The Outreach Metric Dictionary: Numerator, Denominator, Window, Distortion
Look up any outreach metric, write its numerator, denominator, window, and event source, and publish a dictionary two analysts can reproduce from one export.
Key takeaways
- In one published campaign of 1,000 delivered emails, changing only which replies count moves the reply rate from 8.0% to 1.8%, a 4.4x swing on identical data.
- Auto-replies were 51% of all replies across 6,074,430 cold emails in Growsurely's benchmark, so a reply rate that does not exclude them is about half machine output.
- Gmail and Yahoo divide spam complaints by inbox-delivered mail only, so heavier spam filtering can lower the reported complaint rate. Gmail's target is below 0.10% and its ceiling is 0.30%.
- Apple Mail Privacy Protection triggered 10.5% of total opens but 18.2% of unique opens in SendGrid's data, so per-person open rates are distorted more than raw open counts.
- LinkedIn's InMail response rate counts declines and Not Interested clicks as responses on a 30-day window, so a seat can clear the 13% policy floor without one interested candidate.
- In Refolk's index there are about 39 US sourcer titles for every recruiting-operations title, so tool defaults become team definitions unless an ops owner writes them down.
This is a metric dictionary for teams that do person-level outreach: recruiters, sourcers, SDRs, and the ops owner who has to stand behind the dashboard. It covers about twenty metrics across email, LinkedIn messaging, and phone. For each one it gives the numerator, denominator, counting window, and event source, and the specific ways the number gets inflated or deflated. If you paste the rows into team policy, two analysts working from the same export should get the same number.
What does a complete outreach metric definition contain?
A complete definition names five things: the unit (person or message), the numerator event, the denominator, the counting window, and the system of record that emits the event. Leave any one of them blank and two honest analysts will get different numbers from the same export.
The terms, once each:
- Unit: whether you count people ("did this person reply at least once") or messages ("how many messages drew a reply").
- Numerator: the class of event you count, such as a human reply, a positive reply, or a hard bounce.
- Denominator: what you divide by. Usually that is sent, delivered, or contacted, and you need to say how bounces are handled.
- Counting window: how many days after first contact an event still counts.
- Cohort: the group a person belongs to, keyed on first-contact date. Someone contacted in week 12 stays in week 12's denominator even if they reply in week 15.
- Event source: the one system whose records are authoritative for the event, such as the sequencer, the ATS, LinkedIn Recruiter, or Postmaster Tools.
The five layers of a metric definition
- Event sourceThe system that emits the raw event: sequencer, ATS, LinkedIn, Postmaster
- Event classWhat the event really records: human reply, auto-reply, machine open, scanner click
- UnitPer person or per message
- DenominatorSent, delivered, or contacted, with bounce handling stated
- WindowCohort by first-contact date, fixed number of days to count
Use this skeleton for every row in the dictionary. A row with a blank line is provisional and should not go on a shared dashboard.
Metric name: Unit: per person | per message Numerator (event class): Denominator: Bounce handling: Counting window (days): Cohort key: first-contact date Event source (system of record): Machine-event rule: Departs from platform-native definition: yes | no - how: Known inflation: Known deflation: Owner: Last reconciled (date, analysts):
One entry per metric. Fill every line or mark the metric provisional.
Why does the same campaign report 8% or 1.8%?
Because the reply class, the denominator, and the unit are all choices, and the reply class alone moves one real campaign's rate by 4.4 times. A benchmark that quotes a target without stating those choices can't be compared with your number.
Overloop published the inputs for one campaign: 1,000 emails delivered, 18 positive replies, 7 referrals, 25 flat negatives, and 30 out-of-office auto-replies. Here is what five defensible definitions do with those inputs.
| Definition | Numerator | Rate |
|---|---|---|
| All replies ÷ delivered | 80 | 8.0% |
| Human replies ÷ delivered | 50 | 5.0% |
| Positive + referral ÷ delivered | 25 | 2.5% |
| Positive only ÷ delivered | 18 | 1.8% |
| All replies ÷ 3,000 messages, 3-step sequence (illustrative) | 80 | 2.7% |
One campaign, narrowed by reply class
- 1,000Delivered
Overloop's published example
- 80All replies
Includes 30 out-of-office auto-replies
- 50Human replies
Auto-replies removed
- 25Positive + referral
Negatives removed
- 18Positive only
Referrals removed
The unit causes its own distortion. Growsurely puts it cleanly: three emails to one person who replies once is a 100% prospect reply rate and a 33% email reply rate. When a team mixes per-step and per-person counting, multi-step sequences look worse than single sends, because the follow-ups add messages to the denominator without adding people.
