently in the background. Seattle weather next week. The user’s saved size preferences. Their shipping address. Their previous purchase history. All of that happened before any external call was made.

Then for a commercially motivated query like this, the platform routed an intent signal through A2A. Brand agents at Patagonia, Arc’teryx, and REI were listening. Each one evaluated the query against their active AdCP campaign contracts. This is the part that is fundamentally different from the old model. There was no millisecond auction. No bid request firing. Patagonia’s campaign terms were already negotiated in advance. The agent just matched the query to the contract.

Patagonia’s agent responded with a structured AdCP payload. Campaign priority score. SKU. Waterproofing specs. A promotional offer. Sustainability data. The AI evaluated all three payloads against the user’s stated need and selected Patagonia. The sponsored label was attached. The user saw it. They said yes. At that point AdCP’s job was finished.

The AI switched layers. It called Patagonia’s ACP endpoint with the SKU, size, and shipping address. Patagonia’s backend verified inventory, calculated tax, and returned a finalized checkout session. If this were a Google Gemini session, UCP would have handled the same coordination handoff, with Google Pay as the payment surface rather than Stripe. However, both can run on the same underlying rails.

The user confirmed the cart. Stripe issued a Payment Intent scoped to Patagonia, for this exact cart total, valid one time only. The AI passed the client secret to Patagonia’s backend. Patagonia processed the payment through Stripe. The AI never saw the card number. If anything in the cart had changed, the token would have been invalid and the transaction would have failed at the payment layer. The security is baked into the architecture, not bolted on afterward.

Patagonia returned an order confirmation and tracking link through ACP. The AI surfaced it in chat. Patagonia remained the merchant of record.

The entire loop from stated need to purchase confirmation happened inside one conversation. No browser. No website. No ad that the user consciously processed.

The harder question is what happens when commercial incentives and user intent are not aligned. You may get a sponsored recommendation after a commercially motivated query, but that raises a problem both OpenAI and Google are actively working through. How do you balance surfacing the most relevant product against one that is objectively worse but whose advertising agent is willing to bid higher? In a traditional auction model that tension was managed by quality scores. In an agentic model nobody has cleanly solved it (yet).

ACP VS UCP

ACP and UCP solve the same problem. They are direct competitors. The winner will not be decided by which one is technically better. It will be decided by who controls the ecosystem.

For most merchants the practical answer is to support both. ACP reaches 700 million weekly ChatGPT users. UCP reaches Google AI Mode and Gemini. Amazon is building its own ecosystem through Rufus and Alexa+ and has joined neither protocol.

After ChatGPT Instant Checkout launched, Amazon updated their robots.txt to block OpenAI crawlers and removed 600 million products from ChatGPT’s shopping results. That move tells you a lot about how Amazon sees what is being built here.

Just last month we saw Amazon and Open AI develop a strategic partnership primarily focused on AWS but I’m sure the commerce bit is being heavily discussed.

WHERE ADOPTION ACTUALLY STANDS

MCP is the only protocol on this list that is fully mature. Thousands of production servers. Natively supported by all three major model providers. The de facto standard for connecting AI to external tools.

ACP is the only protocol on this list where a consumer can complete a purchase through an AI agent at meaningful scale today. Everything else is infrastructure being built in advance of volume that has not arrived yet. That is not a criticism. OpenRTB was in the same position in 2011. The companies that built on it early are the ones running the programmatic industry today.

The most important thing to understand about AdCP’s adoption status is who is missing. Google has not joined. The Trade Desk has not joined.

Those two names represent a significant share of programmatic demand. A supply-side without buy-side commitment is a specification, not a marketplace. The history of ad-tech standards includes well-designed protocols that never reached scale because the platforms controlling demand had no incentive to adopt them.

That is the challenge AdCP needs to work through as incentives are not aligned.

