Email recency statistics, natural purchases and incremental response https://youleftsomething.com/guides/email-recency-statistics-natural-purchases-and-incremental-response/ ## Quick answer In the 2008 Hillstrom email experiment, no-email conversion was 0.901% for recipients with one-to-three-month purchase recency and 0.295% for ten-to-twelve months. Yet all six tested differences in incremental response across recency bands had 95% intervals spanning zero. Natural purchase likelihood alone did not establish which band gained more from a message. (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026). Customers who bought recently may have a high purchase rate even if they receive no campaign. This analysis therefore keeps the randomized no-email group beside each recency band and evaluates the difference caused by assignment. It reports the uncertainty of comparisons between bands. The experiment can guide a testable targeting hypothesis without turning ordinary repeat purchases into attributed campaign sales. ## Key Takeaways In Hillstrom’s no-email arm, conversion was 67 of 7,438 recipients at one-to-three-month recency, or 0.901%, and 13 of 4,403 at ten-to-twelve months, or 0.295% (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026). The mens-email estimates across the four recency bands were 7.33, 5.03, 7.90 and 6.65 excess buyers per 1,000 assigned recipients (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026). All six unadjusted 95% intervals comparing older-band incremental effects with the one-to-three-month reference spanned zero (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026). For ten-to-twelve-month recency, mens email produced an estimated 6.65 excess buyers per 1,000, with an unadjusted 95% interval of 3.38 to 9.93 (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026). ## Will a likely repeat buyer return without an email? A recent customer can look like the safest person to contact. They know the store, they bought recently and they may already be ready for another order. Those same qualities also make them more likely to buy without a reminder. A targeting decision needs to distinguish those two sources of sales. Figure description: Email recency statistics direct comparison: mens-merchandise email exceeds womens-merchandise email by 3.694 purchasers per 1,000 assigned, approximate 95% interval 1.743 to 5.644, while recency targeting interactions remain exploratory. Figure 1: Compare the actual messages before ranking noisy recency bands. Source: Hillstrom experiment, original direct-arm replay (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026. Original analysis of the cited source. Source periods and calculation limits appear in the text. The public Hillstrom experiment randomly assigned 64,000 recipients to mens email, womens email or no email. The data also records purchase recency before assignment. This analysis groups that recency into one-to-three, four-to-six, seven-to-nine and ten-to-twelve months, then keeps the randomized comparison inside each group. (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026). ## How did natural purchases vary with recency? Among the one-to-three-month no-email recipients, 67 of 7,438 bought, or 0.901%. Among ten-to-twelve-month no-email recipients, 13 of 4,403 bought, or 0.295%. The first rate is roughly 3.05 times the second. (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026). That difference is useful for forecasting natural demand in this sample. It leaves another question open: how much of the purchase rate did the email add? Ranking the emailed groups by total purchases would carry this baseline difference into the decision. The holdout comparison accounts for it by subtracting the no-email rate in the same recency band. ## How did incremental response vary with recency? For mens email, estimated excess buyers per 1,000 recipients were 7.33, 5.03, 7.90 and 6.65 as recency moved from the newest to the oldest band. The largest point estimate appeared in seven-to-nine months. The sequence did not mirror the steady decline in natural conversion. (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026). For womens email, the corresponding estimates were 2.75, 2.54, 3.29 and 4.13. The ten-to-twelve-month band had the highest point estimate for that message. These are estimates from a single experiment, so choosing the largest number after seeing the results would overstate how much the data has established. (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026). ## What does the uncertainty allow a targeting team to conclude? The analysis compares each older band’s incremental effect with the one-to-three-month reference, separately for the two email messages. All six 95% intervals include zero. The observed differences are compatible with several underlying patterns, including no meaningful difference in the treatment effect across those bands. (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026). These are exploratory, unadjusted normal-approximation intervals. They provide a more honest reading than declaring a winning recency band from point estimates alone. A larger or repeated experiment could narrow the intervals, reveal a useful interaction or support keeping the same treatment policy across groups. ## How can recency become a testable sending rule? A merchant considering a reactivation rule can define the bands before the next experiment, randomize contact inside each band and compare added purchases or contribution. The rule should also account for message cost, unsubscribes and any commercial reason to cap contact. Those decisions need the store’s own current inputs. A worked targeting comparison illustrates the distinction. If group A would buy at 1.0% without contact and 1.5% with contact, the message adds 0.5 percentage points. If group B rises from 0.3% to 0.9%, it adds 0.6 points despite producing fewer purchases in total. The inputs are illustrative; the subtraction is the decision the holdout enables. (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026). ## How far do these experimental results travel? The evidence supports