The Surface

On May 5, OpenAI opened its Ads Manager to every US business. No minimum spend. CPC bidding at $3–5 per click. A full competing ad platform, built on licensed publisher content, targeting $2.5 billion in ad revenue this year.

Cannes Lions starts Monday. This is the technology briefing.

The Architecture That Looks Familiar

OpenAI published enough about its ad targeting system to reconstruct the architecture. It works like this:

1. ChatGPT generates an organic response — potentially referencing licensed publisher content.
2. A separate system extracts semantic signals from the conversation: topics, user intent, research depth.
3. Dedicated classification transformers convert those signals into embedding vectors.
4. Advertiser campaigns are also represented as embedding vectors — encoding target contexts, messaging themes, and semantic intent clusters.
5. Cosine similarity scoring matches conversation embeddings against campaign embeddings.
6. Multi-stage filtering: lightweight models screen irrelevant ads first, heavier models score the shortlist.
7. Winning ads render as chat_card elements below the response.

This is embedding-based contextual targeting. No cookies. No user profiles. The content determines the ad, not the person.

It should sound familiar. It’s the same pattern D/Cipher uses.

D/Cipher runs OpenAI’s models to generate semantic embeddings of content pages across Dotdash Meredith’s 40+ brands. Those embeddings power contextual matching for advertisers — privacy-first, no third-party cookies, content-as-signal. It drives over 30% of direct buys and is “over half” of People Inc’s direct digital revenue — roughly $126.5 million per quarter.

Both systems solve the same problem the same way: turn content into vectors, turn campaigns into vectors, measure the distance between them.

The difference is who controls the surface.

DDM pays OpenAI for embedding API access to power D/Cipher. OpenAI licenses DDM’s content to train ChatGPT. ChatGPT answers questions that would have gone to DDM’s sites. OpenAI serves ads against those answers using the same embedding architecture D/Cipher pioneered. The licensing fee is a consolation prize for funding the replacement.

People’s sibling blog identified the arithmetic: if 5% of ad budgets shift from publisher-direct to ChatGPT-direct, that’s $125 million — roughly equal to D/Cipher’s quarterly contribution. The licensing fees don’t offset this.

The Infrastructure Being Built Without D/Cipher

The content-ad paradox is the strategic threat. But the engineering threat is more immediate: the entire agentic advertising infrastructure is being built, right now, without D/Cipher at the table.

Three layers, all absent:

Layer 1: The Agent Registry

IAB Tech Lab’s AAMP Agent Registry is the emerging directory of advertising AI agents. Ten entries so far: Amazon Ads MCP, PubMatic MCP, Equativ MCP, Optable Audience Agent, Dstillery, Burt Intelligence, HyperMindZ, Mixpeek, and IAB’s own registry agent. D/Cipher is not registered.

Registration requires categorization, purpose definition, MCP/A2A support details, domain verification, and endpoint health checks. It’s not a heavy lift. But it requires exposing D/Cipher as an agentic endpoint — something the current architecture doesn’t support.

Layer 2: The Transaction Protocol

AdCP — the Ad Context Protocol — launched in October 2025. Co-founded by PubMatic, Yahoo, Scope3, Optable, Swivel, and Triton Digital. Governed by a pending 501(c)(6) nonprofit with 102+ members. It defines how AI agents discover inventory, initiate media buys, sync creative, and exchange audience signals. Built on Anthropic’s MCP. Already in production — Magnite ran its first AdCP buy in December 2025.

The Prebid Sales Agent is open-source. Python. Implements AdCP’s Media Buy protocol. Supports both MCP and A2A. Deploys in two minutes with Docker. Integrates with Google Ad Manager out of the box. Ten standardized task types: Discovery, Proposal, Negotiation, Approval, Activation, Pacing, Optimization, Reporting, Billing, Attribution.

D/Cipher implements none of these.

Layer 3: The Execution Engine

PubMatic’s Decision Fabric, launched June 1, lets third-party AI models run inside the programmatic supply path — at the exact moment of auction, on live bidstream signals, at ~1ms latency on NVIDIA GPUs. Built on IAB Tech Lab’s ARTF containerization protocol. Partners include inPowered AI, MiQ, Chalice AI, and SWYM.AI. It processes 2.7 trillion advertiser bids daily across 2,000+ publishers and 100,000+ sites.