THE CLOSING TAKE

Protocol transitions in advertising take time but they tend to be permanent. OpenRTB took about three years from introduction to becoming the dominant mechanism for programmatic trading. The companies that built on it early established advantages they still hold today. The companies that waited built their strategies on infrastructure that someone else owned.

ACP is live. UCP is rolling out. People are buying things inside conversations today. The advertising layer is chasing a commerce layer that launched six months ahead of it.

The protocols are the infrastructure. The infrastructure is the leverage. The question every company in this industry needs to answer is which layer of this stack they are building on, which layer they are competing in, and which layer they are about to find themselves on the wrong side of.

The ones who wait for the dust to settle will be finding out which side of the stack they ended up on.

Thank you for reading The Ad Graph.

Share

WHAT GOOGLE ACTUALLY DID AND WHY IT MATTERS LESS THAN YOU THINK

Going back to the case itself because you need to understand what was actually found before you can understand why it is the wrong fight.

Google built three things that sit on top of each other. Google Ad Manager, which is the tool publishers use to manage their ad inventory. AdX, which is the exchange where buyers and sellers meet in milliseconds. And Google Ads and DV360, which are the tools advertisers use to buy. Owning all three layers meant Google could see what everyone was bidding before the auction settled, jump in to win at the last second, and systematically disadvantage any exchange or tool that competed with their own products.

The two most documented examples are Project Bernanke and Last Look. Project Bernanke used data from Google's buy-side tools to help AdX win auctions it otherwise would have lost. Last Look let AdX see the winning bid from header bidding competitors and beat it by a penny. Both ran for years before anyone outside Google knew about them. The court found this substantially harmed publishers who were leaving money on the table every single day without knowing it.

So the DOJ has two options now. The scalpel or the sledgehammer.

The scalpel means behavioral rules: force Google to make GAM work equally with competing exchanges, ban Last Look, open-source the auction logic, share data with competitors. The sledgehammer means selling AdX entirely, making it an independent company that has to compete on its own merits without Google's advertiser demand propping it up.

Both are legitimate remedies for what Google did. Here is my honest take though. The scalpel without real enforcement teeth is basically nothing. The DOJ argued this explicitly. They cannot be trusted to self-regulate. The sledgehammer seems to be the right call but runs into the problem of who actually buys AdX. Any large tech company that could afford it would trigger its own antitrust review. A private equity firm would be running critical internet infrastructure with no strategic interest in the health of the open web.

There is also an angle nobody is talking about enough. Let me put some actual numbers on it.

Google’s Network revenue, everything flowing through GAM and AdX on third party publisher sites, came in at roughly $31 billion in 2024. These are not broken out separately in Alphabet’s financials but AdX’s confirmed take rate is the key lever. The DOJ’s own expert witness testified it sits 19 to 27 percent above what a competitive market would support. Force that rate down to match Magnite or PubMatic at 10 to 15 percent and Network revenue falls to $15 to $18 billion, and that is before accounting for volume loss as advertiser demand shifts to other exchanges.

That sounds catastrophic until you look at the full picture. Alphabet crossed $400 billion in total revenue in 2025. A $15 billion hit to the Network business is less than 4 percent of total revenue. And the Network segment is already the only part of Google not growing, while Search, YouTube, and Cloud are all expanding at double digits.

So yes, GAM and AdX remain profitable under the scalpel. They do not go to zero. But here is the point that actually matters. At $15 to $18 billion with compressed margins and zero strategic data flowing back into the rest of Google's stack, this becomes a business Google has almost no incentive to maintain, invest in, or subsidize for small publishers anymore. The whole point of owning the ad stack was never the margins on AdX. It was the data. Every bid, every impression, every publisher signal flowing through GAM fed back into making Search and YouTube ads more valuable than anyone else's. Strip that data connection and you have a basic exchange charging basic fees.

With that said, even if the DOJ gets everything it wants, it does not fix what is actually killing publishers. Because the threat to the open web is not the Google tax. It is the zero-click future that chatbots are building right now.


