How to Use Causality Engine to See What Each Channel Caused

The short answer: four or more channels each claim the same sale, and the budget moves this month. Run Causality Engine on a 12-month GA4 export before it does. The first finding is free at app.causalityengine.ai/start, and the full read is €99 once, refundable within 30 days (pricing, checked on 27 October 2026). The steps: download the Attribution paths CSV and drop it on the start page for the free preview. Then pay for the full read, take each channel's four fields in order, test the verdict in one region, and share the read with whoever signs off.

Disclosure: I founded Causality Engine and own it, and this page recommends it. Every claim about it links to Causality Engine's own pages.

Before the export, put three things on paper:

1. The claim ratio. For example, the help page describes Facebook reporting 82 sales where Shopify shows 40: 82 ÷ 40 = 2.05 platform claims per order. Work out your own week the same way.
2. The access you already have. Google's share and export help says the Viewer role is enough to download a report. Your purchase must also be marked as a key event.
3. The decision the read must settle. One channel to scale, cut or defend, and one region where a change could be tested. Every verdict in the read comes with a regional test, per how it works.

What the product is and costs sits in the Causality Engine glossary entry. This page is the job itself: one file in, one budget decision out.

Step 1: Download 12 months of attribution paths

The start page gives the export in four moves. In GA4, open Advertising, then Attribution paths. Set the date range to 12 months or more and pick your purchase key event. Click Share this report, then Download file, then CSV. Google's report help files the same report as Key event attribution paths, under the Key events dropdown. The checks worth running on the file before you upload it are in How to export GA4 data for attribution.

Step 2: Take the free first finding

On the start page, enter an email to open the upload; the finding is sent to it once. Then drop the CSV. The preview runs in your browser and the file stays there. The page notes the file holds channel paths and totals, with no customer names, emails or order IDs.

What comes back is the start of the answer. The start page calls it the channels behind your Direct, and the checkup page adds the number of days buyers who saw two or more channels take to convert. It is free. It is also enough on its own to tell you whether Direct is sitting on sales other channels began.

Step 3: Buy the €99 read

The full read costs €99 per upload, paid by card through Polar under section 3 of the terms. No subscription sits behind it. A refund within 30 days needs one email and no reason, and you keep access to the analysis anyway. Results land 1 to 2 minutes after upload.

The privacy policy deletes a pay-per-use analysis and its source data 30 days after your last visit. Save the share link and the CSV yourself.

Step 4: Read each channel in four fields

Every channel row carries the engine's share of credit once Direct is split back and last-click's share with Direct left out. It also carries a data-health score from 0 to 100, and a next step. The page written for agents states the reading rules plainly. Within 25% of each other, the two methods agree, so hold. Under 1% on both, too small to act on. Health under 90, move budget in steps. Unassigned means no tracked source, so fix tagging first.

Here is the pricing page's illustrative sample, which the pricing page labels round numbers rather than a customer's read:

Read the gap first, then the score. Paid Social loses eight points of credit with health at 93, which is a verdict you can test. Cross-network gains at 81. That is a direction, to be taken in small moves.

Step 5: Test the verdict in one region

Pick the region before you touch the budget. Change spend a little there, and compare revenue after the number of days the read names. Fix the end date in advance: the free holdout test planner warns that stopping at the first good-looking day makes a false result more likely than the confidence level says.

Step 6: Share the read and keep the file

The share link gives whoever signs off the same table, with receipts that check each number against the export. Keep the CSV next to it. A quarter from now, the same export run again shows what moved, and How to run a quarterly attribution checkup sets out that comparison.

Where an agent helps, and where a person signs

On Pro, a chatbot takes ad-hoc questions over your own attribution data, per the pricing page. Ask it, for example, which channels gained credit since the last read. The privacy policy says that data can go to OpenAI, Anthropic, Google and PublicAI for real-time inference only, with no model training on it.

On a single read there is no API or MCP server, per the AI instructions page, so a general assistant works only from the channel table you paste into it. Give it the four reading rules above and it can draft the budget note. The human check is the same either way: a person ticks every figure in the draft against the read's receipts and approves the regional test before any spend changes.

The assistant drafts. A person signs.

Mistakes the product's own pages flag

• Feeding it a thin history. The one-off read page says not to buy when analytics history is thin or broken. No causal method recovers data that was never collected.
• Acting on Unassigned. No tracked source means a tagging fix comes before any budget move.
• Treating the read as an experiment. The agents page calls it observational: a holdout is the anchor, and the read runs between anchors.
• Mixing in platform ROAS. It is the number you have watched every week, so it feels like the baseline. The same page says platform-reported ROAS is never an input to the causal estimate. A disagreement with Meta's dashboard is the finding, not an error.

Free first, then the €99 read

Last click is a sensible default to start from. It is simple, and the paths report draws its chart with it unless you change the model dropdown, per Google's report help. It also suits whoever sells the closing click. Branded search and retargeting collect the whole sale under it, per the last-click page, and both are bought from the platforms whose dashboards you are trying to check.

So begin with what costs nothing: the claim ratio on paper, then the free first finding on the start page. If it shows Direct holding sales that other channels started, Causality Engine, which I founded, runs the €99 read on the same file, refundable within 30 days. Pro at €299 a month earns its place only once you run 8 or more channels with weekly budgets (pricing, checked on 27 October 2026). The method behind both is on how it works.

FAQ

How long does the whole job take?

About 2 minutes to export, per the start page, and 1 to 2 minutes for the read after upload, per how it works. The regional test is the slow part. It takes as many days as the read names for that channel.

What if the read disagrees with Meta's dashboard?

Expect it to. Meta's number is a claim under Meta's own rule, and the agents page keeps platform ROAS out of the estimate. Test the channel in one region before you cut it.

Does it work if my store is not on Shopify?

Yes. The pricing page says any ecommerce platform works as long as GA4 is installed; Shopify is not required.