All posts

Your Meta Conversion Number Is Inflated. Here's How to See By How Much.

By Mark Harnett

Meta recently released an official MCP server — the connector that lets AI assistants like Claude or ChatGPT work directly inside your ad account. It's good, and if you manage Meta ads with AI you should connect it. I also built my own Meta MCP, and I still run both. This post explains why, and the answer is really about one word: incrementality.

The number in your report

Open any Meta ads report and the conversion number you see is attributed on a 7-day-click + 1-day-view window. That means Meta takes credit for a purchase if the buyer clicked an ad in the last 7 days — or merely saw one in the last day.

Think about who's inside that number:

  • People who searched for your brand, clicked whatever was in front of them, and were going to buy regardless.
  • Retargeted shoppers who were already mid-checkout when the ad "reached" them.
  • Anyone who scrolled past an impression yesterday and converted today for unrelated reasons.

None of those conversions were caused by the ad. All of them are in the reported number. That's not fraud — it's just what attribution is: credit assignment, not causation.

The number Meta doesn't lead with

Here's the part most advertisers don't know: Meta's own API exposes an `incrementality` attribution breakdown — Meta's modeled estimate of conversions that would not have happened without the ad. It comes out of the same measurement science behind their Conversion Lift program, and it's sitting in the same insights endpoint as the default number.

Almost nobody reports it. The incentives are obvious: the default number is bigger, and bigger numbers make everyone at the table feel better. Agencies get to report more conversions, Meta gets more budget, and the gap between "attributed" and "caused" quietly compounds in your planning.

What I built

I published an open-source MCP server — [mcp-meta-ads-incrementality](https://github.com/mharnett/mcp-meta-ads-incrementality) — that flips the default. When an AI assistant pulls Meta performance through it, the incrementality window leads. Ask for the standard windows in the same call and it computes the gap on every row:

  • **inflation_factor** — reported conversions ÷ incremental conversions
  • **overstatement_pct** — the share of the reported number the ads didn't cause

So instead of "the campaign drove 100 conversions," the report reads: "the campaign was credited with 100 conversions, of which roughly 40 were incremental — the reported number is 2.5× the causal one." Same API, same data, very different budget conversation.

Two honest caveats. Meta's incrementality window is a modeled estimate, not a per-campaign randomized holdout — when the dollars are big enough, a proper Conversion Lift study is still the gold standard. And brand campaigns will always look worse under incrementality than non-brand, which is exactly the point: capturing existing demand is cheap and mostly non-incremental.

Why run both MCPs

Meta's official MCP and mine aren't competitors — they split cleanly along the line of operate vs. evaluate.

Meta's official MCP is the better operator. First-party OAuth (no access tokens to babysit), product catalog management, pixel and CAPI signal diagnostics, campaign creation with everything landing paused by default, plus governance controls for what AI agents may touch. It's the right tool for building and running the account.

The incrementality MCP is the evaluator. The official server reports the standard attributed number — the inflated one. Mine exists so that when the AI answers "what's working?", the causal number leads and the inflation is quantified next to it.

Run the account with Meta's tools. Judge the account with incrementality. The one thing you shouldn't do is let the platform that sells you the ads also define, unchallenged, the number that decides how much you spend on them.

Try it on your own account

The server is MIT-licensed and on npm (`npm install -g mcp-meta-ads-incrementality`). If your reports have only ever shown the default number, pull one month both ways and look at the inflation factor. Whatever it says, you'll know something about your account that most advertisers never check — and you can decide which number deserves to drive next quarter's budget.