Causal Inference
56 articles on causal inference, newest first.
Offer, creative, pages or retention: which drives growth?
Two DTC brand teardowns show the same playbook: a first order loaded with discounts and gifts, hundreds of ads, a page for every angle, and retention behind it. Neither video can say which pillar drives the growth. Your own data can, if you read one pillar at a time.
Ad-matched landing pages: test the page, not the ad
An ad-matched landing page gets the audience its ad already persuaded, so its conversion rate flatters it. Compare pages only within the same ad in GA4, then split one ad's traffic between the matched page and your default page.
Price tests: read margin per visitor, not conversion
A higher price can lower conversion rate and still earn more, so conversion rate is not the verdict. Judge a price test on margin per visitor after ad spend, and run it as a switchback or a product split instead of a before-and-after.
A competitor's ad library shows ads, not results
The Meta Ad Library shows which ads a competitor runs and when they started. For ordinary commercial ads it shows no spend and no results, so a teardown infers winners from survival. Treat what you find as ideas, and test them in your own account.
Your Meta creative test is not an A/B test
A creative test inside Meta compares two ads that were shown to different people, picked by the delivery algorithm. It tells you which ad wins in that system, not whether the creative itself is better.
Does your customer community cause retention?
Members of a customer community buy again more often than everyone else, but that shows who joins, not what joining does. To measure what the community causes, randomly choose which new customers are invited.
Products with more reviews sell more. So what?
Products with more reviews sell more, partly because they have already sold more. To learn what reviews or a sold counter do on your store, hold the product fixed and change only the proof for a random half.
When to re-test a channel you already ruled out
A test result belongs to the product, offer, margin, market and period it was run in. When enough of those change, a verdict that a channel does not work has expired, and the only way to know is to test again.
Machine readable attribution for AI agents
What an attribution result has to carry to be usable by an agent, a Slack bot or a Notion database, rather than only by a human reading a chart image.
When everyone clones the same offer block
If the offer block is what converts, and the offer block is being copied across a whole category, the edge decays. So does the contrast you need in order to measure it.
Incremental sales, not correlations
How to tell whether a measurement platform is identifying incremental sales or simply describing a correlation in a better looking chart. Four questions that separate them.
An AI rebuilt page breaks channel attribution
Rebuilding a product page changes the conversion rate for every channel at once. Any channel comparison spanning that date is measuring the rebuild, not the channels.
The cost of scaling a false winner
A false winner costs more than the spend behind it. It displaces something that worked, sets a target nobody can hit, and teaches the wrong lesson to everyone watching.
What a winning product page is hiding
Six plausible explanations for a page that outperforms, only one of which is the page itself. Ruling out the other five is what makes the lesson worth carrying forward.
Own the test results, not just the winners
Most teams keep the winning variant and discard the rest. The discarded results are the part that would have made the next decision cheaper.
Copying bestsellers is survivorship bias
Browsing what sells shows you the winners only. The near-identical products that failed are invisible, which makes any pattern you spot from that view unreliable.
Test now or wait until the season ends
A product test that runs into peak trading cannot be separated from the season. Either finish before it starts, or accept a result you will not be able to use.
How to decompose what actually sells
A practical sequence for separating concept from execution across a product line, using batches, a holdout and an interval rather than a hunch about which feature mattered.
A bestseller is not evidence of why
Sales prove a product sold. They do not isolate which of its properties caused it, and copying the wrong property is the most common mistake in the category.
Vary one element, learn something reusable
Creative tests that change several things at once produce a winner and no transferable knowledge. Isolating one element is what makes the result apply to the next thing you make.
Explaining your attribution method to non-specialists
Four analogies that survive questioning and two that collapse under it. Describing a causal read without jargon and without misleading anyone.
Defending a finding starts with naming the method
A finding is only as defensible as the design behind it. Naming the design first is what turns a claim into something a sceptic can actually weigh.
Why real-time attribution alerts mislead
A causal estimate is not a live metric. Alerting on it in real time reports noise as news, and the team that acts on those alerts is acting on nothing.
What multi-touch attribution never actually measured
Set coverage aside. Even working perfectly, a path model answers a different question from the one you are asking. The gap is older than the privacy problem.
What a real Shopify channel lift answer requires
Lift means a comparison against a world where you did not run the channel. Three requirements before the word is honest, and where speed genuinely trades against strength.
Unified Marketing Measurement, Minus the Vendor Definition
The category label promises that marketing mix modelling, multi-touch attribution and experiments combine into one trustworthy number. Three methods that disagree do not average into truth.
Incrementality Testing vs A/B Testing: What Each Proves
An A/B test tells you which variant wins. An incrementality test tells you whether running the thing at all beats not running it. Teams routinely use the first to answer the second.
Incrementality vs Attribution: Two Different Questions
These two words get used as competing methods for the same job. They are not. One divides observed credit, the other estimates a counterfactual, and confusing them is how budget gets moved on the wrong evidence.
When the Three Systems Disagree, Which ROAS Is Real?
Your ad platforms, your analytics and your store will never agree, because they are answering three different questions. Knowing which question each answers tells you which number to put in the plan.
Does Journey Orchestration Actually Increase Revenue?
Every orchestration platform can show you conversions its journeys touched. None of that establishes that the orchestration caused them. The question has an answer, and it is not in the platform's own dashboard.
Journey Orchestration After the Third-Party Cookie
Losing third-party cookies changes what orchestration can observe and join. It does not change the fact that the system deciding who gets an experience cannot also be the system that proves the experience worked.
