What Is a Marketing Context Graph? A Full Guide

Marketing has two problems that never show up on the same dashboard, but come from the same root cause.

Problem one: channels don’t work alone. A TV campaign makes people search for your brand. Upper-funnel ads make retargeting cheaper. Most measurement tools can’t see these connections. So a channel’s reported return can swing wildly from quarter to quarter, even when nothing about the business has actually changed.

Problem two: budget decisions lose their reasoning the moment they’re made. Most systems record what ran and how much it cost. They don’t record why. What evidence backed the call. Which option got turned down. Which policy was in effect. So when a CFO asks why paid search spend jumped 28% last quarter, nobody, not even the AI agents now helping make these calls, can explain it.

Lifesight built the Marketing Context Graph (MCG) to fix both problems at once. It’s a governed, time-aware map that shows how channels affect each other and keeps a permanent record of every decision and the evidence behind it. In controlled testing, it cut budget-decision error by up to 43% and correctly answered 100% of “why” questions, versus 25% for a standard dashboard.

For the complete methodology, assumptions, and simulation results, explore Anil Singh’s Marketing Context Graph research paper.

Here’s how it works, and why these two problems turn out to be one problem:

Why Standard Measurement Keeps Breaking Down

Most marketers size their budgets using media-mix modeling (MMM), a method that works from aggregate spend and outcomes rather than individual user tracking. MMM itself isn’t the problem. The problem is what happens once the model produces a number.

Standard MMM output gets read as if each channel operates on its own. A spreadsheet of separate spend columns can’t capture how channels actually interact, and that gap shows up as two deeper problems:

1. Channels aren’t independent, and flat models can’t see the connections.

A channel’s real value is its direct effect on sales, plus every indirect effect it creates through other channels, known as the channel’s halo effect. Worse, a common MMM shortcut, controlling for the mid-funnel channel a campaign influenced, accidentally wipes out that indirect effect and gives credit to the wrong channel. That’s usually why the numbers swing between model runs.

2. Decisions lose their paper trail.

Most systems log the what, not the why. As more decisions get handed to AI agents, that gap becomes a real problem. Regular agent memory is built to recall facts fast. It’s not built to reconstruct the reasoning behind a decision.

Both problems come from the same place: 

Measurement systems that throw away structure.
A marketing context graph puts that structure back.

What a Marketing Context Graph Actually Is

At its core, a marketing context graph is a map of connected things: channels, spend, outcomes, decisions, policies, and evidence. These are linked by clear relationships, like influences, decays into, is supported by, and was decided under. Two things make this useful for measurement specifically.

1. It tracks cause and effect, not just correlation.

A normal knowledge graph just says two things are related. The MCG records the direction and strength of a cause-and-effect relationship, and where that measurement came from, usually a real experiment like a geo-lift test. So the correlation, TV drives branded search, comes with a measured strength, a time window during which it’s considered valid, and a link back to the experiment that proved it.

2. It remembers two different timelines.

Every fact in the graph has two timestamps: when it was true in the real world, and when the team actually found out. This matters more than it sounds. If a model gets updated today with a corrected number, and that correction quietly rewrites last year’s results too, the backtest ends up looking better than it should. The graph avoids this by reconstructing exactly what was known at any point in the past, not what’s known now.

Worth noting: the graph itself isn’t a new statistics method. The actual math, path analysis, Bayesian updating, transporting results across regions, is standard, well-proven causal inference. The graph’s job is to govern the inputs to that math: which relationships to trust, how strong they are, when they were learned, and when they expire.

Measuring the Halo Effect Between Channels

Take TV and branded search. If a model tries to work out both effects at once using only spend data, they get tangled up, because TV spend and the search volume it causes tend to move together.

The fix is to measure the connection directly, with a geo-lift experiment: run TV ads in some regions but not others, and see how much branded search moves. That measured halo effect gets fed into the model as a trusted starting point, so the model only has to figure out the harder part from the data. This is a standard statistical technique. A well-measured starting point can only sharpen a model’s accuracy, never hurt it, and it helps most exactly where the model was most confused.

