Beyond Tables: A Knowledge Graph for Product Management
Tables answer how many. A knowledge graph answers what connects to what. What a product knowledge graph holds, what it lets a PM ask, and how to start without a new database.
Share
A knowledge graph for product management stores the things a product team cares about (accounts, users, features, feedback, decisions, experiments) as nodes, and the relationships between them as typed edges. That lets you ask multi-hop questions a table can't answer cheaply, such as "which features did we ship for accounts that later churned, and did those accounts use them?"
Key takeaways
Google launched its Knowledge Graph in May 2012 with more than 500 million objects and 3.5 billion facts and relationships. Its slogan, "things, not strings", is still the clearest summary of the idea (Google, 2012).
Atlassian's State of Teams 2025 found teams spend over 25% of their workweek searching for information (Atlassian, 2025). Product context scattered across tools is a large part of that.
In Microsoft's GraphRAG study, a graph-based approach beat plain vector retrieval on comprehensiveness in 72-83% of head-to-head comparisons on a 1-million-token dataset (Edge et al., 2024).
You don't need a graph database to start. An edges table and a recursive query in Postgres cover most product questions.
Start with one multi-hop question you can't answer today, not a data migration.
What is a knowledge graph for product management?
Neo4j defines a knowledge graph as "an organized representation of real-world entities and their relationships" (Neo4j, 2024). It has three parts: nodes (the entities), relationships (the edges between them) and an organising principle, meaning a schema that says which kinds of nodes and edges exist.
For a product team, the entities are the nouns you already argue about in planning: accounts, users, features, interview notes, support tickets, insights, opportunities, decisions, experiments and releases. The relationships are the verbs: reported, supports, contradicts, derived from, implements, measured by.
The difference from a table is where the meaning lives. In a spreadsheet, the link between a customer quote and the feature it inspired sits in someone's head or in a free-text "notes" column. In a graph, that link is a record you can query, count and follow.
How is a product graph different from product analytics?
Most "data-driven" product work runs on event analytics: rows of events per user, rolled up into funnels and retention curves. That answers what happened, and how often. A product knowledge graph adds why we think it happened and what we did about it, by connecting behaviour to the qualitative evidence and the decisions around it. Analytics is one input to the graph, not a replacement for it. We cover the split in more depth in product intelligence vs product analytics.
Real-world knowledge graph examples
Google's 2012 launch is the best-known example. Instead of matching the text you typed, search started resolving it to an entity (a person, a city, a film) and showing facts connected to it (Google, 2012). A product knowledge graph applies the same move to your own data: "Acme Ltd" stops being a string that appears in 14 tools and becomes one node with every ticket, call and decision attached.
Why do tables stop working for product questions?
Relational tables are excellent at transactions and totals. They get awkward when the question is about paths.
Take a question a Head of Product might ask before a renewal review: "Which features did we ship this year because of requests from accounts that are now at risk, and did those accounts use them?" That is five hops: account → feedback → insight → decision → feature → usage. In a typical stack, each hop lives in a different tool (CRM, support desk, docs, Jira, analytics). The answer takes exports, lookups and trust in whoever matched the IDs.
Inside one SQL database, each hop is a join. Five joins is fine when the path is fixed. It gets painful when the path length is unknown, as in "show me everything connected to this account within three steps". Recursive SQL and graph query languages exist for exactly that shape of question.
What does disconnected context cost a team?
The everyday symptom isn't a slow query. It's people searching. Atlassian's State of Teams 2025 surveyed 12,000 knowledge workers and 200 executives and found teams spend over 25% of their workweek looking for information; it put the cost across Fortune 500 companies at 2.4 billion hours a year (Atlassian, 2025). Product managers feel this more than most, because their raw material (what customers said, what was decided, and why) is spread across more tools than almost any other function's.
What goes into a product knowledge graph?
Five families of nodes cover most product work. Keep the first version this small.
