Reach, frequency and media pressure: no universal threshold

Why a single threshold is the wrong question

A frequency cap only means something alongside a time window, a population, a budget, an objective and a measurement method. The current corpus does not pit two competing values against each other: it studies an advertising model under a privacy constraint and a dataset of Reach & Frequency contracts. The defensible conclusion is therefore to organise the setting, not to publish a number presented as valid everywhere.^[1][2]

An operational decision must keep four objects in view: the reach expected at each budget level, the frequency being tested, marginal pressure and the privacy constraint. This grid is an original editorial construction. It helps document a setting, but it is neither an algorithm validated by the authors nor an automatic platform recommendation.^[1][2]

Privacy shifts the performance frontier

The preprint on k-anonymity reports a significant drop in performance as soon as privacy is introduced into the model under study. The experiments are designed to examine the trade-off between protecting users and the effectiveness of online brand advertising. This evidence concerns the authors' model and experiments; it does not set an expected loss for any real campaign.^[1]

The documented case is useful because it prevents privacy from being treated as a mere legal constraint bolted on after optimisation. Privacy is part of the measurement system and can move the observed trade-off. Before changing a cap, the team must therefore state the privacy regime, the signals still available and the metric that will remain comparable across the variants tested.^[1]

Budget–performance curves replace the isolated number

The R&F-Inventory dataset links several budget levels to observations of unique users and page views within a single campaign context. It is presented as a large-scale dataset for inventory estimation on Reach & Frequency contracts. Its contribution is methodological: observing a budget–performance curve rather than a single forecast point.^[2]

This case does not provide a table of settings transferable as-is to every platform. It shows the kind of data needed to compare options: several budgets, observations of reach and volume, and the contractual context of targeting, scheduling and frequency control. A team can reproduce this structure on its own data without claiming the same results.^[2]

The privacy trade-off has a measured magnitude

Under the definition of ROAS used by the authors, moving from k = 1 to k = 2 comes with a 33 % drop. That number does not describe every campaign, nor a general rule across platforms. It belongs to a specific experimental setup, with its data, its optimisation function and its metric. Its significance is nonetheless major: it shows that a privacy constraint can materially shift the performance frontier. Protection must therefore not be added after the calculation; it is one of the conditions of the compared result.^[1]

The precision of 33 % could create an illusion of transferability. Yet the authors' ROAS, the level of k-anonymity, the auction mechanism and the campaign portfolio form an inseparable whole. A team must not subtract 33 % from its own forecast. It can use this result to justify a test that keeps the same population, the same window and the same economic definition across variants. Without that comparability, an observed variation does not allow the gap to be attributed to the privacy constraint alone.^[1]

Ten campaigns to compare four approaches

The authors evaluate four approaches to reach measurement and optimisation across ten advertising campaigns. This strengthens the value of the case: the result does not come from a single abstract simulation, but from a repeated comparison on platform data. It is not enough, however, to constitute an industry standard. Ten campaigns can cover varied configurations without representing every channel, objective or market. The analysis therefore retains the comparative protocol and the reality of the field, not a universal recommendation on settings.^[1]

To reproduce the spirit of this test, a team can define several variants before launch, keep the same measurement conventions and document which campaigns are included or excluded. It must record budget, reach, frequency, impressions, privacy regime and the economic metric. The comparison then bears on the distribution of results, not only on the best campaign. A setting becomes a rule only if its advantage remains observable across sufficiently diverse contexts and under the same controls.^[1]

Monotone constraints protect the coherence of the curves

The work on R&F-Inventory compares models and indicates that explicit monotone constraints regularly outperform unconstrained or weakly constrained models, both in predictive accuracy and in structural coherence. That property matters in a budget–reach problem: a curve that behaves incoherently can propose a setting that is mathematically advantageous but operationally implausible. The result does not enshrine a single algorithm; within the authors' benchmark, it establishes the value of constraints aligned with the expected structure of the phenomenon.^[2]

