The Cheapest Way to Get a Causal Attribution Read

Ranked by real cost and effort: the fastest, lowest-priced ways to measure what last quarter's marketing spend actually drove, and when the cheap option is the wrong one.

Updated 8 September 2026 · Joris van Huët, founder, Causality Engine

If the only question you need answered is "where should next month's budget go," the cheapest realistic option for most ecommerce brands is a one-off causal read from a GA4 export: 99 euro, a CSV upload, a result in 5 to 10 minutes. It gives you per-channel incremental ROAS with confidence intervals, a platform-reported versus causal comparison, and budget reallocation suggestions, and it is refundable if it does not move a budget decision.

That is the short answer. The longer answer is that "cheapest" changes depending on what you actually need to prove, and a few of the options below cost far more in labor and calendar time than their price tag suggests.

The two numbers that cost nothing

Before any of the five options, compute two ratios against the one figure no vendor and no platform produces: the orders in your commerce platform. Coverage is the conversions your analytics could attribute to any source, divided by those orders; every cost per acquisition you have ever read is the real one divided by it, and in our own analytics warehouse 48.6% of sessions between 17 October 2025 and 2 September 2026 had no resolvable source (one company's census, not a benchmark). The claim ratio is every platform's claimed conversions added up, divided by the same orders; anything above 1 is the amount by which your suppliers collectively believe they did more work than exists. An hour, a spreadsheet, and you already know more than most of the meeting.

The five realistic options, ranked by fully loaded cost

Price alone is a bad way to rank these. A free open-source tool can cost you six weeks of an analyst's time. The ranking below is by total cost of getting a decision you can defend, not by license fee.

Option	Cash cost	Real effort	Time to result	What you get back
One-off causal read (Causality Engine)	99 euro, refundable	About 2 minutes of operator time	5 to 10 min	Per-channel incremental ROAS with confidence intervals, platform vs causal comparison, budget reallocation, exportable report
In-house Meridian or Robyn	Free software	High: analyst plus data prep	Weeks to months (Robyn's own guide adds up to roughly 12 to 22 weeks for a first study)	Channel contribution, ROI, credible or confidence intervals, response curves, budget scenarios
Freelance analyst	Custom, scoped per project	Medium to high: briefing, access, revisions	Usually weeks	Depends on scope; model files, cleaned data, ROI, diagnostics, recommendations
Incrementality-test vendor	Quoted per engagement; Measured and Haus publish no price list	High: holdout budget, test design, approvals	Two weeks at minimum, longer for upper-funnel media, plus an observation window	Measured lift vs control, scale/cut/reallocate guidance
Enterprise MMM contract	Custom, not published	Highest: procurement, data warehouse, program ownership	Scoped per engagement	Recurring contribution modeling, planning scenarios, dashboards

Option 1: A one-off causal read from a GA4 export

This is the lowest cash cost and the lowest effort when your brand fits. You export GA4 data as a CSV, upload it, and the model reads the spend and conversion history already in that file. The published estimate is about two minutes to export and upload. No pixel, no SDK, no DNS change, no developer ticket, no OAuth connection.

What comes back is per-channel incremental ROAS with confidence intervals, a side-by-side of platform-reported versus causal ROAS (this is where over-attribution usually shows up), budget reallocation recommendations, and an exportable report you can put in front of finance. The first read includes a 14-day Pro trial, and it is refundable if it does not move a single budget decision. Continuous monitoring with automated GA4 ingestion sits in the Pro plan at 299 euro per month.

The honest constraint: there is a fit floor of roughly 5,000 euro or more in monthly paid spend, with meaningful GA4 history. Below that, the signal often is not tight enough to separate from noise well enough to defend a call. The 99 euro price makes this look universally suitable. It is not. If you are spending 1,200 euro a month across two channels with thin conversion volume, a causal read will hand you wide intervals and you will have paid for uncertainty you already had.

The real floor is arithmetic, and it applies to every option on this page, ours included. Multiply a channel's share of revenue (spend divided by revenue) by the return you would honestly defend for it; that is roughly how much total revenue would move if the channel stopped. If that number is smaller than the smallest lift your data can distinguish from noise (about 8% for a typical DTC brand with six months of history and an eight-week test), no method can answer the question at your scale, and the right move is to manage that channel on judgement, openly, rather than on a null result.

Worth being clear about what this is and is not: it is a fast causal attribution read from existing historical data. It is not a randomized holdout experiment. Nobody turns a campaign off. If you need proof that a controlled change caused a lift, jump to option 4.