Attribution Vendors That Explain Their Methodology Openly
Open does not mean the code is public. It means the vendor states what the model assumes, whether the number is experimental or observational, what the interval is, which of your channels it cannot measure, and what it would report if your ads did nothing.
Marketing Mix Modeling vs Causal Attribution for DTC Brands
MMM and causal attribution both try to estimate what advertising caused, at different grain and different cost. For a DTC brand the choice turns on history, budget and how often decisions are made, and the honest version of either states what it cannot validate.
Causal Inference for Marketing Teams: A Non-Technical Guide
Every attribution report answers 'which channel touched the sale'. Causal inference answers 'what would have happened without it'. The difference, the trap that makes observational numbers wrong, and the four things a causal number has to state.
A Dutch Province Geo Test Cannot Reach p Below 0.05
One treated province against eleven placebos gives twelve possible orderings, so the best p-value the design can return is 1 in 12. That is above 0.05 before a euro is spent. The floor by country, and the three fixes.
Can Your Brand Measure Black Friday Lift at All? The Floor
Before spending a cent on a Black Friday test, two tables and one multiplication say whether it can produce an answer. Find your row.
Retargeting Holdout Before Black Friday: The Mystery Box
Retargeting reaches people who already came to your store, which is why it reports well and why nobody knows what it adds. A regional holdout before Black Friday is the only way to find out, and the three possible answers are all useful.
Ask Your Attribution Vendor for a Placebo Test
Thirty-one commercial measurement vendors, audited between 14 August and 2 September 2026, and not one published validation against randomised experiments. Four questions to ask yours before the Black Friday read, including the one about nothing.
Point AI at the Wrong Question and It Will Confidently Lie
A model will answer whatever you ask with total confidence. The trick is asking a causal question, not a correlation one.
How to Prove a Channel Caused Revenue Without Running an Experiment
Geo holdouts are not the only way to prove incrementality. Here is how causal inference on the GA4 export you already own answers the same question in minutes, and when you should still run the experiment.
Is GA4's Data-Driven Attribution a Black Box You Can Trust?
GA4's data-driven attribution is a real model, but it is correlational, unauditable, and many smaller properties are not even running it. Here is how to check yours, and what to do instead.
Structural Equation Modeling for Attribution: Mapping the Full Causal Chain
If you’re still using last-touch, multi-touch, or, god forbid, LLM-based attribution, you’re measuring ghosts. Structural Equation Modeling (SEM) doesn’t just count clicks; it maps the full causal chain from impression to incremental sale.
Difference-in-Differences for Marketing: Measuring Campaign Impact Scientifically
Difference-in-differences (DiD) cuts through marketing noise with causal inference. How DiD measures true campaign impact against flawed attribution models.
Synthetic Control Methods for Marketing: Building Your Counterfactual
Synthetic control methods cut through the noise of broken attribution by building a counterfactual: what would have happened if you had spent nothing.
NULLs in Marketing Data: What LLM-Written SQL Must Handle
A NULL channel satisfies neither equals nor not-equals, and NOT IN returns nothing if its list holds one. A peer-reviewed survey and an error study show how often people and models trip on it, and a four-way count checks your data.
Granger Causality in Marketing: Does Your Ad Spend Actually Cause Revenue?
Granger causality in marketing claims to prove ad spend causes revenue: but does it? Spoiler: No. We break why time series causality fails and what actually works.
Association vs. Causation in Marketing: The Expensive Mistake
Stop wasting your marketing budget. Learn the critical difference between association vs causation and how it impacts your ROI.
What Is Causal Inference and Why Every Marketer Needs It
Stop guessing and start knowing. Learn what causal inference is and how it reveals the true impact of your marketing spend.
The Frontdoor Criterion Explained for Marketing Analytics
Learn how the Frontdoor Criterion, a powerful causal inference method, can fix your broken marketing analytics and reveal true campaign impact.
How a TikTok Ad Can Lead to a Later Meta Conversion
This story is comforting, easy to measure, and catastrophically wrong. It completely ignores the complex, chaotic, and ultimately more profitable reality of customer behavior.
Counterfactual Analysis for Ad Spend: What Would Have Happened Without That Campaign
Discover how counterfactual analysis in marketing reveals the true impact of your ad spend. Stop guessing and start measuring real incrementality.
Directed Acyclic Graphs for Marketing: A Practical Guide
Stop guessing and start knowing. This practical guide to Directed Acyclic Graphs (DAGs) for marketing shows you how to build causal models that reveal true ROI.
Causal AI in Marketing: From Correlation to Confidence
Stop guessing and start knowing. Learn how causal AI moves marketing beyond correlation, and how to test what each channel caused on your own store.
Instrumental Variables in Marketing: Isolating True Channel Impact
Stop guessing your channel impact. Learn how instrumental variables in marketing isolate the true causal effect of your campaigns, moving beyond flawed attribution.
How Causal Inference Reveals Which Channels Actually Drive Sales
Stop guessing which channels drive sales. Causal inference provides the ground truth on channel performance for Dutch ecommerce brands.
Why Correlation-Based Marketing Decisions Waste Budget
Stop wasting your marketing budget. Learn why correlation-based marketing decisions are costing your brand and how causal inference can reveal true ROI.
Your platforms guess.
We run the math.
Upload a GA4 export and see what each channel caused, next to last-click, in 1–2 minutes. The read is yours to keep.
Free, in your browser: your file is not uploaded. The full read is €99, refundable within 30 days. Prices exclude VAT.
Or book a 30-min call.