Your position rose 6 places this quarter. Was it retail media, the packaging change, the 12.12 promo, the reviews that finally crossed 200, or the fact that your strongest competitor went out of stock for two weeks? Spoiler: it was probably a mix. The operational question isn't "which one was it" but "how much weight did each one carry?".

In the previous two articles we built the thermometer (weighted organic position) and the matrix (organic crossed with paid SOV). Those two tell you whether you're winning or losing. But eventually someone — the CFO, the VP of Marketing, your agency — will ask the harder question: "If we're winning, what's the engine?". And it can't be answered well with just the thermometer and matrix. You need a detective.

This article is about the third layer: observational multivariate attribution. How to decompose movement in position into its likely causes, without needing a six-month formal geo experiment. It isn't pure causal attribution, but it's decision-useful — and that's what the business actually needs.

Perfect attribution doesn't exist (and why those who promise it are exaggerating)

Let's start by clearing up something many "attribution solution" vendors won't admit: pure causal attribution is a theoretically unsolvable problem in the context of operational retail media.

Why? Because causal attribution requires a parallel universe identical to yours, except without the ad you're measuring. Since that universe doesn't exist, every attribution method is an approximation — some better, some worse. The good news is some approximations are good enough for budget decisions.

The bad news is the market is full of products that promise "retail media attribution with causal precision" and that are really doing smart correlation and selling it as causality. The quick rule to spot them: if the dashboard gives you a number without confidence intervals, without caveats about confounders, and without specific instructions on validation experiments, it's exaggerating.

The real detective doesn't hand you "the culprit." It hands you a list of suspects ordered by probability, with the chain of evidence for each, and tells you "with the available information, the most likely suspect is X, but I don't rule out Y. To confirm we need to run this experiment." That's the honest stance. That's observational attribution done right.

The 6 variables that almost always matter

In retail media for CPG, six variables almost always have a measurable effect on a product's organic position. They're the model's inputs. If your data captures these six, you've already solved 80% of the problem:

  1. Your sponsored share-of-voice per keyword. How much of the paid space you're buying.
  2. Your relative price to category. Your price divided by the median or average of comparable SKUs. A product 30% above category average faces mechanical resistance from the ranker.
  3. Active promotions and depth. If you're running a 20% promo at the moment of the pick, your sales velocity accelerates and the ranker registers it. The flag (yes/no), duration, and depth all matter.
  4. Content score / listing completeness. Quality and quantity of images, descriptions, filled attributes, video presence. VTEX and Walmart explicitly compute a "listing quality score" that's a direct ranking input.
  5. Review velocity and average rating. How many new reviews you receive per week, what average rating they hold. Reviews are the ranker's social input.
  6. Competitor actions. Their paid SOV, their prices, their stockouts, their launches. Organic position is a relative game: your position rises when competitors fall, not just when you improve.

Other variables matter less (seasonality, retailer events, your own launches), but these six move the bulk. If you have them recorded week by week alongside your organic position, you have the model's raw material.

The observational model in plain language

Let's explain the model without using the words "regression" or "panel" until strictly necessary.

The detective's question is: "Given that your organic position for a keyword changes every week, which of the six variables moves alongside it, and how consistently?".

If your paid SOV rises and two weeks later your position rises, and that happens consistently across 80% of your keywords, there's good evidence that SOV is moving position. If your SOV rises and position doesn't move anywhere, there's good evidence that SOV isn't the engine — at least not in this category at this spend level.

If your price drops 5% and two weeks later position rises, and that happens consistently, evidence that price matters. If your price drops 5% and position doesn't respond, evidence that price isn't the engine.

This is done formally with what statistics calls panel regression with fixed effects per SKU and week. The name is scary but the intuition is what we just described: you're looking for which variables move alongside position, controlling for fixed SKU characteristics and for general market movements each week.

The model's output isn't "paid SOV moved your position 3 places this quarter." The output is "a 10-point increase in paid SOV is associated with a 1.4-place improvement in organic position, with a 0.8 to 2.0 confidence interval". It's a range, not a number. It's a coefficient, not a verdict. It's statistical honesty.

The endogeneity trap (in soccer language)

There's an important trap worth naming before going further: endogeneity. Sounds academic but the concept is simple.

The observational model assumes the variables moved "more or less independently" from each other. But in retail media that's false. You don't invest paid SOV at random — you invest more in the SKUs and weeks where you think it'll work, against the competitors that worry you most, in seasons that sell most. Spend is correlated with everything else.

Simple analogy: imagine you're measuring "does the coach motivate the team?" and you observe that when the coach yells louder, the team wins more. Naive conclusion: yelling works. But if the coach yells louder when he sees the team is losing, and the team sometimes turns the match around for its own reasons, the correlation between "yelling" and "winning" can be confusing cause and effect.

In retail media exactly the same thing happens: you invest heavily in weeks when your competitor is attacking, and sometimes you win because you invested and sometimes you win because the competitor deflated. The observational model can't cleanly distinguish between the two.

