The half-life of a cold-outreach reply

“When do I stop waiting on this lead?” is the kind of question most cold-outreach operators answer on gut feel, because the real answer takes data nobody has bothered to pull. An email goes out, three days pass in silence, and somebody has to call it: bump them with a follow-up, swap in a different sender, try another angle, or just move on. Most of what goes into that call is folklore.

So we pulled roughly fourteen thousand of our creator-outreach lead-campaign pairs this week and asked a tighter version of the same question: if a reply is going to come, how long does it take to arrive? The distribution is heavily front-loaded. About a quarter of all the replies we observe in a 30-day window land inside the first hour of the opener, and roughly 95% are in by day seven.

A lead that hasn’t replied yet may still reply, or may not — the clock is still running. That’s the kind of partially-observed waiting-time data Kaplan-Meier (Kaplan & Meier, 1958) was built for, so that’s what we ran. With Greenwood confidence bands, and a by-hand audit of the suspicious hour-one spike.

The data

The sample is from a recent quarter: roughly fourteen thousand lead-campaign pairs (one creator in one campaign = one row), with about two thousand human replies after we dropped auto-responders and unsubscribes. All timestamps are UTC.

The thing we’re timing is the first human reply. For leads who haven’t replied, we don’t know whether they never will or just haven’t gotten around to it, so we censor those observations at the data-pull timestamp and let the estimator handle the rest. Klein and Moeschberger (2003) is the textbook if you want the formal treatment.

How fast replies arrive

The first hour is huge. A quarter of every reply we ever see lands inside 60 minutes of the opener. That felt too fast for real humans, so we pulled the ~70 replies that arrived in under a minute and looked at them one by one. Most carried hundreds of characters of substantive text. None were auto-responders. None were bounces. Some are presumably creators hitting reply in real time; some are creator-side auto-replies with a rate card attached. Both count as engagement, so they stay in.

Half the mass is in by the end of day one. About 59% of all replies we observe in a 30-day window arrive within 24 hours of the opener. The useful operational read: if somebody hasn’t replied by this time tomorrow, they probably won’t reply this week.

The curve plateaus hard after day fourteen. Between day 14 and day 30 we pick up about 0.2 additional percentage points of reply probability, well inside the ±0.57 pp Greenwood band at day 30. Leads that haven’t replied by day fourteen are, statistically, done.

Two peaks, not one

The cumulative curve hides a more interesting shape. When we bucket the raw arrival times, they come in two peaks — a big one in the first hour, then a second one on day two.

The first peak is the opener. The second is the follow-up email landing and waking up leads who ignored the first one. There’s nothing magical about day two — that’s just where our second step tends to fall. Instantly.ai’s 2026 cross-workspace benchmark attributes roughly 58% of replies to the first touch and 42% to follow-ups (Instantly.ai, 2026). Our split on the campaigns in this sample is 61% opener, 39% follow-up — directionally the same, slightly heavier on the opener, which is about what you’d expect for shorter two-step sequences rather than longer cadences.

So a sequence with fewer than two touches is leaving close to a third of its reply potential on the table. A sequence with more than four is mostly paying for sender capacity without getting reply lift. Step two is the obvious one. Step five is the one that has to earn its keep.

Country tier: similar shape, different rate

Our previous post on this dataset looked at how many creators reply by country (UniSong, 2026). The obvious follow-up question is whether emerging-market creators reply faster than Western ones. They don’t, really. The timing shape is almost the same across tiers. What changes is how many of them reply at all.

Horizon	Western (n ≈ 3,800)	Emerging (n ≈ 950)
1 hour	3.6%	3.3%
24 hours	9.0%	11.0%
72 hours	13.2%	15.4%
7 days	14.5%	17.3%
30 days (plateau)	15.4%	18.3%

Cumulative reply probability by Kaplan-Meier, rounded to one decimal. “Western” groups US/UK/CA/DE/FR/IT/ES/AU/NZ/NL; “Emerging” groups BR/MX/ID/VN/TH/MY/PK/EG. Creators without a country tag are excluded from this table.

Every Western point is a touch lower than the matching Emerging point, and the gap widens a little at longer horizons. Both curves hit about 85% of their plateau by day three, and both are effectively flat after day fourteen. I won’t call the two shapes identical without a formal log-rank test (Mantel, 1966) — the emerging sample is smaller and its confidence band is wider — but the survival-function shape really does look close. What moves between markets is how many creators reply at all, not how fast the ones who do reply get there.

How this compares to published benchmarks

Cold B2B outreach is its own animal. Published studies of email response time tend to cover very different populations:

• Kooti et al. (2015), studying more than two million Yahoo consumer-email users, found a median reply time of 47 minutes. That’s the baseline for “somebody you actually know just emailed you.” Our cold-outreach median among repliers is around 15 hours. Roughly 19× slower.
• Yang et al. (2017), inside a large enterprise, found that messages across the organisational boundary (the closest published analogue to cold outreach) had a median reply time of 134.67 minutes and a reply rate of 2.26%. Internal enterprise email was about twice as fast (65.52 minutes) and three to four times more likely to get a reply (7.76%).

