People's "The Guarantee" dissected D/Cipher from the buy side. Seven structural headwinds. A competitive map. The independence paradox — building your own ad tech recreates the fragmentation that makes platforms attractive. All sharp. But the piece describes what the market sees. What matters is what the architecture is doing underneath.
D/Cipher is not one system. It is two.
Two Layers
Layer 1 is the content understanding engine. It maps every article across 40 brands into a taxonomy of scored relationships — "Hamburg museums" connects to "Hamburg restaurants" but not "Boston museums." The original NLP scoring produced 100 million of these connections. When People Inc ran the entire library through OpenAI in May 2024, 30% of the connections it found were "meaningfully different" from what the NLP had produced. Semantic understanding that keyword matching misses. This layer runs in batch. It doesn't care how much traffic the sites get.
Layer 2 is the behavioral layer — the thing that turned D/Cipher from a contextual platform into an intent platform. When a user reads three articles about kitchen renovations, then clicks on a tile backsplash guide, that sequence generates a real-time intent signal that pure content analysis can't replicate. This is the layer that erodes when sessions decline. And Google sessions have declined 63%.
The guarantee that People described works because Layer 1 works. Pre-computed content understanding is sufficient for upper-funnel brand targeting — proving that an article about Mediterranean cooking reaches food-interested audiences. That's what D/Cipher delivers reliably. That's what has never triggered the performance guarantee's re-run clause.
But the thing that made D/Cipher special — the part that justified building a proprietary ad platform instead of buying one — was Layer 2. Intent, not context. That's the part that's losing its substrate.
The Industry Pattern
Every major contextual platform pre-computes offline and serves cached segments at bid time. The programmatic bidding window is 10 milliseconds. Deep NLP on a page is not happening in 10 milliseconds.
GumGum deploys AWS Outposts co-located with DSPs to serve pre-analyzed contextual segments with sub-10ms latency. Seedtag processes 60 million articles per day in offline batches. IAS crawls 800 million domains and pre-computes roughly 400 contextual segments. What feels like "real-time contextual targeting" is actually a lookup against a pre-built index.
D/Cipher's Layer 1 follows this same pattern. The technology isn't the differentiator — the taxonomy is. Owning 40 brands across verticals gives People Inc a training corpus no standalone vendor can match. The OpenAI enrichment extended it. But it's the same architecture class as every other contextual vendor: offline compute, cached results, deal-ID delivery through PMPs.
What D/Cipher had that GumGum and Seedtag don't was Layer 2 — behavioral signals from 30 million daily users reading across an owned portfolio. That's a signal source, not a technology advantage. And it's a signal source that shrinks with every percentage point Google takes.
Three Futures
| Path | Architecture | Outcome |
|---|---|---|
| Embedding-only | Drop behavioral layer. Compete on taxonomy + OpenAI embeddings alone. | Commodity. Same architecture class as GumGum, Seedtag, IAS. Differentiation narrows to taxonomy breadth. |
| Federate behavioral | D/Cipher+ deploys on third-party sites. Collect session signals from the open web. Requires tag/pixel adoption at scale. | Preserves uniqueness. But this is the walled-garden problem — each publisher building another tag for advertisers to manage. |
| Interoperability | Expose D/Cipher signals through AdCP Signals Activation Protocol. Participate in open contextual exchange. | Solves fragmentation. But surrenders the control that makes D/Cipher a competitive moat. |
Jim Lawson's hire in March 2025 signals Path 2. He co-founded AdTheorent, an ML-powered DSP built on identity-free targeting. His background is buy-side — extending a publisher's signal infrastructure into the open web programmatic ecosystem. The "managed service" framing of D/Cipher+ is consistent with a pre-bid data layer: crawl third-party content via OpenAI, map it to D/Cipher's taxonomy via embeddings, deliver targeting segments through deal IDs. Rebuild the behavioral overlay on a broader substrate.
But Path 2 recreates the exact fragmentation problem that People's "The Guarantee" identified. Every publisher building a proprietary tag is another integration for agencies to manage. This is why AdCP's Signals Activation Protocol exists — it standardizes contextual signal exchange between publisher agents and buyer agents. The Curation Protocol, expected Q2 2026, will standardize contextual targeting requests specifically.
People Inc is not a member of AdCP. They are absent from the interoperability layer that could solve the very problem their strategy creates.
What the Migration Means
The signal migration is already happening. It's not a decision the engineering team needs to make — it's a consequence of a traffic pattern they didn't choose. When Google sessions fall 63%, the behavioral layer degrades whether you planned for it or not. The pre-computed layer absorbs the load because it was always traffic-independent.
The engineering question is whether to let Layer 2 atrophy (Path 1), rebuild it on a different substrate (Path 2), or share it through open protocols (Path 3). Each path has a cost. Path 1 costs competitive differentiation. Path 2 costs adoption — every publisher is trying the same play. Path 3 costs control.
The guarantee works. But the intelligence that backs it is shifting from something only People Inc could build — behavioral intent from a 40-brand portfolio — to something any well-funded contextual vendor could approximate with a good LLM and enough content to crawl. The taxonomy is a moat. The behavioral overlay was a deeper moat. One of them is filling in.
The question isn't whether D/Cipher keeps working. It's whether what makes it work keeps being theirs.