Creative, Retargeting, and Saturation
Review experimental effect sizes for targeting fit, messaging, retargeting timing, frequency, and budget saturation, and apply them to simulations.
Written and reviewed by Eric Han · Co-founder & CTO, reopt
Key takeaways
- Personality-matched Facebook beauty creative increased the odds of purchase by a factor of 1.54 without changing clicks. Two app experiments also increased clicks, so clicks alone are insufficient to assess creative.
- Very early cart reminders can backfire: 13.1% purchased after a reminder sent at 30 minutes, versus 15.2% in the control group; for reminders sent at 24 hours, the figures were 8.3% versus 6.0%.
- Platform creative A/B tests can be distorted when algorithms deliver different creatives to different people.
- Median long-run elasticity for US CPG TV advertising was 0.014, and more than 80% of brands had negative marginal ROI.
- We found no public randomized measurement of a complete saturation curve. Estimate its shape through internal experiments.
Creative and Messaging
| Experiment | Context | Effect | Grade |
|---|---|---|---|
| Matz et al., PNAS 2017 | Facebook, UK beauty, reach 3.13 million | Personality-matched purchase odds ratio 1.54 (1.25–1.90); no click difference. Platform-optimized delivery prompted an internal-validity debate (Eckles et al., 2018; authors replied) | B |
| Same paper | Two app experiments | Click odds ratios 1.38 and 1.30; install odds ratios 1.31 and 1.15 | B |
| Bertrand et al., QJE 2010 | South African lending, direct mail to 53,194 people | Reducing example loans from four to one had an effect comparable to a 200-basis-point interest-rate cut | A |
| Sahni et al., Marketing Science 2018 | Name in email subject | Opens 9.05 → 10.80%; leads 0.39 → 0.51% | A |
| Wernerfelt et al., 2025 | Over 70,000 Meta campaigns | Optimizing clicks instead of purchases raised incremental CAC 31% | A |
| NYU Stern summary | Field experiment, AI creative | CTR up to +19%; −31.5% with AI disclosure. Peer-review status unconfirmed | D |
What “creative accounts for 49%” means
NCSolutions 2023 assigns 49% to creative, 21% to brand, 14% to reach, 11% to targeting, and 5% to recency. These are model contribution shares from machine-learning analysis of about 450 CPG campaigns, not a claim that changing creative raises revenue 49% (measurement vendor, grade C).
Treat platform creative A/B tests carefully. Braun and Schwartz (Journal of Marketing) show how algorithmic delivery to different users can distort both the magnitude and sign of creative A/B results. Creative comparisons and incremental-lift tests are different experiments.
Retargeting
| Experiment | Context | Effect | Grade |
|---|---|---|---|
| Johnson et al., Ghost Ads | One online retailer, two weeks | Visits +17.2%, transactions +12.0%, revenue +10.8% | A |
| Sahni et al., JMR 2019 | Home-improvement supplier | Return visits within four weeks +14.6%; 33% of the first-week effect occurred on day one | A |
| Lambrecht and Tucker, JMR 2013 | Travel | Dynamic creative underperformed generic creative on average; the disadvantage disappeared after review-site visits | A |
Timing Can Reverse the Sign
Li, Luo, Lu, and Moriguchi (Journal of Marketing) randomized the timing of owned messages to cart abandoners. These were email, app, and SMS messages, not paid ads. Figures measure purchasing within one month (40,500 + 23,900 people, grade A).
| Message timing | Treatment | Control |
|---|---|---|
| After 30 minutes | 13.1% | 15.2% |
| After 1 hour | 13.4% | 16.7% |
| After 24 hours | 8.3% | 6.0% |
| After 72 hours | 4.8% | 1.8% |
Treatment and control groups differ by timing, so compare differences within each row, not raw purchase rates across rows.
If automated cart messages go out within an hour, include a holdout and compare with sending after 24 hours. These are owned-message results and should not be transferred directly to paid retargeting.
Frequency and Saturation
| Evidence | Finding | Limitation | Grade |
|---|---|---|---|
| Lewis and Reiley manuscript | Effect peaked around 50 exposures; incremental revenue equaled ad spend near 100 exposures, not profit break-even | Exposure frequency was not exogenous; authors caution against causal interpretation | B |
| Johnson et al. experiment meta-analysis | Visit carryover in the 28 days after campaigns: mean 16.4%, median 2.9% | A few campaigns raise the mean | A |
| Shapiro et al., Econometrica 2021 | 288 US CPG TV brands; median elasticity 0.014; negative marginal ROI for more than 80% | TV data cannot be transferred directly to digital | B |
| Sethuraman et al., JMR 2011 | Advertising-elasticity meta-analysis: short run 0.12, long run 0.24 | Brand advertising from 1960–2008 | B |
Simulation Inputs
| Stage | Input | Evidence and conditions |
|---|---|---|
| Retargeting incrementality | 0 to +12% (author judgment) | Transactions +12.0%, revenue +10.8% at one retailer (A); retain a range until a holdout test |
| Saturation elasticity | Scenario assumption | Closer to 1 for new channels; lower for channels already receiving substantial spend |
| Post-campaign carryover | Median visit lift 2.9% | Mean 16.4% is driven by a few campaigns |
Do not use NCSolutions' 49% creative share as a lift, or TV elasticity of 0.014 as a digital-channel input.
-
Start with a concave curve. Assume doubling the budget does not double conversions. No suitable public parameter estimates were found, so use sensitivity analysis.
-
Treat elasticity as a scenario. Set it closer to 1 for a channel you have just started using and lower for a heavily funded channel. This functional form and scenario distinction are author judgments.
-
Estimate through budget-split experiments. Vary spending across regions or periods and estimate as the log slope between two points.
A Korean reference is Naver Place click elasticity of about 0.55–0.68 (D, before/after comparison, clicks only), covered in Korean Cases.
Missing evidence
We found no public digital-media study directly measuring a saturation curve by randomizing exposure frequency. MMM saturation curves are model assumptions, not direct measurements.
Reported and Incremental Conversions
Use randomized experiments at eBay, Meta, and Google to assess the gap between platform reporting and actual effects, and choose adjustment multipliers for simulations.
Referral Funnel Metrics
Compare industry sharing, invitation conversion, acquisition, and time-to-first-referral metrics, aligning vendor denominators to build a referral forecast.