You have two years of spend and conversion history. A vendor can turn it into a smooth response curve and show where a channel starts to saturate.
The curve may still be answering a question your data never observed.
If spend stayed nearly flat, channels moved together or the market changed whenever budgets changed, more history does not automatically reveal the bend.
Start with the decision, not the dataset
Response curves are useful when they change a recurring budget decision that simpler rules cannot resolve.
Should paid social move from $40k to $50k per week?
Yesterday's mature ROAS is acceptable, but the team does not know whether the next tranche is already hitting saturation.
That is a candidate for response analysis. “We have historical data and want better allocation” is not. It has no declared move, range or cost of being wrong.
If small, reversible changes with mature data, caps and rollback rules already settle the decision, keep the rule until a real budget question exceeds it.
There is no magic spend threshold
Is $10k per segment enough? The dollar amount cannot answer that.
In one product, $10k may generate thousands of mature purchases across repeated markets. In another, it may generate twenty long-lag conversions inside a segment whose ordinary weekly movement is larger than the expected media effect.
The first dataset may contain a usable signal. The second may produce a curve whose shape comes mostly from model assumptions.
Five gates before the curve
- 01
A decision worth changing
Name the current level, proposed level, channel and window. If a different answer would not change a material action, stop here.
- 02
A visible business result
Use mature orders, revenue or margin that stays comparable across periods. At a narrow segment level, the result may be too sparse even when the channel total is usable.
- 03
Media that actually moved
A curve needs useful differences in spend, impressions, reach or frequency. A nearly flat series cannot show where the relationship bends.
- 04
Outside changes you can see
The market need not be stable. Price, promotions, launches, availability, demand, competitors and auction conditions must not silently become “media effect.”
- 05
A way to challenge the answer
A spend step, geo split, staggered change, uplift or holdout can target the range behind one expensive move. It does not have to validate the whole model.
A stable market is not the requirement
The dangerous condition is not change. It is an unrecorded change that moves with media.
Demand changed, and the team can see or design around the change.
Demand changed with spend, but the model has no way to tell the movements apart.
Competitor and auction effects will rarely be observed perfectly. Available competitor signals, query demand, price and promotion history, platform delivery metrics and geo/time variation can still show whether the budget action survives plausible alternatives.
Historical uplift is support, not an entry ticket
An uplift or holdout result is not required for software to fit a curve. It becomes valuable when ordinary history cannot tell two stories apart.
A channel may look saturated because each additional spend step produced less. Or effectiveness may simply have declined while spend happened to rise. Both stories can fit the same history and imply different budget actions.
When that distinction matters, deliberate variation creates evidence instead of waiting for more rows.
Should your team investigate a response curve?
No upload. No lead form. The answers stay in this browser.
Where the series goes next
This article stops at eligibility. It does not teach carryover, saturation functions or marginal return, and it does not authorize a budget move.
Evidence notes
The method was checked against official Google Meridian guidance on data amount, geo variation, controls and scope; the Meta Robyn Analyst Guide on variation, volume and granularity; primary research on separating nonlinear saturation from time-varying effectiveness; and Google research on experiment-informed MMM calibration.
These sources do not provide a universal spend cutoff. This is a routing gate, not a dataset certification, power analysis or promise that a response curve will be causal.