For the tactical decision-maker, structural coherence becomes a preliminary check. A budget–unique-users curve must be inspected before it is optimised: units, expected monotonicity, saturation, breaks and the density of observations. A better average error is not enough if the shape leads to recommending unstable budgets. This critical reading matters especially where data are sparse at the extremes. The model helps choose a testing range; it replaces neither the experiment nor the control of the pressure actually experienced by users.^[2]

A protocol of six decisions before touching the setting

The synthesis leads to six documented decisions: choose the population, fix the window, define the economic metric, state the privacy regime, plot the budget–reach curve and select the frequency variants. Stopping rules and the return to the previous setting are written before the test. The ten campaigns of the first paper justify the multi-case comparison; the dataset of the second justifies observing several budget levels. The protocol brings these contributions together without claiming that their results can be added up or reproduced identically.^[2][1]

The final report must separate observed result, uncertainty and decision. For each variant it states reach, impressions, frequency, cost and the definition of ROAS, then explains the choice made. If the data do not allow the variants to be compared under the same privacy regime, the absence of comparable data does not demonstrate an absence of effect. If the curve contradicts the expected structural constraints, the model is re-examined. This discipline prevents an isolated figure or a platform threshold from being turned into a rule without context.^[1][2]

Fatigue appears as a curve, not as a fixed threshold

The study published in 2026 analyses 387 Chinese consumers who watched live streaming of agricultural products. In this context, purchase intention rises at low to moderate exposure levels, then falls once exposure becomes high. The model links this turning point to two simultaneous mechanisms: authenticity and cultural identity can sustain the response, while psychological saturation holds it back. This result documents a non-linear shape in a specific context; it delivers neither a universal optimal frequency nor causality established by controlled experiment.^[3]

The multigroup analysis reinforces operational caution. The authors conclude that negative marginal utility appears earlier among heavily exposed consumers. They compare 250 people watching at least three sessions a week with 137 less frequent viewers. A media team cannot convert these groups directly into campaign caps. It can, however, segment exposure history, look for a turning point specific to each population and plan a return to the previous setting if the increment deteriorates.^[3]

A tuning procedure, not a magic value

Allianz Spain provides a case of tactical reconfiguration: the insurer confirms using Performance Max with Jellyfish for its motor campaigns, while Google reports 15 % more qualified leads at a cost per lead below that of generic Search campaigns. The figure remains a measurement published by the platform, with no independent audit. The case therefore requires documenting the reference campaign type, the definition of a qualified lead, the measurement window and any other changes made, before attributing the gap to the campaign setting alone.^[6][7]

Start by plotting the unique users expected at several budget levels. Then test several caps while keeping the same window and the same measurement perimeter. Finally, compare the marginal gain with costs, pressure and privacy constraints. This procedure turns the five papers into an execution grid; it remains an editorial proposal to be validated on the campaign's own data.^[1][2]

An increase in frequency should be retained only if the increment remains observable in the target segment and if the comparison respects the same privacy conditions. When the signals are no longer comparable, the right conclusion is not zero effect but evidence unavailable. The setting must then remain experimental, reversible and accompanied by a next review date.^[1][2]

Going further: media pressure and budget allocation

The Alibaba framework brings several business constraints into a single optimisation problem. The Ant Group work adds a joint reading of seven-day activity and coupon cost. Together, these two papers show why a setting cannot be evaluated through one conversion or one isolated cost. The decision function, the constraints and the measurement horizon belong to the very definition of the test.^[4][5]

For a campaign, this means plotting immediate response and delayed response separately, at several levels of spend or pressure. Windows, units and attribution rules are fixed before the comparison. A variant that lowers cost today but degrades later response must not be classed as the winner by default. Both publications justify this structural check; they provide neither a universal curve nor a universal lag to copy.^[4][5]

What these results do not allow you to conclude

The corpus brings together one preprint, two recent works and two industry studies that describe specific contexts. The corpus measures neither universal advertising fatigue, nor an optimal threshold common to all channels, nor causality transferable to a given campaign. It allows a decision and a test protocol to be structured, without turning situated results into a universal rule.^[1][2]