Node type
Examples
Typical edges
Customer
Account, user, segment
belongs to, reported, uses
Evidence
Interview note, support ticket, Slack thread, call transcript, survey answer
reported by, supports, contradicts
Interpretation
Insight, theme, opportunity, assumption
derived from, addresses
Commitment
Decision, PRD, roadmap item, ticket
based on, implements, blocks
Outcome
Release, experiment, metric change
shipped in, measured by, validates
Two rules make the graph useful rather than decorative:
Keep evidence separate from interpretation. An insight node should point back to the raw tickets and transcripts it was derived from, never replace them. Otherwise an AI summary quietly becomes the "source".
Type every edge. "Related to" everywhere produces a hairball. "Supports" and "contradicts" let you count how much evidence sits on each side of a decision.
What can a PM do with a knowledge graph that a table can't?
Map the blast radius before you change something
Before you deprecate a feature, rename a plan or change an onboarding step, follow the edges outward: which accounts use it, which open requests mention it, which decisions and specs assume it exists, which experiments are running on it. "Who will this affect?" goes from a week of Slack messages to one traversal.
Trace any decision back to its evidence
Open a roadmap item, follow based on to the decision, then supports to the interviews and tickets behind it. If the trail ends at "the CEO mentioned it on a call", you've learned something. Our buyer's checklist for customer evidence software shows how to test whether a tool keeps that trail intact. The companion piece on decision graphs for small AI teams covers how to record the decision side so the trail exists in the first place.
Find leading indicators in paths, not totals
A total tells you 30% of trial accounts activated. A path tells you which sequences came before activation, for example accounts that connected an integration and invited a teammate in their first week. Tables give you lagging counts; paths give you the behaviour that predicts the next count. Sequence questions are graph questions.
Give AI assistants grounded context
Language models answer better when retrieval follows relationships instead of only matching similar text. Microsoft researchers built a knowledge graph from source documents and compared it with standard vector retrieval on broad questions over about 1 million tokens of podcast transcripts and 1.7 million tokens of news articles. The graph approach won on comprehensiveness in 72-83% and 72-80% of comparisons, with an LLM as judge (Edge et al., 2024). For product work, that's the gap between "here are five similar tickets" and "here is the theme, the accounts behind it, and what we already decided about it".
Worked example: one churn question, five hops
This example is illustrative. Say a 6-person B2B team has 40 interview notes, 1,200 support tickets, 85 accounts and 18 roadmap decisions from the last two quarters. Seven accounts are flagged as churn risks ahead of renewals.
The question: "Did we build anything for the seven at-risk accounts, and did they use it?"
The table approach. Export accounts from the CRM, match support tickets by email domain, search the docs for account names, cross-check Jira epics, then pull usage per feature from analytics. Realistically half a day, with matching errors at every step.
The graph approach.
Start at the 7 account nodes.
Follow reported to their 46 tickets and 9 interview notes.
Follow supports to the 4 insights those items back.
Find the 2 decisions whose based on edges point at those insights.
Follow implements to the 3 features those decisions produced, then read usage.
Result: of the 3 features shipped partly for these accounts, only 1 of the 7 accounts used any of them in the last 30 days. That's a renewal conversation and probably a discovery gap: the team built what was asked for, not what would have kept them. Each step is a lookup along an existing link rather than a fuzzy match between tools, which is the whole case for a knowledge graph for product management.
Do you need a graph database?
Not at first. Dedicated graph databases such as Neo4j, Amazon Neptune and TigerGraph are built for deep, variable-length traversals over very large graphs. Most product teams have thousands or tens of thousands of nodes, not billions.
A plain edges table (from, to, type) in Postgres plus WITH RECURSIVE queries handles traversal well at that size. The PostgreSQL 18 documentation even includes built-in SEARCH and CYCLE clauses for depth-first or breadth-first ordering and loop detection (PostgreSQL, 2025). Specky's own Product Graph runs on Postgres this way.
Gartner forecast in 2021 that graph technology would be used in 80% of data and analytics innovations by 2025, up from 10% (TechTarget, 2021). Treat forecasts like that as direction, not as a reason to migrate. The useful test is whether your own questions are path-shaped.
