What is a directed acyclic graph (DAG)?

In a recipe, each step depends on the one before it: you can't frost a cake before you bake it, and you can't bake it before you mix the batter. The process flows in one direction, nothing loops back, and the relationship between each step is clear and ordered. A directed acyclic graph works on the same principle; it maps a system of relationships where influence flows from one node to the next, dependencies are explicit, and you can never follow the edges back to where you started.

That kind of structure turns out to be incredibly useful across data science, data flow management, and machine learning. But like any tool, a directed acyclic graph has a specific job. Understanding what that job is—and what it isn't—matters a great deal when you're trying to measure something as dynamic as marketing performance.

Key takeaways

• A directed acyclic graph (DAG) is a data structure made up of nodes (vertices) and directed edges, where connections move in a defined direction and never form cycles or loops back to any previous node.
• DAGs were popularized in causal statistics by computer scientist Judea Pearl as a way to formally represent cause-and-effect relationships between variables.
• In data science and machine learning, directed acyclic graphs are commonly used for task scheduling, data pipelines, topological sorting, and modeling dependencies between variables.
• DAGs work well for representing clear, sequenced relationships, but they're not built for the feedback loops, time-varying effects, and overlapping influences that define real marketing systems.
• Because a directed acyclic graph either captures a single point in time or requires time to be encoded into the data itself, it can't track how marketing effects build, compound, and decay across days and weeks.
• Time series analysis is better suited for marketing measurement because it tracks how variables change over time and how earlier spend shapes later outcomes.
• Marketing mix models (MMMs) built on time series analysis capture the full picture of how campaigns perform, including delayed effects, cross-channel interactions, and shifting efficiency across seasons.

What is a directed acyclic graph?

A directed acyclic graph (DAG) is a data structure made up of two components: nodes and edges. Nodes—also called vertices—are the individual units in the graph. Each node represents a variable, a step, or an entity you're tracking. Edges are the connections between nodes, and in a directed acyclic graph, every edge is directional (it's an arrow pointing from one vertex to another, showing which node influences which).

The "acyclic" part matters just as much as the "directed" part. It means the acyclic graph has no closed loops anywhere in its structure. If you follow the directed edges forward from any given vertex, you'll never find yourself back at the starting vertex. Every path moves forward through the graph, toward the next node in the sequence, never circling back.

Two defining rules hold for every directed acyclic graph: edges point in a defined direction, and no sequence of connected edges forms a cycle. These constraints are what make DAGs so useful for representing dependencies: when node A must influence node B before node B can influence node C, a directed acyclic graph makes those relationships explicit and easy to reason about.

The concept of acyclic graphs comes from graph theory, but directed acyclic graphs were brought into widespread use in statistics by computer scientist Judea Pearl. Pearl used DAGs as the foundation for causal inference, recognizing that if you could represent the causal relationships between variables as a directed graph, you could reason formally about what causes what, not just what correlates with what. His framework gave data scientists a rigorous way to analyze causal relationships at scale, and the directed acyclic graph became a central tool in that work.

How a DAG works (nodes, edges, and directed paths)

It helps to see how the structure of a directed acyclic graph actually functions before getting into where it's used. Every DAG is made up of nodes (vertices) connected by directed edges.

Each node or vertex represents something in your system, such as a variable, a task, a data point in a pipeline, or a step in a process. Each edge is a directed connection between two nodes, with an arrow pointing from the start node to the end node. If an edge runs from node A to node B, it means A influences, precedes, or feeds data into B. A directed path is any sequence of connected edges that flows from one vertex to another without revisiting any node.

The acyclic constraint means that following any directed path through the graph will never bring you back to your starting node. There are no circular dependencies and no vertex that eventually points back to itself through a chain of edges. This property—the absence of any cycle—enables what's called topological ordering: you can arrange all the nodes in a linear sequence such that every edge in the graph points from an earlier node to a later one. Topological sorting takes advantage of this property to determine the valid order in which to process nodes across a system of dependencies.

A family tree is a useful conceptual representation: parents always appear above children, edges represent relationships flowing in one direction, and no child can also be their own ancestor. No cycles, no loops, clear direction. That's the essence of a directed acyclic graph.

How directed acyclic graphs are used in data science and machine learning

Directed acyclic graphs show up throughout data science and machine learning because so many technical problems are fundamentally about dependencies, sequences, and directed data flow. If one task can't run until another task completes, that's a directed relationship between two nodes. If data enters a pipeline at a given vertex and gets transformed before passing to the next step, those nodes are connected by directed edges. DAGs give data engineers and scientists a way to make those dependencies explicit and enforce the correct order of operations.

Task scheduling is one of the most common applications of the acyclic graph structure. Orchestration tools like Apache Airflow use DAG logic to manage complex data workflows: each task is a node in the graph, and the directed edges define which tasks involved in the pipeline must complete before the next tasks can begin. The acyclic constraint ensures there are no cycles or circular dependencies that would prevent the entire workflow from completing. Topological sorting is how these tools determine the right sequence for task execution, working through the directed acyclic graph node by node in a valid linear order.

