Decentralised News · Agentic Finance · Decision research
By Heath Muchena · Source review 2026-11-11 · Instrument version 1.0.0
What Matters
A subsidized purchase is evidence of a subsidized purchase. To judge durable demand, follow comparable customers after the offer ends and subtract the costs of acquiring and serving them. DN’s retention workbook separates initial volume, later contribution and unresolved observation windows. Its illustrative calculation does not establish causal uplift or identify bots in real merchant data.
Rewards, discounts and merchant-funded agent activity can increase visible transactions without creating a profitable returning customer base. The commercial question is not merely who accepted the incentive. It is whether the program changed demand enough to cover its complete incremental cost.
The compute-subsidy and reward-retention outlines are combined here. Both need a defined subsidy, a comparison population and a matured observation window. This extends DN’s sales-funnel work to post-promotion repeat behaviour and contribution, rather than repeating a lead-generation success metric.
DN Evidence Block
Sources reviewed: 2026-11-11 · Evidence type: Primary documentation review and original DN method; examples are explicitly synthetic.
Scope: A bounded merchant acquisition experiment with an explicit incentive and a fixed later observation window.
- Google’s Direct Offers announcement describes commercial offer capabilities, not an independent profit experiment: Google Marketing Live: Search ads and Direct Offers.
- The cohort design and arithmetic below are DN’s proposed method. Actual causal inference depends on assignment, comparable conditions and complete observations.
Not established by this work: No merchant cohort, bot population, subsidy response or retention uplift was measured.
Original DN instrument · version 1.0.0
DN Merchant-Retention Workbook
Estimate descriptive incremental repeat contribution from two completed cohorts. Initial-purchase net contribution is outside this narrow calculation.
Enter your assumptions or load a labelled example. Required blank numbers stay missing; unknown evidence remains unknown. Use nonpersonal labels only. This instrument makes no network requests and uses no persistent browser storage.
Use your own inputs
No current result. Complete inputs or load an illustrative example.
No bot detection, causal estimate, statistical significance or lifetime-value forecast. Entered average contribution can hide variation and must use a consistent cost boundary. The calculation assumes the same contribution per repeat buyer in both cohorts; differing averages require a separate comparison of cohort contribution totals.
1. Define exactly what the merchant is funding
A price discount, free shipping, credit grant and payment for agent compute have different cost boundaries. Record who pays, when the cost is incurred, the eligible action and any cap. Do not label ordinary platform access a merchant-funded acquisition experiment unless the commercial arrangement actually supports that description.
Google’s Direct Offers announcement is evidence that specific offer mechanisms are being described by the platform. It is not proof that a subsidy produces durable demand. DN’s experiment begins only when the merchant can state the actual incentive, its cost and the eligible population.
If the proposal pays for an agent’s compute, specify the metered activity, authorized limit and evidence that connects it to the intended demand. Payment per request can create a different incentive from payment for a completed qualifying order. This is a proposed commercial design to evaluate; the article does not claim that a named partner already offers it.
Preserve the funding rule as a versioned offer. A later increase in a cap or a change in the qualifying action creates a new treatment and should not silently inherit the first experiment’s result.
2. Choose the customer unit and the qualification rule
Choose a customer unit before counting repeats. An account, browser session, household and business buyer can produce different totals. Use an appropriate, privacy-conscious identifier and document its limits. Multiple accounts do not automatically prove abuse; a single account does not guarantee a single economic buyer.
Measure ordinary returning demand as well as suspicious activity. A reward can be unprofitable even when every participant is legitimate, and legitimate buyers can fail to return for reasons unrelated to the offer. Operational abuse classification requires its own evidence and review. A low repeat rate alone does not justify calling customers bots.
Define what makes an initial acquisition eligible and what would disqualify it. A buyer who meets the terms but makes no later purchase is different from a deliberately prohibited transaction. Apply the same evidence standard before assigning those labels; low retention alone does not establish a false identity or automated abuse.
Keep the relationship between customers, accounts and transactions visible. One business buyer can place several legitimate orders through different authorized users. A reward model that counts each as a new economic customer can overstate acquisition without any malicious participant.
3. Separate intended assignment from actual incentive exposure
Checks three and four concern assignment and comparable conditions. Where appropriate and authorized, random assignment can make the incentive comparison more interpretable. If assignment is observational, record how customers entered each cohort and what differences may explain their later behaviour. Highly motivated buyers selecting a reward are not interchangeable with everyone who did not see it.
