“Hi there! I’m here to assist you in finding the best beauty products for your specific needs.”
That’s how Sephora’s AI Beauty Chat starts every conversation. Even if a shopper spent five days comparing two types of shampoo, clicked on the Sephora Spring Savings Event banner three times, and browsed product pages from two different ChatGPT links.
This agent has no context of what a shopper is looking for right now, and it certainly doesn’t have context about how that shopper historically buys across the Sephora website, app, and physical stores. This is (probably) not a model problem. The best foundation models in the world are capable of nuanced, sophisticated reasoning. But reasoning is only as good as the inputs that feed it.
That’s why context graphs are essential for consumer-facing AI agents.
The Context Problem Is a Data Problem
Agents, like humans, are fundamentally limited by the context they have access to. Ask a Sephora sales associate what the “best” shampoo is, and you’ll get a reasonable answer — based on what sells, what they personally like, or what’s being promoted that week. Tell that same person your hair type and goals, what you’ve already tried, and what your family members with similar hair love — and the quality of that answer is completely different.
The same principle applies to AI agents. Without context, even the most capable model defaults to generic or scripted responses. It treats every shopper as a first-time visitor, every support request as a fresh ticket, every user as average. And “average” is the enemy of a great consumer experience.
But there’s a catch: agents can only process a limited amount of context at once. You can’t dump all shopper behavioral history into a prompt and expect a useful response. So the context problem is actually two separate problems:
- Capturing everything an agent could possibly need to know about a shopper
- Delivering only the right slice of that context at the precise moment of action
This is exactly what a context graph is designed to solve.
The Difference Between Enterprise Context Graphs and Consumer Context Graphs
Context graph has become a popular term, and that’s creating confusion in the market. Most of the current discussion centers on enterprise context graphs — knowledge graphs that capture how a business operates, how decisions get made, what policies govern what actions. These are useful for internal agentic workflows: automating approvals, summarizing internal docs, routing tickets, etcetera. But enterprise context graphs are not what powers a great consumer experience. A consumer doesn’t care about your internal org chart. They care whether your agent understands them.
That’s the distinction Conviva is focused on: the consumer context graph — a structured, dynamic representation of who a user is, what they’re doing right now, and how they typically behave. It’s the intelligence layer that sits between your consumers and your agents, translating behavioral signals into actionable context in real time.
The Three Elements of a Consumer Context Graph
A well-built consumer context graph is composed of three interconnected layers:
- Historical behavioral patterns. This is the long-term memory of the graph. Not just what a user clicked, but the patterns that emerge across sessions, channels, and time. Do they typically research deeply before buying? Do they engage differently on mobile vs. web? Are they price-sensitive or convenience-driven? Pattern analytics is what makes this layer possible — moving beyond raw event sequences to recognizable behavioral signatures.
- Real-time context. This is the short-term memory — what the user was doing in the moments immediately before engaging the agent. Were they on a product comparison page? Did they just abandon a cart? Did they click through from a ChatGPT link? The agent needs this right now, not from yesterday’s batch job.
- Synthesized cohort context. This is the most underdeveloped layer for companies, and arguably the most powerful. Even with rich historical data on a specific user, there will be gaps. Synthesized context fills those gaps by mapping each user’s behavioral signature to similar behavioral segment — clusters of users who buy the same way, ask the same questions, and respond to the same kinds of agent interactions. This transforms what might otherwise be a thin user profile into a rich behavioral archetype that the agent can actually act on.
These three layers working together are what turn a context graph from a data structure into genuine intelligence.
Why Census-Level, Stateful Analytics Are Non-Negotiable
Here’s where most approaches fall short.
Sampling doesn’t work for consumer context graphs. If you’re only observing a fraction of your shopper base, you can’t build reliable behavioral patterns, you can’t identify cohorts with statistical confidence, and you can’t detect real-time signals fast enough to be useful.
You need to capture everything that every user did across every session, and then process it in real time. That means tracking not just that a user clicked something, but what state they were in before the click, what they did after, and how that sequence compares to the millions of other sequences happening across your entire customer base simultaneously.
This is the capability Conviva was built for. We’ve spent two decades processing billions of behavioral events per day at census scale for the world’s largest consumer brands — and we’ve built the context graph infrastructure on top of that foundation. No other company in this space combines full census telemetry with stateful behavioral modeling to power a consumer context graph.
The Context Graph Is Your IP — Not the Model’s
One thing I want to be direct about: the consumer context graph is not part of the AI model. It doesn’t train a foundation model that a vendor then owns or shares across customers.
The context graph is a separate knowledge entity — analogous to human long-term and short-term memory — that feeds context to the model at inference time. This means the models themselves are interchangeable. Whether you’re running on GPT-5, Gemini, or Claude tomorrow, your context graph persists. Your behavioral intelligence is yours.
This matters enormously for large consumer brands. The behavioral intelligence Conviva builds and manages on your behalf — the patterns, the cohorts, the real-time signals — represents your core competitive IP. Not ours. Yours. Conviva builds and maintains it; you own the intelligence within it.
What’s at Stake
Today, the gap between a generic “Hi there! I’m here to assist you.” and a truly contextual agent experience is enormous — and it’s almost entirely a context problem, not a model problem. The consumer context graph is the infrastructure that closes that gap.
We’re at the beginning of defining what this category looks like, what standards it needs to meet, and which approaches actually work at scale. Conviva is leading that conversation, and we’re just getting started.
More on how consumer context graphs work — and what it takes to build a good one — coming soon at ProductCon NYC.
Can’t wait? Get in touch.