One important caveat: this only works if the experiment itself was done properly, randomized correctly, run long enough, and free of hidden factors linking TV and sales. The graph makes that assumption easy to check. It can’t fix a badly run experiment. A biased test still gives a wrong answer, just a more confident-sounding one.

Why a Governed Graph Beats a One-Time Assumption

A single experiment is really just an educated guess. A governed graph gets more valuable as evidence builds up over time, in two specific ways.

1. Honest backtesting.

Say a channel’s true return was 2.0 ROAS the whole time, but the team believed it was 3.4 ROAS until a corrective experiment in quarter five set things straight. Now ask: how accurate were our past estimates? A system with no memory of, “what we knew when,” will quietly apply today’s corrected number to every past quarter, and report 0% historical error. That’s not honest, it’s just forgetting. The marketing context graph instead rebuilds the belief as it stood at each point in time, and reports the real number: in Lifesight’s testing, 44% historical error, which hid a 70% overspend on that channel across the five quarters before the fix.

2. Choosing which evidence to trust.

Real measurement programs build up experiments that sometimes disagree, for example, when a platform changes its algorithm and resets a channel’s true effect. A single hard-coded number can’t adjust for that. The graph automatically retires outdated evidence and blends what’s still valid, weighted by how precise each result is.

How you pick which evidence to trust	Error (RMSE)
Hard-coded number, never updated	1.40
Blend every experiment equally, even outdated ones	0.48
Only trust the newest experiment	0.30
Graph: retire old evidence, blend the rest by precision	0.23

True effect = 2.0. Based on 2,000 simulated trials. Lower is better.

Handling Conversions You Can’t See

Not every conversion can be tracked, and the ones you lose aren’t random. They tend to skew toward certain devices, logged-out users, or higher-value customers. Ignoring that skew biases any model trained on what’s left. The standard fix is to give more weight to the conversions you can see, based on how likely they were to be visible in the first place.

The marketing context graph’s job here is to feed and manage that fix: a step-by-step matching process (a direct match, a probable match, a regional estimate, a fallback number) that gives every conversion a documented path, so the correction stays consistent as tracking gets harder over time.

In testing, as untracked, non-random signal loss went from none to a severe 80%, a standard model’s decision error rose from 4% to 33%. The graph-governed approach held close to 4% the whole time, an 89% reduction in the damage caused by lost tracking.

Decision Provenance: Answering, “Why Was This Decision Made?”

This is the second problem from the start of this post: figuring out the reasoning behind a past decision. The marketing context graph treats every budget decision as its own object, connected to the policy that caused it, the evidence that supported it, the options that were turned down, and who approved it.

Here’s a simple example. A coffee brand shifts 20% of its Meta budget to TikTok, based on a past TikTok test, an MMM read, a geo-lift experiment, and a tighter budget policy. Lifesight tested a standard fact-recall system, the kind behind most dashboards and search tools, against the MCG, using a set of questions with known correct answers.

Question type	Standard system	MCG
Simple fact lookup	100%	100%
Multi-step “why” questions	0%	100%
Policy compliance	0%	100%
What alternative was rejected, and why	0%	100%
Which policy applied at the time	0%	100%
Overall	25%	100%

A standard system handles simple lookups fine. Anything that needs to connect a decision to its policy, its evidence, and the options it rejected is invisible to it. The graph answers all of it, because one path from the decision reaches everything connected to it.

The same experiment that sharpens a measurement is also the evidence that justifies the decision it informs. Measurement and decision tracking turn out to be two sides of the same graph, not two separate systems.

Taking a Result From One Region to Another

A geo-lift test usually runs in a specific set of regions with a specific audience. But the decision it’s meant to inform is often for somewhere else entirely. Copying the result over without checking if the two places are actually similar is a common shortcut, and a risky one.