Assignment can also be contaminated when customers share offer codes or move between treatment and comparison channels. Record exposure to the incentive, not only the label assigned at enrollment. A planned comparison may become unsuitable for a strong causal claim if the comparison group receives the same benefit. That does not make the observations useless, but it changes the conclusion they can support. Preserve the intended assignment and the observed exposure as separate facts.
Retain assignment, exposure and redemption as separate states. A customer can be assigned an offer without seeing it or see an offer without using it. Decide in advance which analysis answers the funded question, and avoid switching the denominator after the more attractive result appears.
A small authorized pilot can reveal implementation gaps before a commercial effect can be estimated. Report those gaps as operational findings rather than manufacturing a retention conclusion from a study that did not deliver the intended comparison.
4. Keep the buying conditions and outcome definition comparable
Hold product availability, geography, timing and service conditions as comparable as practical. A treatment cohort recruited during a sale and a comparison cohort observed during a stockout cannot isolate the offer. If groups face different conditions, report a descriptive difference and the confounding factors.
Predefine the outcome before reading the result: paid repeat purchase after subsidy expiry, fulfilled repeat order, or contribution collected within a stated window. Registration, coupon redemption and repeat visits may help diagnose a funnel, but they should not substitute for the commercial outcome that justified funding the experiment.
Record the alternative explanations that the comparison cannot remove: stock differences, service interruptions, seasonal demand or different buyer needs. The appropriate result may be a descriptive cohort difference with specific limits. That can still guide the next test without asserting that the incentive caused the whole difference.
Use the same qualification and outcome rules for both groups. A treated order counted at checkout and a comparison order counted only after fulfilment would make the groups unequal even before customer behavior is considered. The record should show the same endpoint and the same unresolved states.
5. Wait until both observation windows are mature
Give both cohorts the same opportunity to return. A customer recruited yesterday cannot be judged against one observed for three months. Label still-open windows and incomplete refunds instead of treating them as zero repeat demand or zero cost.
Use a fixed end date per customer or a clearly specified cohort calendar. If some buyers are still within the repeat window, report them as not yet fully observed. Excluding those buyers only after seeing that they did not return can inflate the apparent result. Similarly, a late refund can change contribution after the repeat order first appears successful. Define the maturity rule before the result, and publish how many cases remain financially unresolved.
Choose the window to match the intended commercial decision. A low-frequency product may need a different observation period from a repeat consumable; the model does not supply that business assumption. State the choice before looking at the outcome and keep the still-unobserved cases visible.
If the business must make a budget decision before the window closes, report the evidence available at that date and the uncertainty it leaves. Do not replace the missing future purchases with a confident lifetime-value estimate.
6. Include the complete incremental cost once
Include the subsidy itself, incremental support, abuse handling and any costs of the initial subsidized transactions that are not already netted elsewhere. For later orders, define contribution after the agreed fulfilment, payment, refund and service costs. Keep the initial transaction economics separate from the later retention calculation if their cost treatment differs.
Set a cost boundary for program operation: subsidy, incremental service, testing, abuse handling and other resources caused by the offer. Preserve any exclusions. Fixed costs that would exist without the program should not be casually relabeled incremental, while new committed resources should not disappear merely because they are paid on a monthly invoice.
Keep a ledger for the initial transaction and the later repeat activity. If the initial purchase contribution already subtracts its discount and fulfilment expense, do not subtract them again as program cost. If they are omitted from that contribution, include them in the complete business case. The local tool deliberately isolates repeat contribution so its boundary remains inspectable.
Track unresolved refunds and support work separately until the observation rule permits closure. Early order value is not the same as collected contribution after the costs that the offer generates.
7. State what the comparison says about ordinary returning demand
In a synthetic example, 100 treated customers produce 30 qualifying repeat buyers; 100 comparison customers produce 20. The descriptive repeat-rate difference is ten percentage points. Scaling the comparison rate to the treated cohort yields ten additional repeat buyers. The figures demonstrate the calculation and do not report an observed merchant campaign.
Average repeat contribution should use the same denominator in both groups. Contribution per repeat buyer over the window is different from contribution per repeat order. If one buyer places three orders, the model must either aggregate those orders to that buyer or use an order-based calculation consistently. Mixing the two can create artificial uplift even when the underlying cash is unchanged. Retain the underlying distribution so unusually large buyers do not disappear into an apparently typical average.
The counterfactual remains a design question. An observational comparison can contain buyers with different underlying propensity to return. Record why the comparison is informative and what it cannot establish. A randomized, properly implemented design may support stronger inference, but randomization is not a label to attach to a retrospectively assembled dashboard.
