Messaging fatigue and suppression: a design for deciding when not to send

1Context and motivation

Automated messaging makes the marginal message almost free to send. Journeys, triggers, promotions and transactional notices each add contacts, and the usual brake is a blanket frequency cap per channel. A cap treats every customer alike: it stops a customer who would have welcomed one more message, and it allows a customer who is close to unsubscribing to receive several more.

The question this paper addresses is narrower than "what should we send". It is: for this customer, now, would one more message do more good than harm? We describe a design for a fatigue and suppression engine that answers it, and the reasons for each part. We do not report effects, because we have not measured any.

2Definitions

We use three terms.

• Fatigue is the state in which additional contacts reduce a customer's responsiveness or raise the chance that they disengage (unsubscribe, mute, uninstall, complain).
• Saturation is the point beyond which an additional message adds no incremental purchase probability.
• Negative lift is the case in which an additional message lowers expected value, because the loss from disengagement exceeds the gain in purchases.

All three are properties of the incremental effect of a message, not of response rates. A customer who buys after many messages may have bought anyway; a customer who stays subscribed may still have stopped reading.

3Evidence on repetition and contact frequency

The design rests on four findings from published work.

Response to repetition rises, then falls. In laboratory experiments in which people heard a persuasive message up to five times, agreement first increased and then decreased as exposure frequency increased, and counterarguing first fell and then rose¹. In media planning, models that include wearout (ad quality declining under continuous exposure and recovering during breaks) show that pulsed schedules can produce more total awareness than evenly spread ones². Both point the same way: the value of a contact depends on how many came before it and how recently.

Email frequency has a nonlinear effect on profit. In a study of a US home improvement retailer's customers, the number of emails sent had a nonlinear effect on both short-term and long-term profitability; the authors' counterfactual analysis put the profit-maximising number at seven emails in their setting, with four or ten emails losing 32% or 16% of lifetime profit per customer³. The same study found that customers active in opening emails were not necessarily active in purchasing. We take two design points from this: the right frequency is an empirical quantity that differs by setting, and opens are not a safe proxy for value.

Contact decisions have long-run consequences. For catalogue mailing, myopic policies that mail customers most likely to order now overlook how today's mailing changes future orders; a dynamic policy that accounts for these effects can be more profitable⁴. The same logic applies in reverse to suppression: a send that looks profitable today can cost future engagement.

Notifications carry a cost in attention. In a one-week field study of 15 mobile phone users, participants received 63.5 notifications per day on average, and a higher number of notifications was associated with more negative emotions⁵. A larger in-the-wild study found that even a notification with useful content can be perceived as disruptive, depending on its presentation, the sender and the user's current task⁶.

4Design

4.1Inputs

• Contact log: customer, channel, message type, time, campaign, region.
• Outcomes: opens, clicks, sessions, purchases, unsubscribes, complaints, mutes, uninstalls, and non-response.
• Customer context: engagement history, lifecycle state, consent and channel preferences, device mix, language, region.
• Treatment assignment: which customers were randomly held out of which sends, so that effects can be estimated causally.

4.2Models

The engine has three models.

1. Incremental purchase effect of the next message. This is an uplift model: it estimates the change in purchase probability caused by sending, not the purchase probability itself⁷. Tree-based uplift methods split on differences in treatment effect and handle several candidate treatments at once⁸, and causal forests give heterogeneous effect estimates with valid confidence intervals⁹. The effect is modelled as a function of recent contact count and recency, so that the model can represent a response curve that flattens or turns down.
2. Disengagement hazard. A proportional hazards model estimates the instantaneous risk of unsubscribing, muting or uninstalling as a function of covariates¹⁰; here the covariates include contact volume and recency by channel. The engine reports the added hazard from one more send.
3. Value at risk. The expected future value of keeping the customer reachable on the channel, from the lifetime-value model.

4.3Decision rule

For a candidate send, the engine computes net value as: incremental purchase probability times value per purchase, minus added disengagement probability times value at risk, minus the cost of sending. It recommends suppression when net value is below zero, or when the uncertainty interval is wide and the lower bound is below zero for a customer near a contact limit. Hard limits stay in force regardless of the model: consent, quiet hours, and per-channel ceilings set by people.

4.4Targeting on response, not on activity

The engine does not suppress by low engagement alone, and it does not concentrate messages on the most active customers. A field study of retention campaigns found that customers at the highest churn risk were not necessarily the right targets, and that targeting on sensitivity to the intervention worked better than targeting on risk¹¹. We apply the same principle to frequency: the question is who responds to one more message, which is a question about treatment effects.

Table 1 maps each component to the problem it addresses and the evidence behind it.

