State of Email Outreach: which wording correlates with opportunities

Why this matters

Cold email is one of the few channels where a single sentence is presumed to carry measurable persuasive weight. Outreach vendors, sales trainers, and template libraries routinely recommend anchoring a pitch to a specific number: a lift, a time saved, a cost reduced. The advice sounds like it should work. Numbers read as evidence rather than assertion, and a concrete figure signals that a claim has been measured rather than guessed at.

This study tested that assumption directly against reply outcomes rather than against opinion. It measured 287,790 cold emails sent over the window 2026-01-26 to 2026-08-14, sorted into 2 outcome classes (replied, no reply) and 6 content segments describing what kind of proof element, if any, appeared in the body.

The people who should care are the ones writing the playbooks, not just the ones running individual campaigns. A sales leader deciding whether to mandate 'always include a stat' in a team's template is making a claim about population-level behavior, not about one clever email that happened to land. If quoting a percentage systematically lowers reply rates, that mandate is actively costing replies at scale, and the cost compounds every time it is copied into a new template or a new hire's onboarding doc.

It also matters because the intuition behind the advice, that specificity signals credibility, is not obviously wrong. It is plausible that a percentage helps in some contexts (a warm follow-up, a familiar sender) and hurts in others (a cold, unsolicited first touch, where a percentage can read as a canned, templated flourish rather than a genuine observation). This study cannot adjudicate every context, but it can tell us what happened, on average, across a large and heterogeneous set of real sends, which is a different and more grounded question than what sounds persuasive in the abstract.

The finding is worth taking seriously precisely because it cuts against a widely repeated piece of received wisdom. Received wisdom that goes untested tends to persist not because it is correct but because nobody measured the alternative. This is that measurement.

How the measurement works

Unit of analysis. One unit is one sent cold email, matched to whether it received a reply within the measurement window. There were 287,790 such units (287,790 rows, with 0 excluded, so every sent email that met the study's inclusion criteria is represented). An email is the right unit here because reply behavior is a property of the individual message a recipient actually reads, not of the campaign or the sender; aggregating at the campaign level would hide variation in how a given template was worded from send to send.

Outcome classes. Every unit falls into exactly one of 2 classes: replied or no reply. This is a binary, unambiguous outcome, read directly from whether a reply was recorded, not inferred from opens or clicks, which removes a layer of interpretation that plagues weaker email studies (an open does not mean a read, and a click does not mean interest).

Content segments. Each email was also coded into one of 6 content segments describing the proof element in its body: percentage_stat, named_client, case_study_number, timeframe_stat, social_proof, or no_proof. These are not mutually reinforcing categories layered on top of each other; each email was placed in the segment that best described its dominant proof element, which is what allows a clean comparison of reply rate by segment.

Groups. Beyond content segment, the data also carries 45 distinct sending groups, likely representing different campaigns, senders, or send batches. The largest single group accounts for 14.8% of all units, meaning no one group dominates the dataset. This matters for the next section: a finding that holds only because one huge, unusual group skews the average is a much weaker finding than one that holds across many groups of comparable size.

What was not done. No mitigations were applied to the raw counts (0), and no rows were excluded (0). The numbers reported are the numbers observed, not adjusted, reweighted, or filtered for outliers. That is a deliberate choice for a first-pass descriptive study: adjustments can correct for known biases, but they can also quietly launder a result. Reporting the raw observed rates keeps the finding auditable.

The headline result, and what it does and does not mean

Across the full dataset, 3,368 of 287,790 emails (1.2%) received a reply; 284,422 (98.8%) did not. Reply is a rare event overall, which is worth holding in mind before looking at differences between segments: small absolute movements in reply count can look large in relative terms.

By segment:

Emails coded percentage_stat had the lowest reply rate of the six segments. Every other proof-element segment, and even no proof element at all, out-replied it.

What this does mean. Among the emails in this dataset, over this window, quoting a percentage was associated with fewer replies than any of the alternatives tested, including saying nothing evidential at all. That is a real, specific, falsifiable pattern in observed behavior.

