Guide · Updated November 14, 2026

What is a product graph?

A product graph links the records behind a product: the customer evidence, the decision, the work that implemented it and what happened next. With those links, a person or an AI agent can see why a piece of work exists, not just where its files and tickets are.

  • Ticket 4471: first visits booked into 15-minute slotsZendesk
  • Kestrel call: “Every location has its own rules.”Gong
  • Decision: let each clinic set its own rulesMeeting notes
  • PEL-812: booking rules per clinicJira
  • PR 2318, merged into release 4.8.0GitHub
  • Bookings fixed by phone: 23% to 4%PostHog
Six records in six places, and nothing connects the ticket to the release. Sample workspace: online booking software for clinics.

Your tools keep the records, not the reasons between them

Calls, tickets, specs, pull requests and releases already describe your product. Search can find any one of them. It can't tell you which call changed a decision, which limit survived review, or which release should change the next bet.

Here is one example, carried through this guide. Support tickets say new patients keep booking 15-minute slots when a first visit takes 30. That points to a pattern: online booking ignores each clinic's own rules, and 41 clinics hit it. The team decides to let each clinic set its own rules rather than make every first visit 30 minutes, and commits to it with one hard limit: a rule must never turn a patient away without offering a call back. A pull request ships it. Afterward, bookings fixed by phone fall from 23% to 4%, so the decision stands.

Every step there is a record your team already keeps somewhere. The graph is the links between them, and where each link came from: an agent should be able to reach the source, see who owns the decision and whether it was approved, and tell an agreed limit from a suggestion.

Eight records, and the questions they answer

This isn't a workflow your team has to follow. It names the records worth linking. Every question a team asks is a path through them. Choose a question to see which records it needs, filled in from the booking example.

Ask the graph

Start at what shipped and walk back to the decision, the pattern and the first customer who raised it.

  1. 01Step 7 on this path

    Source

    The call, ticket, research note, usage event or document, kept as it arrived.

    Zendesk ticket 4471: “New patients keep booking 15-minute slots, but a first visit takes 30.”

  2. 02Step 6 on this path

    Signal

    One observation from a source, with the account it came from.

    Northmere Family Medicine, $4k ARR: first visits booked too short.

  3. 03Step 5 on this path

    Insight

    A pattern across signals, with the evidence that supports it.

    Online booking ignores each clinic’s own rules. 41 clinics.

  4. 04Step 4 on this path

    Opportunity

    A customer problem worth solving, and what solving it is worth.

    Clinics need booking that follows their own rules. $639k at stake.

  5. 05Step 3 on this path

    Decision

    The bet the team chose, what it ruled out, who owns it and whether it is approved.

    Let each clinic set its own rules, over 30-minute first visits for everyone. Approved by the booking product manager.

  6. 06Step 2 on this path

    Initiative

    The committed work: scope, limits, what is out of scope, and how the team will know it worked.

    Booking rules per clinic (PEL-812): rules by visit type, a referral before a scan. Not in scope: visit pricing.

  7. 07Step 1 on this path

    Delivery

    The issue, pull request, release or agent run that carried it out.

    PR 2318, merged into release 4.8.0.

  8. 08

    Learning

    What customers did afterward, and what that changes about the decision.

Read from delivery back to the source

Booking rules per clinic exists because 41 clinics hit the same booking problem, first raised in Zendesk ticket 4471.

Sample workspace: online booking software for clinics.

What an agent can't learn from the repository

Repository context explains the code. Product context explains the choice. A coding agent can read the code, its dependencies and its tests. It can't read which customers are affected, which alternatives were rejected, which limits were agreed, or which questions are still open. Those live in the graph.

Give the agent the smallest reviewed set the task needs, keep the links back to the sources, and leave consequential product calls with people.

What “product graph” doesn't mean here

The phrase has other meanings. This guide means the graph behind product decisions.

Product catalog graph
Connects products, variants, attributes, offers, and customer behavior for commerce or recommendations. A product-development graph connects the evidence and decisions behind software work.
Software dependency graph
Connects code, services, packages, and runtime dependencies. It explains how software is assembled, but usually not why a product choice was made.
Knowledge graph
Models entities and relationships for a broad domain. A product graph is a purpose-specific graph whose relationships support product decisions, delivery, and learning.
Product context layer
Operates the graph across people and tools by adding permissions, ownership, review state, freshness, retrieval, and write-back rules.
Agent harness
Controls how an agent runs: instructions, tools, state, checks, and feedback loops. The harness can consume reviewed product context from the graph.

How Zentrik does it

Zentrik keeps these eight records linked for product teams, from customer evidence through decisions, initiatives, delivery and what happened next. People review each product decision. MCP and connected workflows carry the relevant context into Codex, Claude Code, Cursor, Jira, GitHub and other tools.

Sources

The definition is Zentrik's synthesis of current product-graph, requirements, protocol and decision-record writing. First published August 16, 2026; updated November 14, 2026 to carry one example through the guide.

  1. Unified Product Graph

    An open schema effort for connecting product information across work systems.

  2. Model Context Protocol specification

    The protocol boundary through which an agent can receive context or use tools.

  3. Jama Software: Product Context Layer

    A requirements and traceability perspective on product context for AI-assisted development.

  4. Michael Nygard: Documenting Architecture Decisions

    The case for short decision records that are superseded, never deleted.

  5. Zentrik: Product context layer guide

    The operating layer around a product graph: review, permissions, freshness and delivery rules.

Questions

What is a product graph?
A product graph links the records behind a product: customer evidence, insights, opportunities, decisions, the committed work, delivery and what happened next. Those links let people and AI agents see why a piece of work exists and which evidence and limits apply to it.
How is a product graph different from a knowledge graph?
A knowledge graph can describe any domain. A product graph is purpose-specific: its entities and relationships are designed to preserve product intent from customer evidence through decisions, delivery, and learning.
What is the difference between a product graph and a product context layer?
The product graph records product entities and their relationships. The product context layer makes those relationships safe and useful across people and AI tools by adding permissions, ownership, review state, freshness, retrieval, and write-back rules.
Why do AI coding agents need a product graph?
Repository context can show how software works, but it often cannot show which customer evidence supports a change, who approved the decision, or which trade-offs and non-goals apply. A reviewed product graph supplies that missing product intent.