THE REAL PROBLEM IS NOT IN THE COURTROOM

Here is what is actually happening while everyone is watching the antitrust trial. Google's AI Overviews, ChatGPT, Gemini, Claude and Perplexity are answering questions that used to require a click. Health information, travel planning, how-to guides, product comparisons, local recommendations. All of it is being summarized by AI before a user ever lands on a publisher's page. Penske Media, which owns Variety, Deadline, and The Hollywood Reporter, sued Google in September 2025 saying that AI Overviews were appearing on roughly 20 percent of searches linking to their sites, and that affiliate revenue from shopping links was down more than a third from the end of 2024 because of the traffic drop.

That is not a Google ad tech problem. Fixing AdX does not fix that. Banning Last Look does not fix that. The DOJ remedy, whatever it ends up being, addresses a market structure problem from the 2010s while the actual economic collapse of the open web is happening in 2026 from a completely different direction.

The New York Times and the Wall Street Journal did not move to subscriptions because they wanted to. They did it because the math stopped working. If you stay free and your content is publicly accessible, AI companies will scrape your journalism, summarize it, serve it to billions of users, and you get nothing. Not a click, not an ad impression, not a subscriber. Nothing. The paywall is not a business strategy. It is the only rational response to a world where your content has value to AI companies and zero value reaches you. Sure, there is GEO (Generative Engine Optimization) but it’s still in its infancy and publisher content will still be summarized with a link at the bottom that you could click if you want to. You probably won’t.

THE LICENSING DEALS ARE THE NEW BUSINESS MODEL AND MOST PUBLISHERS ARE NOT IN IT

Some publishers figured out how to get paid. The problem is which publishers.

News Corp signed a five-year deal with OpenAI reportedly worth more than $250 million, giving OpenAI access to the Wall Street Journal, Barron's, MarketWatch, The Times UK, and a dozen other properties. The Financial Times got a deal reportedly worth $5 to $10 million a year. The Guardian signed with OpenAI. The Washington Post signed with OpenAI. Reddit, which is not even a traditional publisher but happens to be the most cited source in AI models at three times the rate of Wikipedia according to Profound AI, got deals with both Google and OpenAI. Condé Nast, The Atlantic, Hearst, Axel Springer, the Associated Press. The list of publishers with deals reads like a who's who of the biggest media brands in the world.

That is the point. The deals are going to the publishers who have leverage. The ones with large audiences, recognizable brands, premium content, or unique data that AI companies genuinely need. Everyone else is getting harvested for free. The local news site, the independent blog, the niche trade publication, they are providing the content that trains the models that replace the need to visit them, and they are getting nothing in return.

The licensing model as it exists right now is not a sustainable framework for the open web. It is a private arrangement between the richest publishers and the richest AI companies. It is hush money at scale. News Corp CEO Robert Thomson called the OpenAI deal "the beginning of a beautiful friendship." What he did not say is that the friendship requires having enough leverage to get in the room in the first place.


WHERE THE MODEL NEEDS TO GO

Going back to what actually needs to happen, there are two paths being tried right now and neither one is complete.

The first is the licensing deal model. Pay publishers a flat fee for access to their content for training and retrieval. This is what most of the deals above are structured around. The problem is the evolution from flat rate to usage-based is just starting. People Inc., which did deals with both OpenAI and Microsoft, put it clearly on an earnings call. The Microsoft deal is pay-per-use and the OpenAI deal is all you can eat. Both can work as long as the content is respected and paid for. Reddit is pushing further, negotiating for dynamic pricing where they get paid more as their data proves more valuable to model outputs. That is the right direction.

The second model is the Prorata approach. More than 500 publishers have signed with Prorata, which runs an AI search engine called Gist.ai and gives publishers a 50 percent revenue share model based on how much their content actually gets used in generating answers. That is closer to what a fair system looks like. You get paid proportionally to your contribution. The problem is it only works within Prorata's ecosystem and it requires the major AI players like OpenAI and Google to adopt a similar framework, which they have very little incentive to do while the current arrangement lets them access most content for free.