What this does not mean. This is an observational comparison across segments that were not randomly assigned. Emails ended up coded as percentage_stat because someone chose to write one that way, and that choice is entangled with everything else about the sender, the list, the offer, and the rest of the copy. It is possible that percentage stats are disproportionately used in weaker templates, by less experienced senders, or on colder lists, and the number itself is not the active ingredient. This study can tell you the association held at scale; it cannot, on its own, isolate the percentage as the cause. Distinguishing the two would require a controlled test that holds the rest of the email constant and varies only the presence of the stat.

Is the pattern everywhere, or is it one group

A segment-level average can be produced two very different ways. It can be a genuine property of the whole population, showing up in most groups at roughly similar magnitude, or it can be an artifact of one or two unusual groups pulling the average far from what a typical group looks like. Distinguishing these matters because the first is a robust finding you can act on broadly, and the second is a narrower, more fragile one.

The dataset carries this apart by group. Across 45 sending groups, the reply class spread as follows: the no reply class ranged from a minimum of 95.5% to a maximum of 100% across 33 groups counted, with a median of 99.2%. Correspondingly, the replied class ranged from 0% to 4.5% across 33 groups, with a median of 0.8%.

That range on the replied class, a floor of 0% and a ceiling of 4.5%, is wide relative to the overall rate of 1.2%. Some groups produced no replies at all; others replied at several times the overall average. This is normal for cold email, where sending group stands in for a tangle of factors, list quality, sender reputation, subject line, timing, that plausibly matter more than any single line of body copy.

The median group sits close to the overall figures (0.8% for replied versus 1.2% overall), which is reassuring: it suggests the population-level rate is not being dragged to an unrepresentative value by a handful of extreme groups. But the spread itself is the honest caveat. The average unit count per group, 6,395.33, is large enough to trust individual group percentages as more than noise, yet the width of the min-to-max range means that any individual sending group's experience with percentage stats could plausibly run counter to the aggregate finding. The aggregate is a real pattern across the population; it is not a guarantee about any one group within it.

Reading the segment comparison correctly

The prior section established that reply rates vary substantially by sending group, independent of content. That raises an obvious question for the headline finding: is the percentage_stat segment's low reply rate a property spread across many groups, or is it concentrated in one or two groups that happen to overuse percentage stats and also happen to reply poorly for unrelated reasons?

The segment sizes themselves offer a partial answer. The percentage_stat segment is not a small, easily-skewed slice: it contains 79,293 emails, the second-largest of the six segments after no_proof's 171,160. A pattern holding across a segment of this size, drawn from a population spread across 45 sending groups with no single group exceeding 14.8% of the total, is harder to explain away as one group's idiosyncrasy than a pattern found in, say, the 1,934-email social_proof segment or the 6,484-email timeframe_stat segment.

That said, this dataset as summarized does not cross-tabulate segment by group, so it is not possible to state directly whether percentage_stat emails are evenly distributed across all 45 groups or concentrated in a subset. That is the honest limit of what these summary figures show. A reader who wants to fully rule out the one-group explanation should ask for, or compute, the segment-by-group breakdown before treating the percentage_stat effect as uniform across every sender and list.

What can be said is that the segment's low reply rate is not being produced by a tiny sample: 602 replies out of 79,293 is a large enough base that the 0.8% figure is not a fluke of small numbers, in the way a segment with only a few dozen emails might be. Compare this to the smallest segment, social_proof at 1,934 emails: its 1.1% reply rate rests on a much thinner base (22 replies) and should be read with correspondingly more caution than the percentage_stat figure.

What the classification can, and cannot, see

This study has no separate calibration or confidence-scoring metric reported alongside the segment classifications, so this section addresses a narrower but related question: what does the structure of the classification itself let you trust, and what does it not?

The outcome classification (replied vs. no reply) is about as clean as a binary label gets: it is read from an observable event, not inferred, and is not subject to the ambiguity that plagues engagement metrics like opens or clicks. Trust in the 1.2% headline figure should be high on this dimension.