Option 2: Open-source MMM run in-house (Meridian or Robyn)

Google Meridian and Meta Robyn are free and open source. That word describes the code, not the project.

Meridian requires Python 3.11 to 3.13, and Google recommends at least one GPU. It expects a KPI and media data by channel, lets you add control variables such as search query volume to account for organic brand interest, and supports geo-level data, which Google says carries more statistical information than national data. Robyn's official analyst guide lays out a first-study workflow that reads like a small consulting engagement: 1 to 2 weeks to scope, 4 to 6 weeks to collect data, 1 to 2 weeks to review it, 4 to 8 weeks to model, and 2 to 4 weeks to reach recommendations. That is roughly 12 to 22 weeks when the full process is done properly. A sharp team can get an exploratory model up faster, but the guide is describing the responsible version.

The hidden bill is labor. Someone has to:

• Export and reconcile platform data.
• Resolve gaps between spend, impressions, clicks, and conversions.
• Build a complete time series with no unexplained holes.
• Add promotions, price, holidays, distribution, and other confounders.
• Choose the model window and channel granularity.
• Review competing model candidates and pick one.
• Explain the uncertainty to finance without hand-waving.
• Maintain it when new data lands.

There is one more practical wall. Robyn's guide specifies a minimum of two years of weekly history for robust results, and more when you only have monthly data. So if the whole point is to read last quarter, a conventional MMM built from scratch is often the wrong shape entirely. It is not a law that every MMM needs two years, but that is Robyn's own official recommendation, and it matters when you are staring at 90 days of export.

And one thing both platforms now say in their own documentation deserves a line here. Meridian's page on assessing a fitted model states that directly validating the causal inference requires well-designed experiments, that those are likely not practical if you are using an MMM, and that the causal inference therefore cannot be directly assessed. Read that shape slowly: you use the model because you cannot run the experiment, and validating the model would require the experiment. What is offered in its place is fit.

My read: in-house Meridian or Robyn is the right move when you are building a repeatable internal measurement capability and you have the analyst to own it. It is a poor answer to "what did Q2 do."

If you want the underlying concepts before you commit an analyst to this, the plain-English MMM explainer is a reasonable place to start.

Option 3: A freelance analyst

This is the "it depends" option, and I mean that literally. Freelance MMM work is priced per project, and the number depends on whether the freelancer is just running a model or also doing the data engineering, validation, and stakeholder readout. Advertised prices on freelance marketplaces are not quotes for your data; nothing is until the scope is written down.

A freelancer still needs clean historical data. If you hand over a messy export, you are paying their hourly rate to do the cleanup you could have scoped better. The inputs are usually the same structured time series Robyn or Meridian expects: outcome data, channel spend or exposure, and controls like promotions, price, seasonality, and macro factors. Write the exact input list into the statement of work, and specify the deliverable: model files or code, the cleaned dataset, channel contribution and ROI, uncertainty information where the method supports it, diagnostics, assumptions, scenario analysis, and a decision memo.

The failure mode I have watched play out: you pay for a nice dashboard and treat it as causal proof when the model was never validated. A dashboard is not evidence. Ask how the model was checked before you accept the number.

Option 4: An incrementality-test vendor

This is where you go when the decision can be phrased as a controlled intervention. Increase this channel. Cut that one. Hold it out and see what happens.

Measured and Haus, two of the better-known vendors here, publish no price list (their pricing pages showed no figure when we checked on 8 September 2026), so expect a quoted contract rather than a one-off fee. On duration, Haus's own guide says tests shorter than two weeks tend to produce noisy results, treats two to three weeks as typical for demand-capture tactics, expects four to six weeks or more for upper-funnel video, and adds a post-treatment observation window after spend stops.

The vendor fee is not the real cost. The real cost is the media you have to hold out, the targeting or geography you have to restrict, the clean control group you have to maintain, and the willingness to accept a result that says a channel drove no lift. You also need enough conversion volume for the test to reach significance. Launch speed is not time to a usable read.

Two pieces of arithmetic decide whether a geo test can work at all, and no vendor can waive them. The floor from option 1 applies unchanged. And count your units: one treated region against N regions in total means the smallest p-value a placebo-based design can physically return is 1 in N. Twelve Dutch provinces give 0.083, which cannot clear the 0.05 most people report against; the forty COROP regions give 0.025.

This is the strongest option in the list for one specific question: what happened because we changed exposure. It is also slower and more expensive than a historical read, so it is overkill when you just want to reallocate next month. If you are weighing this route, the geo testing guide walks through what a clean design needs.