Consider a dedicated graph database when:
Your analysts regularly write queries with more than five joins to answer basic product questions.
Questions involve variable path lengths ("anything within three hops of this account").
Marketing, support and product data sit in separate schemas and you can't assemble one view of a customer.
Real-time recommendations or personalised experiences time out or lag.
How do you start? A four-step plan
The cheapest way to find out whether a knowledge graph for product management is worth it is one path and five questions, not a platform project.
Map one path. Pick a core flow, such as feedback → decision → release → usage, and draw it as nodes and edges with your data or engineering lead. Stop at 5 to 8 node types.
List the multi-hop questions you can't answer. Write down 5 questions that currently take more than an hour. Good candidates: does a marketing channel affect long-term retention, does adoption of feature A predict adoption of feature B, which decisions rest on a single customer.
Choose the cheapest stack that answers them. Start in Postgres or your warehouse. Move to Neo4j, Neptune or TigerGraph only when the questions demand it.
Pilot on one launch. Connect the evidence, decision, tickets and post-launch metric for your next feature. After 30 days, check whether the graph answered questions faster than the old way. If it didn't, find out which link was missing before adding more.
Modelling everything. A 40-type ontology nobody maintains is worse than 6 types everyone uses.
Untyped edges. "Related to" everywhere gives you a pretty picture and no answers.
No provenance. If nodes don't keep a link to their source (the ticket URL, the transcript timestamp), nobody can check the graph against reality.
Counting demo data. Sample or seed content must be labelled so it is never counted as real customer evidence.
Expecting it to decide. A graph shows what connects. Judging what matters is still the team's job.
Manual upkeep. If nothing writes to the graph automatically, it decays like any wiki.
Where Specky fits
Specky builds this graph for you. It reads customer calls, feedback, Slack, Jira and analytics into a Product Graph, stores each signal, insight, decision and experiment as a typed node, and links them with typed edges such as supports, contradicts and derived from. Specs and tickets keep inline citations back to the evidence, and after launch you can check whether the bet worked. For a tour of what that map looks like in use, read the Product Graph as a living map of your product, or see how product intelligence works.
Frequently asked questions
What is the difference between a knowledge graph and a product graph?
A product graph is a knowledge graph scoped to one product's work. The general idea, entities connected by typed relationships, is the same. The node types differ: a product graph holds accounts, feedback, insights, decisions, specs, tickets, releases and metrics, so the team can trace any piece of work back to the customer evidence behind it and forward to its outcome.
Is a knowledge graph the same as a graph database?
No. A knowledge graph is a data model: nodes, typed edges and a schema. A graph database is one way to store it. You can hold a knowledge graph in Postgres with an edges table and recursive queries, which is enough for most product teams. Dedicated graph databases earn their place at very large scale or with deep, variable-length traversals.
What is a multi-hop question in product management?
A multi-hop question follows several relationships in a row. "Which decisions cite feedback from accounts that churned?" goes account → feedback → decision, which is three hops. Single-table tools answer one hop well. Each extra hop across separate tools adds an export and a manual match, which is where answers slow down and errors creep in.
How big does a product knowledge graph get?
Smaller than people expect. A team with 85 accounts, 1,200 tickets, 40 interviews and a year of decisions has a few thousand nodes and perhaps ten thousand edges. That is a small dataset for Postgres. Size matters less than discipline: typed edges, source links on every evidence node, and automatic ingestion so the graph stays current.
Can I build a product knowledge graph in Notion or a spreadsheet?
You can approximate one with relation properties or ID columns, and it is a good way to learn which links matter. The limits show up with multi-hop questions and upkeep: relations are maintained by hand, and you can't easily query several hops at once. Start there if it helps, but plan for automatic ingestion once the team grows.
GDPR-aligned. EU workspace data is stored in Frankfurt, Germany. GDPR and EU data Built in Vienna, Austria. A product of influence solutions fm GmbH (FN 675374y). Imprint 14-day trial. Card required, and nothing is charged until day 14. 30-day refund after subscribing. See pricing