18 Apr 2026 · Incrementality

Incrementality versus MMP reports in Thailand

An MMP report answers a matching question: given the identifiers we have, which campaign’s rule claimed the install? Incrementality answers a causal question: did spending more in this place or on this network create opens that would not have happened? Those questions are allowed to disagree. In Thailand they often do, and the disagreement is not a bug in either system.

Tourism seasons, Bangkok-heavy digital inventory, and uneven connectivity outside the central region all make geo tests messy. A holdout in the northeast is not a miniature version of the capital. If you treat it as one, you will either over-trust a null result or invent lift from noise. Vault Nodecore still prefers a well-described messy holdout to an MMP screenshot that nobody can rebuild.

When the matching rule should yield

If always-on social keeps receiving MMP credit while a paused province shows no drop in first opens, the matching rule is flattering the network. That is the corrective incrementality is for. Write both numbers. Do not delete the MMP table; annotate it. Finance can live with two tables. They cannot live with a single table that changes meaning depending on who is in the meeting.

PSA-style tests and ghost bids have their own problems on regional inventory. Delivery and cheap video placements do not always have a clean “unbranded” creative to run. If you cannot describe the counterfactual in a sentence, do not cite the test as if it were a laboratory.

When the test is too thin

A three-day pause, a single district, or a campaign that was already exhausting demand will not overrule a quarter of MMP data. Students sometimes arrive hoping incrementality will rescue a CPI they cannot defend. It will not. Thin tests are another form of guesswork. The intensive’s incrementality week includes a real limitation: we will mark a capstone down if it leans on an underpowered holdout to dismiss a reconstructable click-ID handshake.

Thai desks have an advantage if they already report by province or by store cluster. Use that grain. Do not import a US-style national geo split that ignores how people actually move between Bangkok and the rest of the country during festivals. Campaign-to-install attribution remains the daily language. Incrementality is the occasional corrective. Keep the hierarchy explicit, and you will argue less about which dashboard is “true.”