Data pipelines are another major use case. In a data processing workflow, data enters at one vertex and flows through a sequence of transformations before reaching its destination. Each step depends on the output of the previous one, and the directed structure of the acyclic graph makes the full data flow visible: which nodes feed which, where data enters, and how it moves through the entire process in a defined sequence.

In machine learning and causal inference, directed acyclic graphs are used to map out causal relationships between variables. A DAG can represent that variable A influences variable B, and that variable B influences variable C, without implying any reverse relationship. That directed structure is useful for identifying confounders—variables that influence both an apparent cause and its effect—and for reasoning about what would happen if you intervened on a specific node in the graph. (You can deep dive into this with our guide to confounding variables in marketing.)

Directed acyclic graphs also appear in graph theory applications like transitive closure analysis, which tracks all the nodes that can be reached from a given vertex by following directed paths. In distributed ledger technology and blockchain systems, DAG-based structures manage transactions by representing each transaction as a node, with directed edges capturing which transactions must be confirmed before others can be processed. Transitive closure—determining all nodes reachable from a given vertex by following directed paths—is one of the graph operations that supports this. Transitive closure is also widely used in graph theory to map reachability across a directed acyclic graph.

Where DAGs show up in marketing measurement

Marketing has directional relationships. Awareness campaigns drive branded search volume. Branded search drives direct traffic. Direct traffic drives conversions. In that sense, the marketing funnel has a structure that resembles a directed acyclic graph: a set of nodes connected by directed edges, with upper-funnel activity flowing toward lower-funnel outcomes through a sequence of causal relationships.

This is part of why directed acyclic graphs come up in conversations about marketing mix models and measurement methodology. Building a model that accurately represents how your marketing works requires understanding the structure of those relationships: which channels influence which outcomes, in what direction, and with what kind of dependency between nodes. Causal graphs—the visual frameworks that map how marketing variables relate to revenue—are closely connected to the concept of a directed graph, and the directed acyclic graph is a specific and well-defined version of that concept.

But this is also where the limitations of the DAG data structure start to matter for marketing measurement.

Why directed acyclic graphs alone aren't enough for marketing

Marketing doesn't operate at a single point in time, and the causal relationships between marketing variables don't stay fixed. That's the core problem with applying a directed acyclic graph directly to marketing measurement.

Directed acyclic graphs handle time in two ways:

• they either represent a single point in time, or
• they require time to be encoded into the data itself

Neither of those approaches works for the kind of continuous measurement that marketers need.

When you run a connected TV campaign in week one, the effect doesn't stop at the end of that week. It drives branded search in week two, influences organic traffic in week three, and contributes to a conversion in week four. That arc can't be captured in a static acyclic graph, and encoding every time step as a separate node quickly becomes unmanageable, while still not capturing how the efficiency of each relationship shifts as conditions change.

There are also feedback loops that a directed acyclic graph simply can't represent. More revenue often means more budget available for marketing, which drives more revenue. That kind of circular dependency violates the acyclic constraint by definition. The graph has no mechanism to represent these kinds of cycles without breaking its own rules, and marketing systems are full of them.

The relationships between marketing variables are also dynamic in ways a static directed graph can't capture. Channel efficiency shifts with seasonality. Upper-funnel spend changes how effective lower-funnel channels are, and that effect varies depending on how long awareness campaigns have been running, what the competitive environment looks like, and what time of year it is. Those aren't fixed directed relationships that can be drawn as a stable acyclic graph once and left alone. They're evolving dynamics that require a modeling approach built to track change as it actually happens.

These are the reasons why time series analysis—not directed acyclic graphs—is the right foundation for marketing measurement.

Why time series analysis is the right approach for marketing

Time series analysis is built to track how variables change over time, looking at sequences of data points across days, weeks, and months and modeling how earlier values influence later outcomes. For marketing, that distinction is fundamental. Spend today doesn't only drive revenue today, it creates an effect that carries forward across subsequent time periods. A time series model captures that kind of memory naturally because it tracks how a variable at one point shapes what happens next, including the delayed effects of an awareness campaign, compounding brand investment, and the decay of ad effects over time.

Time series analysis can also handle the non-linear, shifting relationships that define real marketing data. Efficiency changes with seasons. Upper-funnel activity in one period shapes the effectiveness of lower-funnel channels in the next. A channel that drives strong results during a peak season might look very different in a slower period. These aren't fixed directional relationships you can encode once in a directed acyclic graph, they're dynamic patterns that need a modeling approach designed to move with the data rather than freeze it.

That's why Prescient uses time series analysis as the foundation of its marketing mix model rather than DAGs. Directed acyclic graphs are powerful tools for the problems they're designed to solve. The issue is that marketing measurement requires something built for time, not just for direction.

Where Prescient comes in

Prescient's marketing mix model is built on time series analysis and designed to capture the full complexity of how your marketing actually works, including the delayed effects, cross-channel interactions, and shifting efficiency that simpler measurement approaches miss. Rather than treating campaigns as isolated inputs with static outputs, Prescient tracks how spend on any given channel ripples across your entire marketing ecosystem over time, including effects on branded search, organic traffic, direct traffic, and Amazon purchases that standard attribution tools never account for.

If you want to understand how your marketing budget is actually performing and where to put the next dollar, we'd love to show you how the Prescient platform is the best tool for the job. See it in action when you book a demo.