In The Integration Machine, I argued that People Inc's proprietary platform was built for uniformity — same CMS, same templates, same ad stack across 40 brands. Now the Inversion strategy demands the opposite: each brand as a distinct product. The tension between these two modes defines the engineering challenge for the next two years.
But the solution isn't in the frontends. It's in the layer beneath them.
What Graphene-Selene Actually Is
Reconstructed from job postings and public interviews: Graphene-Selene is People Inc's internal content services platform — a knowledge graph built on Java/Spring, GraphQL APIs, and graph database infrastructure. It stores content as structured nodes with semantic relationships: articles connect to topics, entities, authors, products, and each other. Every piece of content across all 40 brands lives in this graph.
Today it serves two consumers. The website rendering pipeline pulls content from the graph, applies brand templates (Java + Freemarker), and delivers HTML through Fastly's CDN. D/Cipher reads the same graph to extract taxonomy signals, map 12,000-term contextual classifications, and generate the embedding vectors that power contextual ad targeting without user tracking.
Two consumers. Both mature. Both revenue-generating.
Three more are arriving.
Five Consumers, One Graph
| Consumer | What It Needs | Status |
|---|---|---|
| Website rendering | Full article HTML, SEO metadata, ad placements, brand templates | Serving today |
| D/Cipher | Taxonomy nodes, contextual embeddings, intent signals, 100M connections | Serving today |
| Product apps | Structured data — recipes with ingredients, reviews with ratings, collections per user, cross-brand content for MyRecipes | Partially built |
| AI agents | WebMCP endpoints, structured responses, licensing metadata, PCM metering hooks | Not built |
| Internal AI tools | Corpus search, semantic similarity, content relationship queries, RAG-ready chunks | Early stage |
The product app consumer is the Inversion layer. MyRecipes needs structured recipe data across six food brands — ingredients, cook times, nutritional info, user-saved collections. The People App needs article metadata optimized for a TikTok-style swipe UX, not a traditional web page. Investopedia's tools need financial term definitions as structured data, not rendered HTML. Each Inversion product needs content served differently than the website does.
The AI agent consumer is the protocol stack layer. When Chrome ships WebMCP endpoints and Microsoft PCM goes live, agents will query content directly — no browser, no session, no pageview. The graph needs to serve structured responses with licensing metadata and metering hooks attached.
The internal AI consumer is the one Jon Roberts described to Digiday: "Team by team, finding the big, expensive, slow bottlenecks, and then seeing where and how LLMs can apply." Christina DiGangi's AI Editorial Operations team is deploying these tools across brands. They need the content graph to support RAG-style queries — semantic search, content chunking, relationship traversal — so editors can ask "what have we published about this topic across all brands in the last year" and get structured answers.
The Vector DB Signal
The strongest public signal about where this is heading: People Inc's SVP Platform Engineering job posting explicitly lists "Vector DB" alongside GraphQL and Gen AI as target technologies. This means the Graphene-Selene team is extending the content graph with vector embeddings — storing semantic representations of content alongside the structured graph relationships.
This is the architectural move that makes all five consumers viable from one substrate. The graph stores structure (articles → topics → entities). The vector layer stores meaning (this article about retirement planning is semantically similar to that article about 401k rollovers, even across Investopedia and The Balance). D/Cipher already uses embeddings for ad targeting. The vector layer generalizes that capability to every consumer.
What the Competitors Are Doing
People Inc isn't alone in discovering this pattern:
| Publisher | Platform | Architecture | Key Tradeoff |
|---|---|---|---|
| Vox Media | Duet | Killed Chorus. Headless WordPress + custom GraphQL + React frontend. | Gave up CMS control for WordPress talent pool. Custom GraphQL = still proprietary at the API layer. |
| Future plc | Vanilla | Tightly coupled CMS + contextual ad tech (Hybrid) + eCommerce (Hawk). 200 brands, article every 2 seconds. | Tight coupling is the moat AND the constraint. Hard to serve new consumers without refactoring. |
| Washington Post | Arc XP | Headless-first, REST + GraphQL APIs, React PageBuilder. Externalized as a product. RAG-based "Ask The Post." | Most headless, but React lock-in for first-class path. Shared roadmap with third-party customers. |
| CNN | Stellar | Proprietary. Piloted on Underscored, rolled to full newsroom. Publishing speed: minutes to seconds. | Built for editorial speed, not content API flexibility. No AI integration on roadmap. |
Two approaches in the market. Decouple (Vox, Arc XP) — separate the content layer from the presentation layer, accept the complexity of maintaining both. Integrate (Future, CNN) — keep the content system tightly coupled to its primary consumer, accept that new consumers will be hard to add.
People Inc is currently in the Integrate camp. Graphene-Selene was built to render websites. The Inversion strategy requires moving toward Decouple — but without losing the D/Cipher integration that makes the tight coupling valuable.
The Recommendation
The content graph is the single highest-leverage investment People Inc's engineering org can make, because every active priority routes through it:
- D/Cipher expansion needs richer embeddings and more content connections
- Inversion products need structured content APIs tailored to each product's data model
- Agentic web readiness needs WebMCP-compatible endpoints with licensing metadata
- AI editorial tools need semantic search and RAG-ready content chunking
The architectural move: expose Graphene-Selene as a multi-consumer content API platform, not just an internal system that renders HTML. GraphQL is already the query language. The vector layer is being added. The missing pieces are consumer-specific API surfaces — product apps get recipe schemas, agents get structured responses with metering, AI tools get chunked content with relationship context.
Vox Media abandoned its proprietary CMS because the maintenance cost exceeded the competitive value. People Inc can't do that — D/Cipher's entire revenue model depends on owning the content processing pipeline. The answer isn't to abandon the graph. It's to open it up — internally — to every system that needs it.
The graph is already there. The question is whether it stays a publishing system that happens to feed D/Cipher, or becomes a content platform that happens to publish websites.