Reply rate vs response rate
Treat these as different words. In the dictionary, keep "response rate" for platform-native metrics, where the platform decides what counts. LinkedIn, for example, counts a decline as a response. Keep "reply rate" for metrics you define yourself, where the numerator is a written reply you have classified.
Reading external benchmarks
Apollo reports a Sales.co figure of a 2.09% average cold email reply rate, "with only 14.1% of those being genuinely positive." That 14.1% is a share of replies, not of sends. Multiplied through, it comes to about 0.3% positive. I could not confirm whether the study divided by sent or delivered. Before you compare an outside figure with yours, rewrite it in your dictionary's terms.
Who sets outreach metric definitions on a team?
In practice it's the default of whichever tool emits the event. The people who own definitions are vastly outnumbered by the people who produce the numbers, so a definition nobody writes down becomes whatever the sequencer computes.
Refolk's index of professional profiles shows how thin that ownership layer is:
| Population | Market | Count | Most common title |
|---|---|---|---|
| Recruiting ops titles | US | 147 | Recruiting Operations |
| Recruiting ops titles | UK | 19 | Recruitment Operations |
| Sourcer titles | US | 5,735 | Sourcing Specialist |
| SDR/BDR titles | US | 47,256 | Sales Development Representative |
These are title counts, not reporting lines, and the "most common title" column comes from samples of 25 people (US) and 19 (UK). They still show the mechanism. With roughly one definition owner for every few dozen producers, each person's or tool's default becomes the de facto standard.
What the tools and platforms define by default
| Metric | Denominator | Window | What it counts |
|---|---|---|---|
| Gmail spam rate | Inbox-delivered | Daily | User reports |
| Yahoo spam rate | Inbox-delivered | Not stated | User reports |
| LinkedIn InMail response | InMails sent | 30 days | Reply, accept, Not Interested, Quick Reply |
| LinkedIn credit-back | Per InMail | 90 days | Accept, decline, reply |
| Reply.io reply rate | Contacted (excludes bounce/OOO/auto) | Not stated | Human replies |
Sources disagree on denominators. Reply.io documents "Reply rate = Replies × 100 ÷ Contacted," with Contacted defined as contacts who received emails and are not in Bounced, Out of Office, or Auto Replied status. Warmerly divides by delivered. Outreach detects out-of-office replies "based on specific phrases which appear in either the Subject or Body," and its "Prospect replies" trigger also fires when the prospect answers a call made as part of the sequence. I could not establish the Salesloft and Apollo defaults from primary documentation. To check yours locally, export ten prospects and recompute the tool's displayed rate by hand.
Comparing your dictionary with peers
Before you publish, it helps to compare your dictionary with ops owners at similar companies. In Refolk's index the US top employers for these titles include Plaid, Perplexity, Harvey, Hudson River Trading, and Susquehanna. In the UK they include Anthropic, Jane Street, British Airways, and Bauer Media Group. The title vocabulary splits by country: "Recruiting Operations" was 16 of 25 in the US sample, and "Recruitment Operations" was 12 of 19 in the UK. A keyword search on one spelling misses half the population. Refolk takes the request in plain English, so you can describe the role and the reporting work rather than guess the spelling.
Email delivery and spam complaint rate definitions
Delivery metrics are counted per message, with sent as the denominator. Spam complaint rate is the one you can't compute exactly yourself, because the mailbox providers divide by inbox-delivered mail, which you never see.
| Metric | Numerator ÷ Denominator | Window | Event source |
|---|---|---|---|
| Delivered rate | Delivered ÷ sent | Per send day | Sending service |
| Hard bounce rate | Hard bounces ÷ sent | Per send day | Sending service |
| Unsubscribe rate | Unsubscribed persons ÷ delivered persons | Team cohort window | Sequencer |
| Gmail spam rate | Reports ÷ inbox-delivered | Daily | Postmaster Tools |
| Internal complaint rate | Feedback-loop reports ÷ delivered | Per send day | Provider feedback loops |
How to calculate spam complaint rate
- Provider version: user spam reports ÷ mail delivered to the inbox. Gmail calculates this daily. Yahoo states the same inbox-delivered denominator. You read these figures off the provider's tools and don't recompute them.