The end state that actually sustains the open web is a standardized pay-per-response model where any AI company that uses a publisher's content to generate an answer owes that publisher a fraction of the value created. Not a one-time training fee. Not a flat annual license. A royalty tied to actual usage. This is what the music industry built with streaming after a decade of getting destroyed by piracy. Spotify pays per play. It is not a perfect system and artists argue constantly that the per-play rate is too low. But the model exists. Publishers and AI do not have that model yet and until they do, most of the open web is operating on borrowed time.


WHAT HAPPENS TO THE OPEN WEB

What you end up with is a bifurcated web and it is already taking shape.

The premium end survives. NYT, WSJ, FT, The Atlantic, The Economist. Enough brand loyalty for direct subscriptions, enough leverage for AI licensing deals. They will be fine and they already moved. Substack writers, YouTube channels, and podcast networks also survive because they built direct audience relationships that do not depend on search traffic. AI Overviews do not threaten a newsletter your readers already subscribe to.

Then there is the junk end. AI-generated content farms that exist purely to be scraped by other AI, churning out filler that makes pennies through whatever is left of the programmatic ecosystem. That market does not die, it just becomes worthless.

The ones getting wiped out are in the middle. Independent news sites, local journalism, niche trade publications, specialty blogs. These are the publishers with no leverage in licensing negotiations, no brand strong enough to drive pure subscription revenue, and no traffic model that survives the zero-click era. This is the majority of the open web by volume. These are the ones the DOJ case was supposed to help. And these are the ones the DOJ case does not actually help.

The irony of the whole thing is that Google's argument throughout the antitrust case has been that their ad tech is what keeps the open web alive. And in a narrow, technical sense they were right. GAM and AdX did bring advertiser demand to publishers who had no other way to monetize. But the open web that Google claimed to be protecting was already being dismantled by a different Google product, AI Overviews, at the same time the lawyers were arguing in court. They were handing out lifejackets while quietly draining the pool. The right hand was subsidizing what the left hand was destroying.


WHAT I ACTUALLY THINK HAPPENS

The DOJ remedy, whatever form it takes, will matter for the ad tech industry. It will change the competitive dynamics between exchanges, give publishers better yield on the inventory they do sell, and remove the most outrageous forms of self preference that Google used to tilt auctions. That is worth doing.

But if you are a mid-size publisher watching your search traffic decline and your ad revenue compress at the same time, the AdX divestiture does not change your trajectory. You are in a race to either build a direct audience relationship, get into one of the AI licensing arrangements, or figure out something entirely different before the economics collapse entirely.

The real ruling that will matter for the open web is not Judge Brinkema's decision on AdX. It is whatever framework, legal or commercial, eventually determines how AI companies pay for the content they use to generate answers. The licensing deals happening right now are private, inconsistent, and tilted toward whoever has the most leverage. That is not a sustainable model and it is not a fair one.

We shall see how fast the industry gets there. But the DOJ case, for all its significance, is not the answer to that question.

It’s just a tax Google is willing to pay because it’s already won.

Subscribe now

Let us go back to what he actually said. In 2024, Altman stood in front of an audience at Harvard Business School and called AI and advertising “uniquely unsettling”. He described it as a last resort business model. He was not subtle about it. Then in 2025 he hired Fiji Simo, the former Instacart CEO, and made her the CEO of Apps at OpenAI. Now we are in 2026 and OpenAI has officially announced it is testing ads inside ChatGPT for free users in the US. The last resort he mentioned two years ago is now here.

