This is the first in a three-part series on what AI needs to analyze marketing data reliably. Reliable marketing AI rests on providing three things:
- Knowledge of your data: AI needs a dynamic map of what your data means and how it should be used.
- Knowledge of your business: AI needs a comprehensive context on what’s happening in your business and industry to accurately interpret the data.
- The ability to act: AI needs the tools to execute the query and prove how it got there.
This first article focuses on laying out the knowledge of marketing data that AI tools need in order to eliminate hallucinations, and ensure a question gets the same answer every time.
Conversations about AI in marketing analytics usually start with capability. What can it connect to? What analysis can it run? Can it generate reports, explain performance, or act autonomously? But all of these questions must come after a more basic question. First we need to ask whether the AI has a complete understanding of marketing data.
Marketing metrics are already complex, and that complexity is only amplified when they’re used across platforms, teams, regions, and reporting systems. Ambiguity has been the root problem for many reporting issues, but when AI is introduced into the equation the risk grows because the system can act on those definitions faster, across more queries, reports, and workflows.
The problem with trying to prompt around AI’s missing knowledge of marketing data is that it assumes every marketer already knows the correct definitions well enough to write them into a prompt, which defeats the purpose of making data self-serve in the first place.
Adding prompts through an LLM’s memory runs into a similar problem. It’s true that you can get further by putting shared definitions into a project's memory. However, beyond the time it takes to set up and maintain, there are limits to what memory can actually do.
LLM memory is built to store natural language. You can write "exclude agency fees from cost," but there's no way to attach an actual calculation, transformation, or field-resolution step to it. The LLM reads the instruction and reasons through it in natural language each time, rather than running a fixed formula. That means the same instruction can produce different answers on different days.
For AI to be reliable, it can’t depend on a static data map, as this leaves ambiguity which results in hallucinations. Marketing data doesn't stay still, so AI needs access to a live map of your data which shifts to accurately reflect this.
Here are a few of the changes that a static semantic layer doesn’t account for:
- Platforms rename or restructure fields. GA4's rename of "conversions" to "key events" in March 2024 changed the underlying API field too, so a query still pointing at the old name stopped returning data.
- A platform redefines a metric without renaming it. For example, the calculation method changes or an attribution window shifts, and the field is now returning different data under the same name.
- Definitions shift internally. A team might stop including agency fees in CAC, and a quarter-over-quarter comparison six months later ends up comparing two different formulas.
- Metrics get retired with nothing to replace them, so reports built on them return zero, and zero can pass for a real result if nobody's watching for it.
An intelligent semantic layer (like the one Adverity Atlas uses) creates a dynamic map of your data to ensure AI uses the right fields, definitions, and logic at query time. Definitions are discovered automatically by the intelligent semantic layer, then confirmed by your team, and applied consistently everywhere they're used. Atlas continuously monitors for drift (whether that’s due to a change from the data source or from within the business), so any changes are flagged for approval rather than slipping through unnoticed.
Instead of asking each user to explain what "cost" means, which currency conversion to use, or how CAC should be calculated whenever a new investigation starts, the agreed method is governed centrally and applied automatically. This distinction is especially important for enterprise teams because AI reliability depends on consistency. If one user calculates CAC using media spend in USD, another uses local currency with agency fees included, and a third leaves the AI to infer the method from the schema, the organization gets three different answers that all look credible. Shared metric definitions remove that ambiguity before analysis starts by giving AI one governed calculation method to follow.
AI can collapse time to insights, but only if every answer starts from the same understanding of the data involved. That means knowing what metrics mean, how metric calculations work across different platforms, and which rules govern how data can be used - the list is long. For AI to run accurately, it needs a comprehensive understanding of everything listed below.
The governed definitions AI needs before it calculates anything.
Canonical marketing concepts
Shared concepts such as cost, impressions, clicks, conversions, revenue, ROAS, CAC, CTR, CPM, pipeline, and attribution.
Governed metric definitions
The agreed meaning of each KPI, including what is included, excluded, transformed, and used for reporting.
Calculation logic
The formulas behind metrics such as ROAS, CAC, CTR, CPM, conversion rate, blended performance, or pipeline efficiency.
Attribution definitions
The agreed logic for conversions, revenue, touchpoints, lookback windows, and channel contribution.