- Internal version: complaints received through feedback loops ÷ delivered. It reads lower than the provider's figure for the same complaints, because delivered includes mail that went to spam. Label it as a proxy.
The thresholds, as each provider states them:
- Gmail: keep the rate reported in Postmaster Tools below 0.10% and never reach 0.30%. Secondary sources quote Google's FAQ as saying that above 0.3%, a sender is not eligible for mitigation until the rate stays below 0.3% for 7 consecutive days.
- Yahoo: keep the rate below 0.3%. Yahoo runs a Complaint Feedback Loop.
- Microsoft: has not published a numeric threshold. Its monitoring tools are SNDS and JMRP.
- Bulk sender line: all three providers apply their bulk requirements at 5,000 or more messages per day. Enforcement dates have changed over time, so re-check each provider's sender guidelines page rather than relying on a remembered date.
Postmaster Tools covers mail to personal Gmail accounts, and only when volume is high enough. Cold email to Google Workspace business addresses doesn't show up there, which matters a great deal for recruiting teams writing to work addresses.
Are email open and click rates still accurate?
Opens and clicks record that a machine fetched something, not that a person paid attention. Define them as raw event counts, report machine events in their own row, and never use them as a success metric for person-level outreach.
| Metric | Numerator ÷ Denominator | Window | Event source |
|---|---|---|---|
| Raw open rate | Persons with any open ÷ delivered persons | Cohort | Sending service |
| Machine-open share | Flagged opens ÷ all opens | Cohort | Machine-open flag |
| Unflagged open rate | Persons with an unflagged open ÷ delivered persons | Cohort | Sending service |
| Raw click rate | Persons with any click ÷ delivered persons | Cohort | Sending service |
| Filtered click rate | Persons with a rule-passing click ÷ delivered persons | Cohort | Sending service + rules |
What machine opens do
SendGrid found that Apple Mail Privacy Protection (MPP, which prefetches mail on the recipient's behalf) "triggered 10.5% of the total opens and 18.2% of the unique opens" it processed after launch. Unique opens take the bigger hit because MPP fires one open for each opted-in recipient, including people who would never have opened the email. Per-person open rates are therefore distorted more than raw counts.
The flags each sending service gives you are partial:
- SendGrid's
sg_machine_openfield in the Event Webhook is limited to Apple machine opens. SendGrid also notes that a user-initiated open may still register assg_machine_open: true. - Mailgun's
client-info.botfield takes the values apple, gmail, generic, or empty.
So report flagged opens separately rather than deleting them. Deleting them deflates human opens, and keeping them in inflates unique opens.
What scanner clicks do
Security scanners at the recipient's company follow links before a person sees the message. As reported second-hand, Omnivery found that at large companies 80 to 95 percent of clicks came from security scanners. SendGrid "does not distinguish between legitimate and illegitimate clicks," but it does document the signatures. A machine-event rule should exclude a click when any of these hold:
- It arrives at approximately the same time the message is delivered. One open-source rule set uses about 30 seconds from send.
- The same IP appears across many clicks.
- The user agent matches a known scanner.
- Every link in the message was clicked, including unsubscribe.
No industry working group has published a figure for scanner share. If nobody on your team can write and test these rules, you need a deliverability engineer, and Refolk can find people who have worked on sending infrastructure from a one-sentence request.
How to calculate positive reply rate and the other reply metrics
Positive reply rate is the number of people whose human reply shows interest in a conversation, divided by delivered persons in the first-contact cohort. Every other reply class gets its own row so nothing gets quietly folded in.