The content-segment classification is a different matter. Deciding whether an email's dominant proof element is a percentage_stat, a case_study_number, a timeframe_stat, or something else requires a judgment call, and emails can plausibly contain more than one type of proof element (a case study that also cites a percentage, for instance). How that overlap was resolved, whether by a strict priority rule, human coding, or automated detection, is not stated in the numbers available here, and it is exactly the kind of detail that changes how much weight the segment comparison can bear. A classification error rate that misassigns even a modest share of emails between percentage_stat and no_proof, for example, would compress the true gap between those two segments' reply rates toward the observed one, meaning the real underlying gap could be larger than measured. It would take a very large and systematic misclassification to invert the finding.

The honest position: treat the segment labels as descriptive of the dominant, most salient proof element in each email's body, understand that borderline cases exist, and do not treat the gap between percentage_stat (0.8%) and the next-lowest segment as precise to a tenth of a percentage point. Treat it as a real, directionally consistent gap, of a magnitude documented here, whose exact size depends on classification decisions not fully visible in the summary numbers.

What someone acting on this should actually do

Stop treating a quoted percentage as a default upgrade to a cold email. The data does not support the common vendor advice that adding a stat reliably improves reply rates; among these 287,790 emails, the opposite held. Emails that quoted a percentage replied at 0.8%, below every other segment measured, including no_proof at 1.3%.

Do not overcorrect into stripping every number from every template. The finding is an association across a large, heterogeneous set of real sends, not a controlled experiment isolating the percentage as the causal factor. A team with a genuinely strong, contextually earned percentage (one specific to the recipient, not a generic industry-benchmark figure) may see a different result than the aggregate. What the data supports is skepticism of the blanket rule, not a blanket ban.

Run your own controlled test before committing. The clean way to settle causation for your own list is an A/B test: same email, same send, same list, with and without the percentage claim, everything else held fixed. That isolates the one variable this study's observational design cannot. If your test shows the same direction, the vendor advice is wrong for your context too. If it shows the opposite, your context differs from the aggregate, and that is useful information about your specific list or offer.

Look at what beat the percentage stat, not just what lost. named_client (1%) and case_study_number (1.5%) segments outperformed percentage_stat in this data. Both are concrete and specific in a different way than a percentage: they name a real, checkable entity rather than an aggregate figure. That distinction, specificity tied to a verifiable referent versus specificity as a bare statistic, is a more promising place to focus a template rewrite than simply deleting numbers.

Do not generalize past cold, first-touch outreach. This dataset is cold emails. A percentage in a warm follow-up, a proposal, or a renewal pitch is a different communicative context and was not measured here.

Reading the published dataset yourself

The figures in this study are drawn from a dataset of 287,790 rows (0 excluded, leaving 287,790 units), classified into 2 outcome classes and 6 content segments across 45 sending groups. Anyone checking the work should look for a few specific things.

Check the denominators before trusting a percentage. A segment's reply rate is only as trustworthy as its sample size. social_proof, at 1,934 emails and 22 replies, is a thinner base than percentage_stat's 79,293 emails and 602 replies. Weight your confidence in each segment's figure accordingly.

Check group concentration. With 45 groups and a largest-group share of 14.8%, no single group can single-handedly produce the aggregate result, but the group-level spread (0% to 4.5% on the replied class) means the aggregate should not be read as describing every group uniformly.

Ask for the segment-by-group cross-tabulation if you want to rule out concentration effects. As noted earlier, this summary does not include it, and it is the single most useful additional cut for verifying that the percentage_stat effect is not an artifact of one or two atypical sending groups.

Note what has and has not been adjusted. 0 mitigations were applied to this data, and 0 rows were excluded. The numbers are raw observed counts. If you rerun this analysis on a different window or a different list, expect the exact percentages to move; the question worth re-testing is whether the ranking of segments, percentage_stat at or near the bottom, holds up, not whether the precise figures replicate to a decimal.

Treat the outcome label as reliable and the segment label as judgment-dependent. The replied/no-reply split is a direct read of an observable event. The content-segment assignment required a coding decision this summary does not fully document, and that is where a skeptical re-analysis should start.