Marketing and data engineering use the same words for different things. This collects every one I have run into, what each side usually means by it, and which guide goes into it properly. Search it, link to a single definition, or send someone the line that settles the argument.

Most of these came out of a six-part field guide sponsored by Databricks. They funded the work and reviewed it for factual accuracy, and had no say over any of the definitions. This page is not sponsored.

37 terms

Activation

You mean
Launching a campaign, or putting an audience to use.
They mean
Delivering eligible records to a destination with the right identifiers, permissions, freshness and format.
From Guide 1

Agentic

You mean
Agents that draft audiences, journeys and briefs inside your CDP or email platform, often with a person approving the launch.
They mean
An AI model that picks its own next step, uses tools, checks the result and repeats until a goal is met.
Needs first
Written goals, agreed tool access, and a decision on which actions need sign-off.
From Hello, I'm human

AI-native

You mean
Used for new products and for older ones with AI added, so the label alone tells you little.
They mean
A product designed around a model from day one. Take the AI out and the product stops making sense.
Needs first
Willingness to change the process along with the software.
From Hello, I'm human

Audience

You mean
Interchangeable with segment.
They mean
Often the exported copy of a segment, the version pushed to the destination, not the logic behind it.
From Guide 1

Batch

You mean
The overnight update, or however the data gets refreshed.
They mean
A scheduled job with a window, a rerun path, and somebody who looks at it in the morning before anything downstream consumes it.
From Guide 3

Brain

You mean
Context brain, marketing brain, company brain, brand brain. Holds brand rules, approved claims, legal limits and what worked. When it "learns", it saves corrections. The AI itself is not retrained.
They mean
A store of documents, rules and past decisions that an AI looks up before it acts. Borrowed from the "second brain" note-taking idea.
Needs first
Someone who curates what goes in and removes what is stale, and reviewers whose reasons get recorded.
From Hello, I'm human

Computed trait

You mean
A useful number on a person, like lifetime value.
They mean
Something recalculated on a schedule, at a cost, for every person, forever.
From Guide 1

Context engineering

You mean
Treated as a new discipline and role: getting the right data, content and permissions to each agent at the right moment.
They mean
Deciding everything an AI sees each time it acts: instructions, data, examples, tools and memory. Prompt engineering covered only the instructions.
Needs first
Business rules written down, and data clean and labelled enough to hand to a machine.
From Hello, I'm human

Context layer

You mean
CDPs use it for customer profiles, events and consent served to agents. Others use it for metric definitions and policies.
They mean
No settled meaning. Usually the service that supplies an AI with facts and rules at the moment it acts.
Needs first
One agreed source for each customer fact, with consent attached.
From Hello, I'm human

Copy

You mean
Downloading or exporting the data somewhere.
They mean
Any additional persisted version. Replication, materialised tables, caches, extracts, destination-side storage.
From Guide 1

Correction window

You mean
Not a phrase you have used.
They mean
How long there is to catch a mistake before a customer sees it. It shrinks to nothing as the latency does, and it is the real cost of speed.
From Guide 3

Customer

You mean
A person.
They mean
An entity at a grain. Person, account, household or device, and which identifiers represent it.
From Guide 1

Data source

You mean
The tool or application where you see the data.
They mean
The system that produced it. Which system, how it arrived, and which version counts.
From Guide 1

Deduplication

You mean
Removing the doubles.
They mean
Identity resolution. A policy with thresholds. Too keen merges two people, too cautious splits one.
From Guide 1

Event

You mean
Something a customer did that you can target on.
They mean
A timestamped record with a schema, written once. The schema can change, and that affects ingestion and everything downstream.
From Guide 1

Field

You mean
A box in the segment builder.
They mean
A column with a type (what values are allowed), a nullability rule (can it be empty), and an owner.
From Guide 1

Forward deployed engineer

You mean
A vendor's engineer inside your team who sets up the product and reworks workflows around it. Part builder, part consultant.
They mean
An engineer embedded with a single customer, adapting the platform to their problem. Now a common role at AI labs.
Needs first
Internal owners who learn alongside them, so the setup survives their exit.
From Hello, I'm human

Freshness

You mean
How recent the data is.
They mean
The age of the data at the moment it is read, which is a different number from how often the job runs and usually a worse one.
From Guide 3

Golden record

You mean
The truth.
They mean
The output of rules a person wrote, which can be changed.
From Guide 1

Governance

You mean
The reason things take longer.
They mean
The reason the company has not had to disclose a breach.
From Guide 1

Harness

You mean
Widened to mean a coordination layer across a team's tools, with shared business context and guardrails. Engineers would call that orchestration.
They mean
All the code around a model that turns it into an agent: the loop, the tools, what it sees and what it may do.
Needs first
Permissions for each tool, a log of what ran, and an owner for the rules.
From Hello, I'm human

In-session state

You mean
The site knows what I just did.
They mean
Data served fast enough to answer in the time a page has to render, often from a separate serving layer kept in step with the warehouse.
From Guide 3

Integration

You mean
Two tools are connected.
They mean
A direction of travel with auth, field mapping, schema handling, rate limits, retries, monitoring and an owner.
From Guide 1

Latency

You mean
The delay before you see something.
They mean
Several delays that add up, measured at points you have not specified.
From Guide 1

Live

You mean
It is working now.
They mean
It is in production rather than in a test environment.
From Guide 1

Lookalike

You mean
Expanded reach, built from a seed audience.
They mean
A downstream black box they cannot inspect, explain or debug. Expect hesitation about depending on it.
From Guide 1

Micro-batch

You mean
Near enough to live.
They mean
The same kind of job, run every few minutes. Depending on the setup, you can end up paying for the schedule rather than the volume.
From Guide 3

Model

You mean
A prediction, like propensity to churn.
They mean
The data model. Tables and how they relate. The worst offender on this page.
From Guide 1

Profile

You mean
Everything known about someone, in one view.
They mean
A join, several tables stitched on a shared key, computed at a moment and already slightly stale.
From Guide 1

Raw data

You mean
An unformatted export.
They mean
The untouched landing layer, preserved so downstream can be rebuilt.
From Guide 1

Real time

You mean
Soon enough that it does not feel broken.
They mean
Sub-second streaming, an always-on architecture, and a cost line that never stops.
From Guide 1

Refresh

You mean
The numbers change.
They mean
A pipeline run with dependencies, a schedule, and someone on call for it.
From Guide 1

Segment

You mean
A group of people to message.
They mean
Needs to know if you mean the definition or the result. Those cost differently.
From Guide 1

Source of truth

You mean
The dashboard you trust.
They mean
The designated system, governed dataset or semantic definition. There can legitimately be different ones per subject.
From Guide 1

Streaming

You mean
Instant.
They mean
A pipeline that never stops, where a failure becomes an incident rather than a quiet rerun, and on always-on infrastructure the cost runs whether anything happened or not.
From Guide 3

Sync

You mean
The data shows up in the tool.
They mean
A job with failure modes, retries, and last night's failed records. Scheduled, change capture or event-driven.
From Guide 1