Example values below use Overloop's inputs and assume one email per person, so persons and messages coincide.
| Metric | Numerator ÷ Denominator | Overloop example |
|---|---|---|
| All-reply rate | Persons with any reply ÷ delivered persons | 8.0% |
| Human reply rate | Persons with a human reply ÷ delivered persons | 5.0% |
| Auto-reply share | Auto/OOO replies ÷ all replies | 37.5% |
| Positive reply rate | Persons classed positive ÷ delivered persons | 1.8% |
| Referral rate | Persons classed referral ÷ delivered persons | 0.7% |
| Negative rate | Persons classed negative ÷ delivered persons | 2.5% |
Auto-replies are the biggest single distortion. Across 6,074,430 emails, Growsurely found "auto-replies are 51% of all replies." Instantly's figure, reported second-hand, was 45.1%. Practitioners advise that you "classify every reply by hand or with a reviewed rule". Phrase matching alone, which is how Outreach detects out-of-office replies, will misfile some.
My position on denominators: take auto-replies out of the human-reply numerator, but keep the person in the denominator. Reply.io's default removes auto-replied contacts from Contacted, which raises the rate. If you adopt that default, write it down as a departure.
Classifying one reply into one person-level class
- Machine checkArrived within minutes, or matches OOO phrases in subject or body? Class auto/OOO
- Left-company checkAuto or human reply says the person has left? Class left-company
- Intent classPositive, referral, negative, or unsubscribe, by reviewed rule
- Person rollupPerson takes the class of their latest human reply inside the window
- CountNumerators read person classes, never raw reply counts
AUTO/OOO - machine-generated; out-of-office, vacation, or system notice. Not a human reply. LEFT-COMPANY - any reply stating the person has left. Not a human reply. Route to re-addressing. POSITIVE - a human agrees to talk, asks for times, or asks a question that implies interest. REFERRAL - a human names or forwards to another person. Counted separately from positive. NEGATIVE - a human declines, says not now, or says wrong fit. UNSUBSCRIBE - a human asks to stop contact. Also counted in unsubscribe rate. Suppress. ROLLUP - one class per person: the latest human reply inside the counting window. TIE-BREAK - if unclear, a team lead classifies; the rule is updated, not overridden.
Paste into policy. Add two real example replies under each class before sign-off.
Platform messaging, phone, and meeting rate definitions
Platform-native metrics like LinkedIn's InMail response rate are set by the platform, count declines as responses, and run on their own clocks. Record them verbatim in the dictionary and add a separate human-interest row beside each one.
LinkedIn's definition: "An InMail response is defined as any InMail to which the recipient sends an InMail reply, accepts the InMail invitation, or replies as 'Not Interested.'" Quick Replies also count. Third-party summaries say auto-replies count on Sales Navigator. I have not seen that on a primary page, so check it in your own exports.
| Metric | Numerator ÷ Denominator | Window | Event source |
|---|---|---|---|
| InMail response rate (native) | Responses ÷ InMails sent | 30 days | LinkedIn Recruiter |
| InMail accept rate (team) | Accepts ÷ InMails sent | 30 days | LinkedIn export |
| InMail decline share (team) | Declines + Not Interested ÷ responses | 30 days | LinkedIn export |
| InMail human interest (team) | Persons classed positive ÷ persons sent | Team window | LinkedIn export + taxonomy |
LinkedIn's three clocks
- Response rate: only responses within the first 30 days after the send date count. This is quoted second-hand; I did not find the primary page.
- Credit-back: you get a credit back for every sent InMail that receives a response within 90 days of the send date.
- Policy floor: at or above 13% on 100 or more InMails sent within every 14-day assessment period. Falling below it leads to a warning, and after that bulk InMail is disabled for 2 weeks.
The clocks don't line up. A response on day 45 is a "no response" in the rate but still earns its credit back. A pending InMail counts as neither accepted nor declined, so it stays in the denominator as a non-response until the window closes.
Phone, meeting, and ATS rows
| Metric | Numerator ÷ Denominator | Unit | Event source |
|---|---|---|---|
| Connect rate | Dials answered by a person ÷ dials | Per dial | Dialer |
| Person reach rate | Persons reached live ÷ persons dialed | Per person | Dialer |
| Meeting rate | Persons with a held meeting ÷ delivered persons | Per person | Calendar or CRM |
| Reply-to-meeting | Persons with a held meeting ÷ persons classed positive | Per person | CRM |
| Sourced-to-screen | Candidates reaching screen ÷ candidates sourced | Per person | ATS activity feed |
Count meetings as held, and keep booked as a separate row. If a sequencer's "replied" status can fire on an answered call, as Outreach's trigger does, attribute the outcome to a channel before it reaches the meeting row.