Option 5: An enterprise MMM contract

Highest commitment, highest cost, and rarely the answer to a single-quarter question. Enterprise vendors scope pricing around data sources, markets, brands, channels, refresh cadence, and services, and the number is quoted, not published. Recast's site, for one, carries no pricing page at all (checked 8 September 2026).

The hidden costs are procurement, legal and security review, data warehouse work, recurring refresh fees, internal program ownership, and the meetings required to agree on KPI definitions across markets. You buy this when you need recurring, multi-market, multi-channel planning, not when you want to know if last quarter's Meta spend earned its budget.

What I would do first

If your paid spend is above roughly 5,000 euro a month and your GA4 history is reasonably complete, run the 99 euro read first. It is refundable, it takes minutes, and it either moves a decision or it does not. Worst case you take the refund. Best case the platform-versus-causal comparison shows you a chunk of platform-reported ROAS with no incremental read behind it, or a channel last-click wanted to kill that turns out to be pulling weight.

Then decide whether you need more:

• You need experimental proof of a specific change. Budget for an incrementality test. Accept the holdout cost.
• You are building an internal measurement program. Commit an analyst to Meridian or Robyn, or scope an enterprise contract.
• You have a one-off, messy dataset and no in-house time. Brief a freelancer carefully and demand validation, not a dashboard.

Whatever you pick, ask the vendor, us included, one question before you pay: which of my channels are not measurable at my current spend? Between 14 August and 2 September 2026 we audited thirty-one commercial measurement vendors for a published validation of their method against randomised experiments, with the sample, the design and the discrepancies disclosed, and we found none. Hold Causality Engine to the same standard. The read states its counterfactual and its interval, and it is a model on observational data, not an experiment.

When the cheap option is the wrong one

Plainly: the 99 euro read is not the right call if you have less than roughly 5,000 euro a month in paid spend, if your GA4 history is weak or full of gaps, if you need experimental proof rather than a read of existing data, or if the decision depends on offline sales, retail distribution, pricing, promotions, or long-term brand effects that never show up in a GA4 export. In those cases, the extra cost of a test, a specialist, or an enterprise model is buying something the cheap read cannot give you, and the decision is worth protecting.

For everyone else, the ranking holds. Lowest cash cost and lowest effort at the top, and the price climbs fast the moment you need a controlled experiment or a recurring program.

FAQ

Is a 99 euro causal read the same as an incrementality test?

No, and it is worth not confusing them. The read applies causal inference to your historical GA4 data and returns incremental ROAS with confidence intervals without touching your live campaigns. An incrementality test creates treatment and control groups and measures lift during a real test window. The read is faster and cheaper; the test is stronger evidence for a specific controlled change.

Can I just use last quarter's data in Meridian or Robyn?

You can load it, but Robyn's official guide recommends at least two years of weekly history for robust results. A single quarter usually will not give a conventional MMM enough variation to produce a defensible answer. That mismatch is a big reason a purpose-built read of recent GA4 history is a better fit for a last-quarter question.

Why is open-source MMM listed as more expensive than a paid tool?

Because the license is free but the analytical labor is not. Data preparation, model selection, validation, and stakeholder explanation take real weeks of skilled time. When you price that time honestly, running Meridian or Robyn in-house usually costs more than a 99 euro read, unless you are amortizing it across an ongoing measurement program.

What do I actually need to run the cheapest option?

A GA4 export in CSV format covering the period you want analyzed, and roughly 5,000 euro or more in monthly paid spend with meaningful history. A Shopify export can be added when you have it, but GA4 alone is enough for the read. Longer history produces a cleaner result.

Sources and further reading

• The Price of Being Found (Causality Engine, Edition 2.10, September 2026): the coverage rate and the claim ratio (Chapters 8 and 9), the measurability floor and the placebo floor (Chapters 15 and 16), the vendor audit and the platforms' own validation statements (Chapter 17)
• Causality Engine pricing: the 99 euro read, the 14-day Pro trial, Pro at 299 euro per month
• Meridian README: Python and GPU requirements
• Meridian introduction: control variables and geo-level modeling
• Robyn analyst's guide to MMM: phase timings and the two-year minimum
• Haus, how long should you run an incrementality test: test durations and the post-treatment window
• Sixty-second versions of these ideas on YouTube Shorts

Vendor prices and features quoted in this article were taken from each vendor's own website on 8 September 2026 and may have changed since. Check the vendor's pricing page before relying on a figure.