Component	Problem it addresses	Evidence
Effect modelled as a function of recent contact	Response to repetition rises and then falls	1, 2
Frequency learned per setting, not fixed	The profit-maximising frequency differs by setting	3
Long-horizon value in the decision rule	Myopic contact policies overlook future effects	4
Uplift model for the next message	Response rates confuse effect with baseline behaviour	7, 8
Heterogeneous effects with intervals	One cap treats different customers alike	9
Proportional hazards model for disengagement	Unsubscribes and uninstalls follow contact patterns over time	10
Attention cost counted for push and SMS	Notifications are perceived as disruptive	5, 6
Targeting on response, not risk or activity	High-risk or high-activity customers are not always the right targets	11
Randomised holdouts per send type	Effects cannot be estimated from observational logs alone	12

Source: our design, with the works cited in the table.

5A worked model of the marginal message

Table 2 shows the decision rule for one customer over a week. The inputs are illustrative, not estimates: value per purchase is 40 currency units, value at risk if the customer disengages is 30 units, and the cost of a send is 0.01 units.

Message k	Incremental purchase probability (input)	Added disengagement probability (input)	Gain: uplift × 40	Loss: disengagement × 30	Net value: gain minus loss minus 0.01
1	0.020	0.001	0.80	0.03	0.76
2	0.012	0.002	0.48	0.06	0.41
3	0.007	0.004	0.28	0.12	0.15
4	0.004	0.006	0.16	0.18	-0.03
5	0.002	0.009	0.08	0.27	-0.20

Source: Worked model; inputs in the table.

Under these inputs the engine sends three messages and suppresses the fourth and fifth. A blanket cap of five would send both; a cap of two would forgo the third, which still has positive value. The point of the table is not the number three, which follows entirely from the illustrative inputs, but the shape: falling uplift and rising disengagement cross, and where they cross differs by customer. Estimating both curves per customer, with honest uncertainty, is the work the models above have to do.

6Integration with orchestration

The engine does not choose content or channel. It answers "may we send to this customer on this channel now", and returns a net-value estimate, a confidence interval and a recommended weekly envelope per channel. Upstream decision systems (journey orchestration, send-time selection, campaign tools) consult it before every send. When several campaigns compete for the same customer, the envelope is shared, so a promotion and a lifecycle message draw on the same allowance instead of each assuming the customer is untouched.

7Evaluation protocol

Before any outcome claim could be made, the engine would need to pass the following tests.

• Randomised comparison. Customers are randomised between the existing cap policy and the engine's policy, with outcomes compared over a horizon long enough to see disengagement: purchases and revenue, messages sent, unsubscribes, complaints and uninstalls. Pre-registration, sample size planning and guardrail metrics follow standard practice for online controlled experiments¹².
• Holdout design for learning. A small, persistent random share of each send type is held out, so that the uplift and hazard models are trained on effects rather than on correlations.
• Calibration. Predicted disengagement hazards are compared with observed rates on held-out periods, by channel and cohort.
• Distribution of contact. Messages per customer by cohort and region are reported, to detect suppression that falls disproportionately on some groups.

8Fairness and privacy

Suppression withholds information as well as noise. A customer suppressed for low predicted response also stops receiving offers, and if low predicted response reflects thin data, for example for a new region or a recently acquired cohort, the engine can lock in under-service. The design sets a minimum contact floor for consented service and offer messages where a cohort's estimates are uncertain, and reviews features that may proxy protected attributes. On privacy, the engine uses contact and response logs the customer's consent covers, and does not infer sensitive states from behaviour. Customers can set their own frequency and channel preferences, and those settings are hard limits, not features.

9Limitations

• No measurements. We have not measured the effect of this engine on revenue, list health or complaints. The benefits implied above are hypotheses for the evaluation protocol.
• Disengagement is partly invisible. Customers who stop reading without unsubscribing do not appear in the hazard model; non-response is a weak and delayed signal.
• Cross-channel effects. Fatigue on one channel may affect response on another. The design shares an envelope across campaigns but models disengagement per channel.
• Holdouts cost revenue. Persistent holdouts withhold messages that may have value, and small merchants may lack the volume for precise per-customer estimates.
• Value at risk is itself estimated. Errors in the lifetime-value model propagate into the suppression decision.

10Conclusion

We have described a fatigue and suppression engine that estimates, per customer, the incremental value of one more message, the added risk of disengagement and the value at stake, and suppresses sends whose net value is negative. The design rests on published work on repetition and wearout, contact policies, uplift and heterogeneous effects, and hazard models, summarised in Table 1. The next step is a randomised comparison against the existing cap policy on one channel, with disengagement tracked over a long enough horizon to see it.