The graph borrows a well-established idea from causal inference called transportability: note exactly how the source region and the target region differ, then check whether the result can safely carry across that difference before using it. If it can, the graph carries it over. If it can’t, it says so instead of guessing.

How different are the regions	Copy-paste	Fresh estimate, adjusted	Governed transport
Identical	0.03	0.22	0.03
Somewhat different	0.50	0.22	0.03
Very different	1.00	0.22	0.04
Highly different	1.50	0.23	0.05
Extremely different	2.00	0.23	0.05

Error (RMSE) in the estimated effect. Lower is better. 2,000 simulated trials per setting.

Just copying the number over gets worse and worse as the regions become more different. Running a fresh estimate in the new region and adjusting for audience differences avoids bias, but it’s noisy, since the new region doesn’t have much of its own test data yet. Governed transport, which borrows the cleaner signal from the original experiment and reweights it, stays accurate the whole way through, about 4 to 6 times more accurate than a fresh, adjusted estimate on its own

This advantage isn’t permanent. Once a region has run enough of its own solid experiments, the two approaches end up in about the same place, and transport stops adding much. It’s most useful for the common case: a region that hasn’t run a full experiment of its own yet.

The graph also runs a quick check before transporting a result: a small, low-cost test in the new region, just to confirm the original result still holds. As unexpected differences between regions grow, the graph gets more likely to simply say, “I don’t know,” instead of guessing, going from a 4% pass rate on genuinely similar regions up to declining every time once the difference is too large to ignore. A simple copy-paste approach, by contrast, confidently answers every case with about the same amount of error, whether the transfer was ever valid or not.

Results at a Glance

One pattern shows up again and again: when a marketing system doesn’t have much underlying structure to work with, the graph and a standard flat model perform about the same. The graph’s advantage grows specifically as real-world complexity grows. These numbers come from Lifesight’s simulation of 8,000 budget decisions, run 800 times each.

• Budget decisions: using the graph to trace indirect effects cuts decision error by up to 43% as channels become more interconnected, and makes no difference when they aren’t, exactly as it should.
• Backtesting: a standard system can hide 44 percentage points of real historical error by quietly rewriting the past. The graph reports it honestly.
• Decision tracking: the graph correctly answers 100% of why questions on a realistic test, versus 25% for a standard system.
• Cross-region transport: governed transport stays 4 to 6 times more accurate than a fresh, region-only estimate, and it knows when to say no instead of guessing.

Honest Limits

A marketing context graph is only as good as the experiments feeding it. A weak or biased test still produces a confident, wrong starting point. The graph makes that assumption visible and easy to check, but it can’t fix a badly designed experiment.

Transporting results depends on marking regional differences properly up front. And its advantage fades once a target region builds up its own solid test data.

There’s also a real setup cost. Building the graph means mapping out channels, policies, and relationships before you get any value from it. That’s a real investment, and it’s worth planning for.

All the results here come from controlled simulations with a known correct answer, on purpose, so error could be measured precisely. The next step is testing the same approach on real campaigns and real decisions.

The Bigger Picture

Channels that interact with each other, and decisions that lose their reasoning, look like two separate headaches. They actually share one root cause: measurement and decision systems that throw away causal, time-aware structure. A marketing context graph puts that structure back and manages it, while leaving the actual statistics, path analysis, Bayesian updating, transport, to well-established causal inference methods that are already trusted.

The result is measurement that gets more accurate as the real marketing system gets more complicated. Backtests that tell the truth about what was known and when. And a system that can explain why a decision was made, not just what happened.

This is the thinking behind Lifesight’s Agentic Unified Marketing Measurement Platform. Most vendors give their agents a dashboard to read. Lifesight gives its agents a causal engine to think with, the same governed structure described here, built into the platform that 300+ brands rely on for measurement they can trust and explain. Book a Demo