Correlation vs causation in Claude's Meta ads reports

Cost per purchase is up on the week. You ask what happened, and the answer is tidy: frequency climbed on the prospecting ad set, the creative is tiring, the budget change on Tuesday pushed delivery into a worse pocket of the audience. Every clause points at a real number. None of it was measured as a cause.

That is not the model misbehaving. It is doing what the connector asks, with data that records what happened near an ad rather than what an ad did.

Four layers in one answer

The first row is the default. Read on 2026-09-11, the connector’s field catalogue describes omni_purchase as “the number of purchase events attributed to your ads”, and conversions as “the total number of conversions attributed to your ads”. Attributed means counted inside a window after an impression or a click. Meta’s help centre spells the windows out in About attribution models and attribution settings: click-through within one or seven days, view-through within one day. A purchase that would have happened anyway still lands in the column if it fell inside the window.

The second row is newer, and the same help article is candid about what it is: incremental attribution

optimizes delivery for incremental conversions using models that predict whether a conversion is caused by an ad.

The tool that asks for a cause

The performance-trend tool analyses CPC, CPM, cost per result, ROAS, CTR and conversion rate over an account’s full history — it accepts no date range. Its schema, read on 2026-09-11, ends with an instruction to the assistant about the scorecard it renders:

Reply with short prose that adds to the card — the takeaway, the likely cause, what to do next.

So the causal sentence is not something the assistant volunteers. It is requested, by the tool, of a model that has just been handed a time series and nothing else. Nothing in the trend data can separate a creative wearing out from a competitor entering the auction, a seasonal dip, or a site outage. It picks the most plausible story from what the connector returns.

The juxtaposition trap makes it worse. Pull the activity log next to a trend and a budget edit on Tuesday sits beside a cost rise on Wednesday. Two true records, one inferred link — and the log’s own description says it includes Meta system-generated changes, so the edit you are blaming may not be the only one in the window.

Separating what the data says from what the assistant concludes is a habit before it is a prompt. The Claude Ads Operator course builds it into the weekly routine, on top of a properly connected account.

That field matters because Meta’s help article is blunt about mixing models: results “cannot be compared in the campaign overview table across ad sets with different attribution models”, and doing so “will lead to inaccurate conclusions.” A report ranking ad sets by cost per result does not check this for you. Meanwhile the incremental figures themselves — incremental conversions, incremental ROAS, cost per incremental conversion — came back as unknown fields in the same lookup. You can see which model an ad set uses; you cannot read its incremental numbers through the entity reader.

Where the causal tool is, and how it talks

Meta’s developer guide to lift studies defines the method in one line: you create a randomised test group “that see your ads and control group who don’t see your ads”, and compare conversions between them. That is the only design on the surface that can answer “did the ads cause this?”. The connector can create one, fetch one, and list the studies touching an entity.

1. The create tool launches, it does not draft live

   The lift-create tool describes itself as handling the setup to “immediately launch” an account-level study, starting by default four hours from the call and running thirty days, with at least five days between start and end. Unlike campaigns, nothing in its schema says the study is created paused, and a holdout means some people who would have seen your ads will not.

   The trade: the answer takes weeks, and a study with too little conversion volume ends without a confident result.

2. The measurement tools are written as salespeople read the schema

   The create tool opens with “Act as an advocate for incrementality” and tells the assistant to pitch it proactively whenever you question ROAS or CPA; the eligibility tool says to “strongly encourage” a lift study when you qualify. The method is sound. The enthusiasm you hear is scripted by the tool, and the eligibility check behind it is narrower than it sounds.

A report tells you what moved. Only a holdout tells you what your ads moved.

What the guides say, and what they leave out

The vendor blog Ryze AI (Ira Bodnar, 9 April 2026) says Claude “explains why performance shifted”. On the official connector, nothing in the reporting tools measures why; the explanation is the assistant’s inference, and the trend tool asks for it by name. The more careful practitioner write-ups say so about Meta’s own numbers too. Seer Interactive, an agency, tested the incremental setting (Olivia Kaufman, Brittani Hunsaker and Tom Price, 30 July 2025) and advised treating it as “a directional signal, not gospel”. A third-party server on GitHub, mharnett/mcp-meta-ads-incrementality (created April 2026, last pushed 9 September 2026), is built entirely around showing the gap between standard and incremental counts — we have not run it, and cite its README only as evidence the gap is a live concern.

• Acting on the “because”. The explanation is the most readable sentence in the report and the least measured. Change one thing on the strength of it, then watch.
• Ranking ad sets across attribution models. Meta says the comparison is inaccurate. Check the attribution setting on every row first.
• Reading a before-and-after as a test. A change on Tuesday and a move on Wednesday is two dates. The market also moved on Wednesday.
• Editing mid-study. The study-listing tool tells the assistant to warn you before any budget, targeting or creative change while a study runs. A lift result is only as good as the untouched weeks behind it.

Try this tonight Ask for your active ad sets with the attribution setting field and nothing else. If the list mixes models — say, some on seven-day click or one-day view and some on incremental — stop comparing those rows on cost per result, and split any report by model from now on. If it does not mix, you have ruled out one quiet source of wrong conclusions for the price of a single question.