What do you do about it? Three honest answers:

  1. Report coefficients with that caveat explicit. "There's likely positive endogeneity bias, so the real coefficient is ≤ X."
  2. Cross enough SKUs and weeks for noise to average out. With 200 SKUs × 52 weeks × 5 retailers you have 52,000 observations — that absorbs a lot of idiosyncratic noise.
  3. Validate with point experiments. Once a year, geo holdout on your largest line to check that the directions of the observational model hold up under a controlled experiment.

Why "directionally positive" is decision-useful

There's a common critique of the observational model from the purist side: "your coefficient is biased, it isn't causal, it doesn't help me make decisions."

The answer is: it depends on what decision.

Decision 1: "Do I shut down the retail media line?" For this decision you need pure causal attribution — a geo experiment. The observational model isn't enough, because the decision is discrete (continue or stop) and costly to reverse.

Decision 2: "Should I raise the bid 20% on these 50 keywords?" For this decision, a biased but directionally positive coefficient is plenty. If the model says "raising SOV is associated with position improvement, in the range of 0.8-2.0 places per 10pp of SOV," the decision is reasonable: raise the bid on the priority keywords and measure. You don't need the exact coefficient to do it.

Decision 3: "Should I reallocate budget from retailer A to retailer B?" Here the observational model is exactly the right tool — it's built for incremental allocation decisions. Perfect cause-effect would be nice but isn't necessary.

Most retail media decisions are categories 2 and 3: incremental allocations week to week, quarter to quarter. For those decisions, "directionally positive with confidence interval" is the right currency. Perfect causal attribution is a year-end goal, not an operational one.

The phrase to remember: better than total darkness. It isn't surrender — it's the honest stance on what observational attribution can and cannot do.

When you DO need a real experiment

Three situations where the observational model isn't enough and you need a controlled experiment:

  1. When the decision is big and irreversible. Cutting an RMN's budget 50%, shutting down a line, changing agencies. Here the 4-8 weeks and costs of a geo holdout are worth it.
  2. When the model says something that goes against your strong intuition. If the model says "your Mercado Libre Ads spend isn't moving anything" and your gut says "impossible, it should be," the controlled experiment settles it. One of the two is wrong.
  3. When the spend is so large that the experiment cost is marginal. If you spend 20% of your revenue on retail media, spending the additional 1% to run an annual geo holdout is trivial. It's the system's calibration check.

For everything else, the observational model is the right tool.

How to run a geo holdout, in five lines: divide your geographic footprint into comparable market groups, keep normal spend in some and pause (or reduce) in others for 4-8 weeks, compare the sales trajectory between the groups. It's the methodology of Meta Conversion Lift and Google Ads Geo Experiments. It works cleanly for off-site retail media (programmatic display). For on-site sponsored search it's more complicated, requires retailer coordination, and many RMNs don't yet have the tools to support it well.

The IAB/MRC standard: what the industry is agreeing to report

In January 2024, IAB and the Media Rating Council (MRC) published the first Retail Media Measurement Guidelines — the industry's formal answer to the "every retailer is its own black box" problem. The companion Executive Playbook, developed with Boston Consulting Group, translates the definitions into actionable steps. IAB Europe published its version in April 2024, built on the US standard.

Three points in the standard that matter for the observational model:

What IAB/MRC doesn't solve: the structural conflict of interest. The retailer can be 100% IAB-compliant in how it counts impressions and attributes conversions, and still be reporting only what suits it to show. The standard aligns vocabulary; independent verification is still necessary.

How to feed the model in practice

The observational model lives or dies by the quality of its inputs. If you record three of the six variables instead of all six, the model is much weaker. If you record all six but only across 50 keywords and 3 SKUs, there isn't enough variation to estimate reliable coefficients.

ePerfectStore.com feeds the model with consistent scraping of the six variables at scale:

With those inputs, a brand can run the model internally with its analytics team, or use the pre-processed output ePerfectStore delivers as a dashboard. What matters isn't the specific algorithm — it's the consistent availability of data. Most brands don't have the model because they don't have the data, not because they don't have the analysts.

In summary

Question What applies
What's the model? Observational attribution: relating organic position to 6 variables week by week.
Is it causal? No. It's directional, with known endogeneity bias.
What's it for? Incremental decisions: raise/lower bid, reallocate budget across retailers or keywords.
What's it NOT for? Big discrete decisions (shutting down a line, changing agencies). Use a geo holdout there.
When do you experiment? Annually on your largest line + whenever the model contradicts strong intuition.

The observational attribution model isn't magic and it isn't pure causal attribution. It's a detective tool: it gives you suspects ordered by probability and tells you which to chase first. For most retail media decisions — incremental budget allocations across retailers, keywords, SKUs — that's exactly the information you need.

Where the model falls short, point experiments take over. Where the model is enough, it avoids the cost and slowness of experiments. The combination of the two — continuous observational model + annual point experiments — is the closest thing to a complete retail media measurement operating system that exists today.

In the last article in the series we close with the layer that pays for itself every day: real-time operational scraping to detect competitor stockouts, price changes, and smart bidding opportunities.

Sources

Want to run observational attribution on your LATAM portfolio? ePerfectStore.com captures the six key variables across every retailer, week by week, ready to feed your model.

Back to Blog