How platform-specific data maps into common marketing concepts.
Cross-platform field mappings
How different platforms represent the same concept, such as Google Ads cost_micros, Meta spend, TikTok cost, or LinkedIn costInLocalCurrency.
Field transformations
Required conversions such as micros to currency units, percentages to decimals, or local currency to reporting currency.
Platform-specific marketing logic
How each platform structures, labels, reports, models, or attributes performance.
The usable map of the data environment: what exists, where it lives, and how it can be used.
Available sources
Which connected platforms, warehouse tables, files, and business systems exist.
Schema structure and field relationships
Which tables, fields, data types, IDs, dates, dimensions, and metrics are available.
Cross-system relationships
How data connects across ad platforms, CRM, ecommerce, analytics, finance, and planning sources.
Join paths and reliability
Which IDs, fields, naming conventions, or inferred relationships can connect datasets, and where joins may create duplication, gaps, or misleading matches.
Data grain
Whether data is stored at event, daily, weekly, monthly, campaign, customer, market, or channel level.
The rules that stop AI from using the right-looking data in the wrong way.
Source-of-truth guidance
Which system should be used for each metric or use case, such as platform data for spend, CRM for qualified leads, ecommerce for transactions, or finance for booked revenue.
Aggregation rules
How metrics can be safely rolled up, combined, or compared.
Data freshness and coverage
Which sources are current, which periods or markets are complete, and where data lags.
Calendars, currencies, and time zones
Which reporting calendar, fiscal period, currency, and time zone should apply.
Current schema drift
How API updates, renamed fields, connector changes, attribution shifts, and warehouse model updates affect the data environment.
A simple prototype can make this look easier than it is. It’s possible to connect an AI tool to a warehouse, map a handful of familiar fields, and get plausible answers from well-known tables.
We’ve already established the complexity of maintaining reliable metric understanding across changing data sources, platform updates, renamed columns, and evolving business definitions. But don’t underestimate the foundational build either. Building out sound logic across data sources is complex, time-consuming, and tedious even before maintenance enters the picture.
Marketing data creates specific challenges:
- Metrics often need transformations before they can be compared.
- Relevant data often sits across systems that do not share clean keys.
- Sources operate at different levels of detail, making joins and aggregation risky.
- Data quality and coverage vary by source, market, and period.
- Campaign names often encode brand, market, objective, audience, and test structure in their naming conventions.
- Critical business rules often live in people’s heads rather than in a governed system.
A single performance question often spans multiple platforms. An intelligent semantic layer helps AI understand how those pieces fit together, and what limitations apply to the answer it is about to produce before it explains the result. That includes the relevant sources, tables, fields, grains, join paths, freshness, coverage, and quality caveats that shape the analysis.
A standard semantic layer defines metric meaning inside a data model someone has already built and maintained by hand. Atlas's intelligent semantic layer does this too, but it also:
- Automatically discovers schemas and field relationships
- Resolves marketing-specific meaning across platforms
- Builds the live map of how systems connect
- Keeps all of it current as the environment changes, so teams are not stuck manually configuring and maintaining it forever
In Atlas, the pre-encoded marketing intelligence makes AI more reliable before any investigation runs. For example, Atlas includes more than 40 canonical marketing concepts, with field mappings across platforms. A concept like cost can be resolved across systems before an investigation starts, rather than leaving the AI to infer the right field from column names alone.
It also captures marketing knowledge that is specific to how teams organize their reporting. Naming conventions, preferred definitions and reporting structures are all discovered by Atlas and provide your AI with a stronger understanding of how performance should be compared and interpreted.
And because definitions are consistent, the semantic layer improves over time. When a team sets a definition once, that definition becomes part of the foundation for every subsequent investigation.
Built on more than a decade of enterprise deployments representing over $80 billion in managed advertising spend, Atlas embeds the marketing knowledge organizations need to move AI from experimentation to production.
The promise of AI in marketing analytics is speed. But speed alone does not solve the reporting problem if every answer still needs to be checked manually.
A semantic layer built once and left alone starts drifting and breaking from the moment it's approved. Marketing data doesn’t hold still. In order to reflect a shifting reality, your marketing AI needs a semantic layer that can keep up, so every user gets the same answer from the same agreed definition, with no gap left open for the AI to fill in on its own.