Our 15-hour median is slower than both. That tracks — cold creator outreach carries a much higher burden-of-proof on the sender side. What’s more interesting is that our reply rate (~13.7% after censoring, across this sample) sits well above the 2.26% Yang et al. saw for external enterprise email, and well above the 3.43% cross-workspace average in Instantly’s 2026 report (Instantly.ai, 2026). Creator cold outreach really does look structurally different from generic cold B2B email.

What to do with this

If a lead hasn’t replied by day three, only something like 15% of the reply potential on that lead is still in play. That’s the point where further effort — a second sender, a different offer, moving them to a nurture track — needs a better reason than hope.

Day 14 is a safe “done” line for reporting. Between day 14 and day 30 the curve moves less than a confidence interval, so a 14-day maturity cutoff for campaign reply rates holds up. We used the same cutoff in the country-reply-rate piece and it generalises cleanly.

The step-level math lines up with what operators have argued for a while: one touch isn’t enough, and past four it’s hard to justify the sender capacity you’re burning. The interesting argument is between two and four.

One more: timing shape is stable across markets, total reply rate is not. So optimise your sequence cadence once, globally, and optimise your country mix separately. Don’t try to tune cadence per-market — there’s nothing there to tune against.

Next post is the country-vs-platform logistic regression we promised last time (proper pseudo-R², odds ratios, the whole thing), then a cut by creator follower tier. If there’s a slice you’d like to see, the contact page is open. The outreach machinery this analysis runs on top of is the creator outreach service we offer brands.

---

Methodology

Unit of analysis. One lead × one campaign per row. The same creator showing up in two campaigns counts twice.

Event and censoring. The event is the timestamp of the first AI-classified human reply to a given lead, after auto-responders and unsubscribes are filtered out. Leads that haven’t replied are right-censored at the data-pull timestamp.

Estimator. Kaplan-Meier product-limit (Kaplan & Meier, 1958), computed over the hazard at each observed event time. Confidence intervals are Greenwood’s formula on the log-survival scale. Halfwidths at the reported horizons: ±0.30 pp (1h), ±0.45 pp (24h), ±0.53 pp (72h), ±0.55 pp (7 days), ±0.57 pp (14 days), ±0.57 pp (30 days).

Stratification. Country tiers come from the country metadata we capture on each lead, normalised to ISO country names and bucketed into Western (US, UK, CA, DE, FR, IT, ES, AU, NZ, NL) and Emerging (BR, MX, ID, VN, TH, MY, PK, EG). Leads without a country tag are reported separately rather than folded into either tier. We did not run a formal log-rank test on Western vs Emerging here — the claim in the post is similar shape, different rate, which is the weaker of the two statements that test would rule on. The follow-up will include the log-rank and a Cox proportional-hazards fit with robust clustering by campaign and sending account.

Hour-zero scrutiny. Replies arriving within one minute of the first send (~70 of ~1,900) were checked individually against the AI reply-classifier and average body length (600–800 characters). None were auto-responders or unsubscribes. Some are likely creator-side auto-replies carrying a rate card or media kit. Those count as engagement, so they stay. Excluding them drops the 1-hour cumulative reply probability by about 0.4 pp, which doesn’t move any of the published claims.

Sample-size reporting. Lead counts are rounded to the nearest 100 in running prose and the nearest 50 in the country-tier table. The underlying computations use the unrounded counts.

What we can’t claim. One agency, one observation window, one product-category mix. Read the numbers as a description of what the distribution looks like in our sample, not a universal benchmark for cold email. The comparisons with Kooti et al. (2015), Yang et al. (2017), and Instantly.ai (2026) are reference points, not controlled replications — the populations and email categories are different. We also can’t rule out informative censoring: leads we sent to very recently may differ from earlier leads in ways we haven’t modelled.

References

Instantly.ai. (2026, January 12). Cold email benchmark report 2026: Reply rates, deliverability and trends. Retrieved from https://instantly.ai/cold-email-benchmark-report-2026

Kaplan, E. L., & Meier, P. (1958). Nonparametric estimation from incomplete observations. Journal of the American Statistical Association, 53(282), 457–481. https://doi.org/10.1080/01621459.1958.10501452

Klein, J. P., & Moeschberger, M. L. (2003). Survival analysis: Techniques for censored and truncated data (2nd ed.). Springer.

Kooti, F., Aiello, L. M., Grbovic, M., Lerman, K., & Mantrach, A. (2015). Evolution of conversations in the age of email overload. In Proceedings of the 24th International Conference on World Wide Web (pp. 603–613). ACM. https://doi.org/10.1145/2736277.2741130

Mantel, N. (1966). Evaluation of survival data and two new rank order statistics arising in its consideration. Cancer Chemotherapy Reports, 50(3), 163–170.

UniSong Creator Studio. (2026, April 23). Country was the strongest signal in our creator cold outreach sample. https://unisong.studio/blog/creator-cold-outreach-reply-rates-by-country/

Yang, L., Dumais, S. T., Bennett, P. N., & Awadallah, A. H. (2017). Characterizing and predicting enterprise email reply behavior. In Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval (pp. 235–244). ACM. https://doi.org/10.1145/3077136.3080782