Cannibalization

Shopee
Whether extra ad spend grows your order book or just moves orders you'd have won anyway into the ad-attributed column.

Cannibalization asks one question: when this shop spends more on ads, does the order book grow — or do orders that would have arrived anyway simply move into the ad-attributed column? The page shows the evidence DataGlass has, the gates that evidence passed or failed, and nothing beyond that.

Find it at Ads → Cannibalization, on Shopee shops only. The panel is built from per-order ad classification, which no other marketplace produces.

Every figure here is observational. A shop with no cannibalization at all that spends defensively produces the same pattern as one with a lot of it, so the randomized policy effect is reported as not identified and never inferred. Read this page as evidence to weigh, not as a verdict on your ads.

What the two measures mean

MeasureReads as
CI (cannibalization index)Book displacement per unit of log spend — how much of the non-ad order book moves as spend rises.
BCE (bounded cannibalization estimate)The share of the attributed gain that is paired with a residual loss of non-ad orders.
Book exposureBCE applied at the dose this shop actually spends, in percentage points of the order book.

CI is drawn as a 12-week rolling estimate, recomputed daily. BCE is a single verdict at one date over a 180-day reference window, not a series — so it appears as one figure beside the chart rather than a second line on it.

Outcomes mature for 16 days before they count, and the current BCE calibration is fitted on CPC history; CPM delivery needs an impression-based refit before it can be read the same way.

What each status means

The header carries one status, and it is the honest reading of two independent gates — whether the shop could be measured at all, and whether anything was found.

StatusMeaning
Exposure flaggedThe temporal displacement screen passed its support and stability gates.
No signalWe looked. No publishable residual-loss signal passed the BCE gates.
Unstable signalA raw signal existed but did not survive a stability route.
UnresolvedThe signal is too extreme to publish as bounded exposure.
Not admittedWe could not look — the shop does not meet the historical support floor.

No signal and Not admitted are different answers and are never collapsed into one. The first means the test ran and found nothing; the second means there was not enough classified history to run it.

Reading the cannibalization decomposition

When the index is admitted, four figures sit under the chart. They matter because the scalar alone invites the wrong conclusion: an index driven by the demand term is a shrinking order book, which defensive spending also produces.

FigureWhat it tells you
Scored weeksHow many weeks of mature history the estimate rests on.
Ad shareThe share of the order book that is ad-attributed.
Ad termThe part of the index attributable to ad delivery.
Demand termThe part attributable to underlying demand moving.

If the shop is not admitted there is no quotable figure, and the page shows the exclusion reason instead — that reason is the answer.

Where to take cannibalization next

Cannibalization is a shop-level read. To act on it per campaign, pair it with True ROAS, which nets each campaign's revenue down to profit, and with an optimization group objective and constraint, which is where a spend decision is actually made.