a specific warning about interpretation. A strong gradient in natural purchases can coexist with uncertain differences in incremental response. It supplies a reproducible way to examine that issue, with all arms and all four recency bands retained in the download. The campaign took place in 2008 and concerns retail email. A current checkout-recovery or iMessage program needs its own eligibility rules and randomized measurement. Use this analysis to design that test and to read its result, while keeping the historical estimates attached to their original campaign. (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026). ## Which source figures support the findings? The table retains the source measure and its population. Independent results in the next table use the stated formula; their precision follows the recorded inputs. Source measure | Value and unit | Population or denominator 1-3 months No E-Mail assigned | 7,438 recipients (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026) | Recipients in the specified randomized arm and recency band 1-3 months No E-Mail buyers | 67 buyers (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026) | Recipients in the specified randomized arm and recency band 1-3 months Mens E-Mail assigned | 7,469 recipients (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026) | Recipients in the specified randomized arm and recency band 1-3 months Mens E-Mail buyers | 122 buyers (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026) | Recipients in the specified randomized arm and recency band 1-3 months Womens E-Mail assigned | 7,486 recipients (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026) | Recipients in the specified randomized arm and recency band 1-3 months Womens E-Mail buyers | 88 buyers (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026) | Recipients in the specified randomized arm and recency band 4-6 months No E-Mail assigned | 4,766 recipients (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026) | Recipients in the specified randomized arm and recency band 4-6 months No E-Mail buyers | 26 buyers (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026) | Recipients in the specified randomized arm and recency band 4-6 months Mens E-Mail assigned | 4,673 recipients (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026) | Recipients in the specified randomized arm and recency band 4-6 months Mens E-Mail buyers | 49 buyers (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026) | Recipients in the specified randomized arm and recency band 4-6 months Womens E-Mail assigned | 4,753 recipients (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026) | Recipients in the specified randomized arm and recency band 4-6 months Womens E-Mail buyers | 38 buyers (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026) | Recipients in the specified randomized arm and recency band 7-9 months No E-Mail assigned | 4,699 recipients (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026) | Recipients in the specified randomized arm and recency band 7-9 months No E-Mail buyers | 16 buyers (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026) | Recipients in the specified randomized arm and recency band 7-9 months Mens E-Mail assigned | 4,688 recipients (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026) | Recipients in the specified randomized arm and recency band 7-9 months Mens E-Mail buyers | 53 buyers (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026) | Recipients in the specified randomized arm and recency band 7-9 months Womens E-Mail assigned | 4,627 recipients (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026) | Recipients in the specified randomized arm and recency band 7-9 months Womens E-Mail buyers | 31 buyers (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026) | Recipients in the specified randomized arm and recency band 10-12 months No E-Mail assigned | 4,403 recipients (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026) | Recipients in the specified randomized arm and recency band 10-12 months No E-Mail buyers | 13 buyers (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026) | Recipients in the specified randomized arm and recency band 10-12 months Mens E-Mail assigned | 4,477 recipients (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026) | Recipients in the specified randomized arm and recency band 10-12 months Mens E-Mail buyers | 43 buyers (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026) | Recipients in the specified randomized arm and recency band 10-12 months Womens E-Mail assigned | 4,521 recipients (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026) | Recipients in the specified randomized arm and recency band 10-12 months Womens E-Mail buyers | 32 buyers (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026) | Recipients in the specified randomized arm and recency band Recency interaction intervals containing zero | 6 intervals (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026) | Six comparisons against the 1-3-month reference band ### How were the additional comparisons calculated? These are calculations from the source inputs. Each retains its original unit and comparison population. Calculation | Result | Formula 1-3 months Mens E-Mail excess buyers | 7.33 buyers per 1,000 assigned (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026) | (122 / 7469 - 67 / 7438) * 1000 1-3 months Womens E-Mail excess buyers | 2.75 buyers per 1,000 assigned (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026) | (88 / 7486 - 67 / 7438) * 1000 4-6 months Mens E-Mail excess buyers | 5.03 buyers per 1,000 assigned (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026) | (49 / 4673 - 26 / 4766) * 1000 4-6 months Womens E-Mail excess buyers | 2.54 buyers per 1,000 assigned (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026) | (38 / 4753 - 26 / 4766) * 1000 7-9 months Mens E-Mail excess buyers | 7.9 buyers per 1,000 assigned (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026) | (53 / 4688 - 16 / 4699) * 1000 7-9 months Womens E-Mail excess buyers | 3.29 buyers per 1,000 assigned (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026) | (31 / 4627 - 16 / 4699) * 1000 10-12 months Mens E-Mail excess buyers | 6.65 buyers per 1,000 assigned (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026) | (43 / 4477 - 13 / 4403) * 1000 10-12 months Womens E-Mail excess buyers | 4.13 buyers per 1,000 assigned (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026) | (32 / 4521 - 13 / 4403) * 1000 Recent natural conversion rate | 0.9 % (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026) | 67 / 7438 * 100 Less-recent natural conversion rate | 0.3 % (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026) | 13 / 4403 * 100 Natural conversion rate ratio | 3.05 ratio (Hillstrom experiment, independent analysis (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026) | (67 / 7438) / (13 / 4403) ## Is the message comparison clearer than the recency ranking? Across all assigned customers, mens-merchandise email exceeded womens-merchandise email by 3.69 purchasers per 1,000, with an exploratory unadjusted 95% interval of 1.74 to 5.64. The comparison uses 267 of 21,307 assigned customers versus 189 of 21,387, directly comparing two randomized messages rather than subtracting two separately displayed holdout results. (Hillstrom experiment, original direct-message comparison (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html), Nov 2026). That result concerns these two historical messages in this audience. Their labels identify merchandise, not recipient gender. It does not establish that one category always works better, but it suggests a different immediate decision from selecting a recency band after inspecting four noisy estimates. The existing recency interaction intervals do not establish a winning band, while the whole-audience message comparison estimates a positive difference in this campaign. ### Why should the common holdout not be counted twice? The contrast between two email effects is (mens rate − no-email rate) − (womens rate − no-email rate). The common no-email term cancels, leaving the direct message-rate difference. Treating the two holdout comparisons as independent and adding their variances would include the same control uncertainty twice. The direct comparison uses only the two randomized message arms' binomial variances. This is useful when a merchant compares several recovery variants against one holdout. The holdout is essential for whether contact adds purchases. It is not a second source of uncertainty in a contrast where the same control subtracts away. Preserve the actual arm counts and the precise contrast before calculating an interval. ### Does either message need a proven recency interaction to be useful? No. A message can add purchases across several bands without evidence that its effect differs between them. An inconclusive interaction therefore does not imply the message has no effect within each band. Conversely, a recent group's high total conversion does not establish that its incremental effect exceeds older groups. A local decision can separate message selection from segmentation. First define the eligible audience and compare the actual messages with persistent random assignment. Then assess a predeclared recency interaction if a differentiated policy would materially change the decision. If the evidence cannot resolve that interaction, retaining a simpler common policy can be a defensible choice rather than inventing a precise targeting hierarchy. ### What should be checked before adopting that simpler policy? Keep contribution, contact cost, opt-outs and contact-frequency constraints beside the purchase contrast. The public file has no message-exposure or click field, and the campaign lasts two weeks. A stronger purchase response in 2008 does not establish current deliverability, margin or a long-run response. A new store's category mix and creative content may produce a different result. The direct message comparisons by recency (https://youleftsomething.com/guides/email-recency-statistics-natural-purchases-and-incremental-response/data/message-vs-recency.csv) and raw-data reconstruction (https://youleftsomething.com/guides/email-recency-statistics-natural-purchases-and-incremental-response/analysis.py) retain the whole audience and all four bands. The intervals are exploratory and unadjusted. They make the choice of contrast inspectable without declaring a current merchant's message or targeting winner. ## Methodology and Sources All 64,000 assigned recipients. Predeclared recency bands of 1 to 3, 4 to 6, 7 to 9 and 10 to 12 months. Contrast each email arm against no email within band, then compare subgroup effects to 1 to 3 months using independent-binomial standard errors. One 2008 randomized retail email campaign, not an abandoned-checkout or iMessage experiment. Subgroup estimates are exploratory with unadjusted Wald intervals. No subgroup is declared a proven winner. Recorded purchase behavior and spend are the campaign outcomes. Actual contribution margin and contact cost are absent. Derived percentages are rounded to two decimal places. The source period is 2008 randomized retail email campaign. Independent calculations were completed in November 2026 and use the formulas shown above. ## Glossary Random assignment: Allocation to a message or comparison group through a randomized process. Holdout: The comparison group that did not receive the campaign email. Incremental response: The difference in purchase response between assigned groups. Recency: Time since the last recorded purchase. Purchase profile: A classification based on the customer’s recorded category history. Confidence interval: An interval describing uncertainty around an estimated difference. Percentage point: The absolute difference between two percentages. Interaction: A comparison of treatment differences between customer groups. ## Sources Hillstrom experiment: MineThatData E-Mail Analytics and Data Mining Challenge (https://blog.minethatdata.com/2008/03/minethatdata-e-mail-analytics-and-data.html) Check the source figures The data file retains source labels, values, populations and limitations for this analysis.