Do not convert the ten-buyer difference into a claim that ten bots were removed or ten customers were permanently retained. It is a difference at the stated endpoint and window.
8. Compare incremental repeat contribution with program cost
Assume a common 40 CU of average repeat contribution per qualifying buyer in both cohorts, measured over the same window, and 500 CU in total incremental program cost. The scenario yields 400 CU incremental repeat contribution and a net difference of minus 100 CU. It would require 12.5 incremental repeat buyers at that average contribution to cover 500 CU; whole-person planning would require at least 13 under these assumptions.
This shortcut assumes the same contribution per repeat buyer in both cohorts. If the cohorts have different average contributions, compare their separately measured repeat-contribution totals instead.
That calculation excludes any separately measured net gain from the initial purchase. If initial incremental contribution is relevant, add it explicitly in the complete business case without double counting its costs. The local tool focuses on the repeat-demand component. Its result remains descriptive unless the experiment design supports a causal interpretation.
Review whether the incentive changed the mix of purchases, support burden or refund behaviour. A higher repeat count with much smaller baskets can still reduce contribution. Likewise, a small number of valuable repeat customers may cover the program even when headline retention appears modest. Preserve the full contribution definition and inspect the distribution, not only its average.
The unchanged workbook reports this narrow scenario and checks whether the observation windows are marked complete and comparable. Its inputs are declarations, not independent verification. A positive value therefore requires review of the underlying cohort and cost evidence before it becomes a budget recommendation.
9. Agree the stop, narrow or extend rule before scaling
The ninth check is the decision rule. Agree the minimum evidence and acceptable loss budget before scaling. A program with an unresolved observation window should not be declared a success from early redemptions. A negative result may justify stopping, changing eligibility or testing a more focused buyer segment; it does not automatically identify the correct remedy.
The deliverable should be a dated cohort record: enrollment rules, assignment, offer terms, customer unit, observed outcomes, cost boundary, unresolved cases and decision. Repeating that record after a material offer change creates a useful research asset. A dashboard that only counts rewarded transactions cannot tell a merchant whether agent-assisted acquisition deserves another budget allocation.
A useful next experiment might narrow eligibility to a segment with a recurring need, reduce the subsidy or change the moment at which the offer is made. State the new hypothesis and hold other conditions as stable as practical. Do not recycle the original results as evidence for the revised program: the changed incentive may attract a different population. The maintained record should show which commercial decision each cohort actually tested.
Assign a decision owner and retain the budget boundary. The next step might be a smaller offer, a different qualifying action or an end to the program. Choose the response from the observed mechanism and commercial objective rather than treating every negative result as a reason to increase the reward.
Methodology, limitations and corrections
Specify the customer unit, incentive, outcome, cost boundary and observation window before enrollment. Record assignment and confounding differences; use permitted randomization where suitable, rather than claiming that a comparison alone removes selection bias.
Follow both cohorts through the same matured window and reconcile returns, refunds and support costs. The worked example is synthetic and demonstrates arithmetic only. Publish a descriptive result unless the design and sample support stronger inference. Do not extrapolate one repeat period into lifetime value without evidence.
For corrections, identify the claim, source and date when contacting Decentralised News. A new record-creation timestamp does not refresh its underlying evidence.
Change log: version1.0.0, 2026-11-11. Initial documentation review, original instrument and labelled examples.
Frequently asked questions
Does low retention prove reward abuse?
No. Abuse needs independent evidence. A legitimate customer may simply have no later need.
Can different-size cohorts be compared?
Rates can be scaled descriptively, as the tool does. Comparable conditions and assignment still matter.
Why exclude initial-purchase contribution?
It keeps the tool’s boundary explicit. Add that separately to the full business case and avoid counting costs twice.
When is a cohort mature?
When every included customer has had the defined observation opportunity and relevant outcomes are sufficiently resolved for the stated measure.
Primary sources
- Google Marketing Live: Search ads and Direct Offers. Announced offers are a platform description, not evidence of incremental merchant profit or a completed compute subsidy experiment. Reviewed 2026-11-11.
Turn the finding into a decision
The DN Research Desk can scope a bounded subsidy and retention-contribution evaluation, subject to source access and a written quote. Start with one unresolved decision and the records available to answer it. This article does not promise a completed audit, integration or independently verified result.
Related DN research
Disclosure: Analysis and instrument results are editorial. Where labelled affiliate links appear, DN may receive compensation when you sign up or buy through them. Commercial relationships do not determine findings. No tested integration or guaranteed outcome is implied.