There are no published sourcing funnel metric definitions from any talent-acquisition operations body. In Greenhouse, stage names are custom per job, so one requisition says "Phone Screen" and another says "Recruiter Screen." Map every local stage name to a canonical stage in the dictionary. Watch the top of the funnel too. One practitioner page merges "applied or sourced" as "they applied, or replied to outreach," which means "sourced" sometimes already includes a reply. Ariana Moon, a senior recruiter at Greenhouse, has written about conversion rates that look too high, and a merge like that is a likely cause.
How do outreach metrics mislead? Failure modes and false positives
Most metric errors come from eight repeatable mechanisms. Each one leaves a signature you can check in a single sitting.
- Complaint rate looks healthy when it isn't. Signature: the rate falls while Postmaster shows Gmail is automatically sending a significant share of your messages to spam. Check: read complaint rate next to the delivery-error and compliance panels, never on its own.
- Postmaster blind spot. Signature: "no data" for a team that mostly writes to work addresses. Workspace recipients don't appear, so complaints can be happening unseen. Check: segment sends by recipient domain type and treat the business-domain share as unmonitored.
- Machine-open flag over-counts. Signature: human open rate collapses after someone deletes every flagged open. SendGrid notes that a user-initiated open may still register as a machine open. Check: report machine opens in their own row.
- Scanner clicks. Signature: clicks within seconds of delivery, from one shared IP, on every link including unsubscribe. Check: apply the timing, IP, and user-agent rules before any click reaches a dashboard.
- InMail response rate read as interest. Signature: a healthy response rate with no booked screens. A seat can clear the 13% floor on Not Interested clicks alone. Check: split responses into accept, decline, and reply.
- Seat-transfer contamination. Signature: a new recruiter's rate starts well above or below where their activity would put it. LinkedIn states that "the new seat holder will be accountable for the previous recruiter's response rate." Check: note the seat transfer date on the dashboard and start a fresh team-defined series from it.
- Auto-replies counted as replies. Signature: reply rate is roughly double the human reply rate, which fits auto-replies being 51% of the reply folder. Check: measure the share of replies that arrive within minutes and match OOO phrases.
- Stage drift. Signature: conversion rates that can't be compared across jobs, or that never change historically. Custom stage names break comparisons, and Greenhouse's Harvest API returns each candidate's current stage only. Check: pull historical transitions from the activity feed, not the current-stage field.
A complaint rate that falls while placement worsens is not good news. It means fewer people are able to complain.
How to build, test, and keep the dictionary current
Build the dictionary in eight steps that end in a reconciliation test: two analysts, one export, the same numbers. Then re-check it whenever a platform changes a threshold or a window, because several rows inherit their definition from the platform.
Build and publish the team metric dictionary
- List metrics and their system of recordWrite down every dashboard metric and the one system that emits its events. Done when no metric has two candidate sources.
- Fix the unit and the cohortMark each metric per person or per message and key cohorts on first-contact date. Done when every row names its unit.
- Set the denominatorsChoose sent, delivered, or contacted and record bounce handling. Done when departures from tool defaults are written down.
- Write the reply taxonomyDefine human, auto/OOO, left-company, positive, referral, negative, and unsubscribe, with a reviewed rule. Done when team leads sign off on examples.
- Write machine-event rulesSpecify machine-open flags and scanner-click timing, IP, and user-agent rules. Done when machine events land in their own rows.
- Align counting windows with platform clocksSet a window in days per metric and note LinkedIn's 30, 90, and 14-day clocks and Gmail's daily rate. Done when every row has a window.
- Document departures from native definitionsLabel each platform-native metric and how it differs from the team version. Done when no tile mixes the two unlabelled.
- Run the two-analyst reconciliation testTwo analysts compute every metric from one raw export separately. Done when the numbers match or every gap is traced to a wording fix.
Running the reconciliation test
Give both analysts the dictionary and a raw event export, not a pre-aggregated report. A mismatch almost always traces to one of three things: a person counted in two cohorts, a reply class applied differently, or a window counted from send date in one place and first-contact date in another. Fix the wording, not the spreadsheet, then rerun.