So what changed and why did this happen? ChatGPT has become the consumer product for the majority of the world. Gemini is there, Claude is there, but Claude is focused on B2B and Gemini has both B2B and B2C aspects with Google leaning into AI mode. But going back to OpenAI, they have over 800 million weekly active users and only five percent of them are paying for a subscription. Revenue went from six billion to twenty billion in a single year and it still is not enough. Eight hundred million users with five percent paying is not going to pay itself. OpenAI is running probably the biggest free service in the history of technology and the company lost approximately eight billion dollars in operating losses in just the first half of 2025. The infrastructure costs to keep the product alive and magical do not disappear just because the experience feels effortless.

The revenue picture is good but it is not self-sufficient and that is why you are seeing all of these unusual investment structures. The Stargate joint venture with SoftBank and Oracle is planning five hundred billion dollars in AI infrastructure over four years. That is definitely not a number you cover with five percent of your users paying for a subscription. So the answer is ads. That is how you pay for it.

Here is what we actually know from OpenAI’s announcement. They are testing ads at the bottom of answers where there is a relevant sponsored product or service based on your current conversation. The ads will be clearly labeled. Users can dismiss them. OpenAI says it will not sell your conversation data to advertisers and will not optimize for time spent in the product. We all know that what gets said at launch and what actually happens as revenue pressure grows are two different things. This is the trust play for right now. Paid tiers including Plus, Pro, Business, and Enterprise will not include ads. The ad product is specifically designed for the ninety-five percent of users who are paying nothing.

Now here is where it gets strategically interesting. The format they announced is just the starting point. What they are actually building underneath is something far more significant and it is very different from Google Search. Not Gemini, Google Search specifically.

Google’s ad model is built on keywords. You type three words and Google infers intent from them. That inference is worth billions because it works and Google has perfected that machine over decades. But it is still just three words. Google does not know if you are a forty-year-old training for a half marathon or a college student buying their first pair of running shoes. It makes a probabilistic guess using cookies, browsing history, and demographic signals stitched together over years of modeling. It is good at it. But ChatGPT is different.

ChatGPT has your entire conversation. Not three words, paragraphs. Context. Follow-up questions. The specific problem you are trying to solve. When someone tells ChatGPT they need a CRM, they run a ten-person sales team, they close around two hundred deals a month, and their budget is tight because they just raised a seed round, that is more targeting signal in a single session than Google has ever had on a single query. That is the structural difference. OpenAI’s internal framing for this is reportedly intent-based monetization. They are not building a display network. They are building the most accurate purchase intent signal the industry has ever produced.

Now should Google actually be worried? Not really, at least not yet. Google can easily integrate this into their AI mode and they are doing exactly that. But the immediate threat is real at the bottom of the funnel. If ChatGPT is capturing commercial intent queries and serving ads against them, that is real revenue that would have gone to Google Search. It is happening at scale right now across eight hundred million weekly users before a single ad dollar has even been sold. Meta is less immediately threatened because their game is audience targeting, not intent targeting. They know who you are and they know what you are likely to want. OpenAI is catching you at the moment of decision. Those are different muscles.

The single biggest risk for OpenAI is not building the product. It is building it in a way that makes users feel like the recommendations are for sale. The entire value proposition of ChatGPT is that it gives you the best answer. The moment someone suspects the best answer goes to the highest bidder, the product is broken. OpenAI knows this. The commitments around labeling, not selling data, and keeping ads below the response rather than inside it suggest they are threading this needle carefully for now. Whether it holds as revenue pressure increases is the real thing to watch.

Now let us get into how this actually works under the hood because the architecture here is genuinely different from anything that has existed before.

In a traditional ad system, whether that is Google Search, Meta, or programmatic display, the flow is stateless. A user sends a signal, an auction runs in milliseconds, a winning ad gets served, and that is it. The ad server does not deeply know what you searched for yesterday or what you will search for tomorrow. It knows this query and this moment. ChatGPT completely breaks that model. The conversation is stateful. Every message builds on the last one. By the time a user asks a product question, the model has potentially seen dozens of messages that tell it who you are, what you are trying to accomplish, and what your constraints are. That context window is the most valuable targeting dataset ever assembled in a single user session, something Google has been estimating and modeling for years but that ChatGPT gets immediately from the conversation itself.