Decision Fabric is the layer where D/Cipher should live — a containerized execution environment purpose-built for running contextual intelligence models at auction time. D/Cipher’s embeddings running inside PubMatic’s bidstream would be a natural integration. It hasn’t happened.

What Agents See When They Look for D/Cipher

This is the engineering question that matters for Monday.

When a buyer agent — built on PubMatic’s AgenticOS, or WPP’s Buyer Agent for Video, or any of the 102 AdCP members — tries to find premium contextual targeting for an auto campaign across lifestyle and home content, here is what it can discover:

• Optable Audience Agent — registered in AAMP, integrated with PubMatic, agent-driven audience planning with first-party data
• PubMatic Inventory Discovery Agent — queries live inventory, returns structured data with expected impressions, pricing, confidence levels
• Magnite Seller Agent — built on AdCP, production since December 2025
• Prebid Sales Agent — open-source, deploys in two minutes, implements full Media Buy protocol

D/Cipher? The agent can’t find it. No AAMP registration. No AdCP endpoint. No MCP server. No /.well-known/agent.json manifest. D/Cipher exists as a managed service inside DDM’s direct sales organization, invisible to every agentic buyer in the ecosystem.

This matters because the distinction between “direct” and “programmatic” is exactly what agentic advertising dissolves. When AI agents buy media, every transaction looks like programmatic — structured, protocol-driven, machine-readable. The “direct relationship” premium only holds if humans are making the buying decisions. Optable’s research found that only 18% of publishers are ready to scale agentic. But that 18% will capture a disproportionate share of AI-mediated spend.

What I’d Build Before Monday

Three things, ordered by engineering effort:

1. Register D/Cipher in the AAMP Agent Registry. This is administrative, not architectural. Define the agent categorization, map D/Cipher’s capabilities to the registry schema, verify the domain, submit. Minimum viable presence. Do it this weekend.

2. Publish a D/Cipher MCP server. Expose D/Cipher’s contextual intelligence as an MCP endpoint. The PubMatic MCP server spec is public. The IAB Tech Lab SDKs are open-source. D/Cipher’s embedding-based classification is already API-shaped — wrapping it in MCP requires an adapter, not a rewrite. This makes D/Cipher discoverable by any agent that speaks MCP, A2A, or AdCP. Target: operational by Cannes Week.

3. Pilot D/Cipher inside Decision Fabric. PubMatic’s containerized execution layer runs third-party models at ~1ms on NVIDIA GPUs against live bidstream. D/Cipher’s contextual embeddings running at auction time — not pre-computed, not batch-processed, but live — would be the highest-value integration available. This is the medium-term play. Start the conversation at Cannes. Ship in Q3.

None of these require abandoning D/Cipher’s core architecture. The embedding pipeline stays. The contextual intelligence stays. The privacy-first design stays. What changes is the surface area: D/Cipher becomes findable, transactable, and executable by agents — not just by humans with phone numbers.

The Surface Is the Stack

D/Cipher was ahead of its time. Embedding-based contextual targeting, privacy-first, no cookies — this was the right architecture before anyone else built it. The Inference Tax is real. The intelligence is genuine. In a world where humans read websites, D/Cipher wins.

But the surface is shifting. Users ask ChatGPT instead of searching Allrecipes. Buyer agents query protocol registries instead of calling sales reps. Auction-time AI models run inside SSP infrastructure instead of returning batch recommendations.

OpenAI built D/Cipher’s architecture on its own surface, using content licensed from the company D/Cipher was built to serve. Criteo is running 1,000+ brands with conversion rates approaching 2x traditional search. PubMatic is running 30+ fully autonomous agentic campaigns across five countries. The WPP coalition is writing governance rules for CTV buying with Disney, Netflix, and Fox — no publishers at the table.

Intelligence without surface is a feature. Intelligence with surface is a platform. Intelligence with surface, infrastructure, and protocols is a business.

D/Cipher has the intelligence. Monday is day one of a week where everyone else shows their surface. ■