Keeping it current
- Re-read the provider pages behind the threshold rows (Google's sender guidelines, Yahoo's best practices, LinkedIn's InMail policy) on a fixed review cadence. Those are the rows most likely to change without notice.
- Version the dictionary. Every dashboard should show which version it computes against.
- When a sequencer or ATS updates its metric help pages, recompute ten prospects by hand to confirm the default hasn't moved.
Before the dictionary goes into policy
- Every row states per person or per message.
- Every row names exactly one event source.
- Every denominator can be known from delivery events alone, or the exception is documented.
- Auto-replies, referrals, and negatives each have their own row and appear in no other numerator.
- Machine opens and scanner clicks are reported in their own rows, not deleted.
- Platform-native metrics carry a label and sit beside a team-defined human-interest row.
- Complaint rate is displayed next to the delivery-error panel, with the Workspace blind spot noted.
- LinkedIn seat transfer dates are marked on the dashboard.
- Two analysts produced identical numbers from the same raw export.
Questions practitioners ask
How do you calculate positive reply rate?
Divide the number of people whose human reply shows interest in a conversation by the number of people with at least one delivered message, grouped by first-contact date. Count referrals, negatives, and auto-replies in their own rows, not in the numerator. In Overloop's published example of 1,000 delivered emails, that gives 1.8%. Folding in the 7 referrals would push it to 2.5%, so say in writing which version your team reports.
What is the difference between reply rate and response rate?
Response rate is usually a platform-native metric where the platform decides what counts. LinkedIn counts an InMail decline or a Not Interested click as a response. Reply rate should mean a metric your team defines, where the numerator is a written reply you have classified. Keep the two words separate in your dictionary so nobody compares a LinkedIn response rate against an email human reply rate as if they measured the same thing.
Are email open rates still accurate?
As a count of fetch events, yes. As a measure of human attention, no. Apple Mail Privacy Protection triggered 18.2% of unique opens in SendGrid's data, and machine-open flags can also mark genuine opens as machine opens. Report raw opens, flagged opens, and unflagged opens as three separate rows, and don't use any of them as a success metric for person-level outreach. Replies and meetings are better outcomes.
How is spam complaint rate calculated by Gmail and Yahoo?
Both divide user spam reports by mail delivered to the inbox, not by everything you sent. Gmail calculates the rate daily and shows it in Postmaster Tools. Google's guidance is to stay below 0.10% and never reach 0.30%. Yahoo says to stay below 0.3%. You can't reproduce these figures yourself because you can't see inbox placement. Track an internal complaint rate on delivered mail as a labelled proxy.
Should auto-replies be removed from the reply rate denominator?
My recommendation is no. Take them out of the human-reply numerator and leave the person in the denominator. Reply.io's default does the opposite and removes auto-replied contacts from its Contacted denominator, which makes the rate higher. If a person's place in the denominator depends on how they replied, the rate stops being comparable between campaigns. Whichever choice you make, write it in the dictionary.
Should reply rate be counted per email or per person?
For outcome metrics, count per person. Three emails to one person who replies once is a 100% prospect reply rate but a 33% email reply rate, so mixing the two makes multi-step sequences look worse than single sends. Keep a separate per-message, per-step row only for diagnosing which step in a sequence draws replies. Label it clearly so it never sits on a dashboard next to the per-person figure.
Try it on the search you came here for
Stop building boolean strings. Just describe the person.
Type one sentence. I plan the search, read GitHub, public LinkedIn and Crunchbase records, and the open web as it is right now, and hand back a ranked list with the reason next to every name.
01Describe them
One plain sentence. Role, city, stack, stage, whatever matters to you.
02I read the web live
GitHub, public LinkedIn and Crunchbase records, the open web. Not a database that went stale last quarter.
03You read the shortlist
Ranked, with the reasoning under every name. Open a profile, ask a follow-up, narrow it down.
- Staff backend engineers in NYC who shipped Rust in production
- Series A fintechs in SF under 50 people, growing headcount this year
- Maintainers of fast-growing Rust web frameworks on GitHub
- No boolean, no filters, no seat to buy. One box.
- Read at search time, so a profile updated yesterday counts today.
- Every step visible as it runs, every name with its reason.
500 free credits on sign-up. No card, no demo call. See real searches.