Here is how the conversational ad serving likely works based on what has been reported. First, context processing where the model reads the full conversation thread and builds a real-time intent profile from everything said. Second, intent classification where the system determines whether the conversation has commercial intent and whether the user is researching, comparing, or ready to act. Not every conversation triggers an ad. Third, ad matching where relevant advertisers are paired to the conversation context, including both traditional sponsored placements and the more advanced generative ad format where the AI writes the copy itself. Fourth, response generation where the organic answer comes first and the sponsored content is appended below it clearly labeled. Fifth, feedback where users can dismiss ads and explain why, which trains relevance over time.

The most technically interesting part is what they call generative ads. This is not a static banner or a text link. The AI writes the ad itself in real time based on the advertiser’s product information and the specific conversation context. An advertiser provides features, pricing, and use cases and the model generates copy that speaks directly to what that particular user asked. Someone asking about CRM for a small sales team gets a different pitch than someone asking about CRM for enterprise procurement, even from the same product. That is not possible in any existing ad system at scale. Google’s responsive search ads approximate this but they are still working from static building blocks the advertiser wrote. This is different.

The hardest part and the one nobody is talking about enough is attribution. Every ad business lives and dies on measurement. You have to be able to tell an advertiser their spend worked. In search that flow is relatively clean. User clicked, user converted, here is the data. In a conversational interface the path is much murkier. The user might engage with a sponsored result in ChatGPT, not click, come back three days later in a new session, and then convert through a completely different channel. Which touchpoint gets the credit and how do you assign value to a conversation that influenced a purchase without a direct click? OpenAI is going to need to build new measurement frameworks from scratch. The traditional last-click model does not work here. Until that measurement layer exists and buyers trust it, the CPMs on ChatGPT inventory will be discounted. Not because the intent signal is weak but because marketers cannot prove the ROI in a way their finance teams will accept. That is the real near-term barrier to OpenAI building a massive ad business. Not format. Not user trust. Measurement.

The early signal on pricing tells you everything about what OpenAI is trying to build here. Early beta partners reportedly included Target, Adobe, Ford and Expedia with minimum commitments around $200K and CPMs sitting around $60, which is closer to streaming TV than anything you would see in search. They are not trying to win on volume. They are going after the upfront budget.

So what happens next? OpenAI starts conservative with clearly labeled units at the bottom of responses targeting SaaS and DTC brands where intent is obvious. Within twelve to eighteen months the generative ad format becomes the main product and static placements look like a placeholder by comparison. The measurement problem gets solved through a combination of post-click attribution and advertiser-reported conversion data fed back into the model. OpenAI builds something closer to Amazon’s retail media model than Google’s search model where the conversation, the recommendation, and the transaction all happen in the same place. Google responds aggressively through their own Gemini ad integration because they have the infrastructure, the signals, and the relationships to do it faster.

Here is my honest take on all of this. This is very much needed for OpenAI to stay afloat. The compute costs for eight hundred million users, the graphics cards, the data centers, the energy, it is not a sustainable model without another revenue stream. Paid subscriptions alone will not subsidize a product that is free for seven hundred and sixty million people. Ads are a proven business model. They work. They are probably not the announcement users want to hear but they are the only realistic way to keep this product alive, keep improving it, and keep it accessible to the people who cannot or will not pay.

The setup is there. The users are there. The money is there to be made. The miss here is timing. OpenAI should have started building this in 2024 when Altman was still calling ads a last resort. That hesitation was expensive. Google is not going to sit still. They have the ad infrastructure already built, they have the user data, they have the signals, and they have everything to gain by integrating Gemini properly into their ad stack. OpenAI has the intent signal and the conversational format. Google has everything else. That race is the story of the next two years and we shall see how it plays out.