Retention Cohort Spreadsheet

Make Every Percentage Reversible

A cohort chart can show that retention fell. A solo founder needs to be able to open the number and find the accounts that caused it.

Use this workbook after the activation funnel has established which accounts reached first value. It follows those accounts to the next natural occurrence of the customer’s job and asks whether the product delivered the value again. It keeps three signals separate: repeated value, value recovered through founder assistance, and an eligible cycle that missed the promised job.

The spreadsheet is working when a surprising cell leads back to an account, an event or observation, and a plausible product decision. It is not working when a smooth curve can no longer explain who returned or why.

Write the Retention Claim First

Fix the meaning of a row before adding formulas. At the top of the workbook, write:

RETENTION CLAIM
Target segment:
Retention unit — person, account, workspace, project, or another value receiver:
Cohort entry event — the first honest value behavior:
Repeated value event:
Natural cycle and next eligible opportunity:
Cohort start period:
Assistance that must remain visible:
Product and event-definition versions:
Current decision this workbook must inform:

The cohort entry event should be activation, not signup. Someone who never received first value has an activation problem; mixing that account into a retention denominator makes both problems harder to see.

Choose the retention unit that receives the outcome. A reporting product may retain an agency even when a different employee sends the next report. An API may retain a production service even when no person returns to its dashboard.

Use the customer’s clock. Weekly client reporting, monthly close, and the next incident are different schedules. Define whether a cycle is an exact date, a bounded window, or the next observed occurrence of the job. D7 is meaningful only when day seven corresponds to another opportunity for value.

Keep One Account Ledger

The account ledger holds stable facts that should not be copied into every cycle. Use one row per retention unit:

account_id | segment | acquisition_source | cohort_started_at
cohort_entry_event | cohort_entry_at | activation_assistance
product_version | event_definition_version | current_status

Keep test accounts out. Preserve the original segment and definition when they change rather than rewriting history. If the product serves several distinct jobs, split the cohorts; a shared login does not make their natural frequencies comparable.

The ledger should also name the evidence source. A durable product event, database record, completed customer artifact, support note, or manual observation may be sufficient. Manual evidence is not inferior at this stage, but unmarked inference is.

Record One Row Per Account-Cycle

A long cycle log is easier to audit than a wide sheet with a growing number of week columns. Add one row for each expected value cycle, then update it when the account reaches that opportunity:

account_id | cycle_number | cycle_due_at | eligible_state
value_event_at | result_state | assistance_minutes | assistance_action
failure_or_hesitation | product_version | evidence_source | note

Use a small, controlled vocabulary. For eligible_state, record not_due, eligible, no_opportunity, or unknown. When the account is eligible, record one of these result states:

• repeated — the value event occurred without founder help;
• repeated_assisted — the value event occurred, but founder labor carried part of the promise;
• missed — the opportunity arrived and the value event did not occur;
• resurrected — value repeated after at least one missed eligible cycle;
• unknown — the evidence cannot distinguish absence from an instrumentation or observation failure.

An agency with no client report due this week is not automatically retained, and it is not automatically churned. Mark why the opportunity did not occur. not_due and no_opportunity remain outside that cycle’s denominator, while unknown blocks a trustworthy rate until the evidence is repaired. Repeated absence of the expected job may reveal weak qualification or a mistaken natural frequency; do not quietly discard it to improve the rate.

Keep prompts and founder rescue visible. A reminder, import repair, manual output review, custom configuration, or customer-success call may prove that the problem matters. It does not prove that the product carries the workflow on its own.

Derive the Cohort View

Build the summary from the cycle log rather than typing percentages by hand. For each start cohort and cycle, show the activated cohort size, accounts whose next opportunity has arrived, unassisted repeats, assisted repeats, missed accounts, resurrected accounts, unknown records, accounts without an opportunity, and accounts not yet due.

Two rates are useful when their boundaries stay visible:

unassisted cycle retention
  = repeated accounts
    / activated cohort accounts with an eligible cycle

observed value repetition
  = repeated + repeated_assisted + resurrected accounts
    / activated cohort accounts with an eligible cycle

The second rate can show that customers still want the outcome. It must not be presented as self-serve retention. Keep raw counts beside both rates when a single account can move the percentage sharply. If an unknown result could change the decision, repair the record instead of reporting either rate.

A compact matrix earns its place here because the reader needs to compare the same cycles across start cohorts:

COHORT START   ACTIVATED   C1                 C2                 C3
2026-06-01     8           5R / 1A / 2M      4R / 1A / 3M      4R / 0A / 4M
2026-06-08     6           4R / 1A / 1M      3R / 1A / 1M / 1N —

R = repeated without help   A = repeated with assistance
M = missed                  N = not due

Do not let the matrix become the evidence store. A cell such as 4R / 1A / 3M must filter directly to the eight account-cycle rows behind it.

Read a Flat Rate More Carefully

In a modeled review, carry five activated small agencies forward from a weekly client-digest funnel. In the first reporting cycle, four send another client-ready digest. Three do so independently. Beacon Studio needs the founder to repair an advertising export; Cedar & Co misses the cycle because nobody owns client approval.

In the second cycle, four of five again send a digest. The aggregate rate is unchanged, but the accounts have moved. Beacon succeeds without help after the supported import path is clarified. Cedar returns after inviting an approval owner. Dovetail misses because its only recurring client contract has ended. Elm sends its digest only after the founder repairs a newly changed export. Aster sends both cycles without help.

“Eighty percent retained” conceals all of this. Only one account has repeated both cycles without assistance or interruption. One apparent recovery came from a workflow-owner change. One loss challenges the claim that the segment has recurring work. Import fragility remains, even though it moved to another account.

The next step is not a general retention campaign. The founder can qualify for an active weekly reporting obligation, support the promised export formats, and ask whether the next small cohort completes two consecutive client-digest cycles without repair. Reminders would blur the approval problem; broader acquisition would feed more accounts into a path that is not yet repeatable.

Diagnose Before Choosing Work

Read the sheet in sequence.

First, inspect accounts excluded from the retention cohort because they never activated. A large group there points back to qualification, setup, trust, or first value. Do not call it retention work.

Next, inspect eligible misses. Compare their segment, recurring job, entry path, product version, and first failure. A coherent segment that returns while another does not may justify narrowing. A drop after one impressive result may mean novelty rather than recurring utility. A missed cycle caused by a broken event pipeline requires evidence repair before product interpretation.

Then inspect assistance and resurrection. If customers return only after a prompt or rescue, the value may be real while the product remains incomplete. A resurrection at the next true deadline may mean the original clock was wrong. A resurrection immediately after outreach may measure the founder’s relationship rather than the product.

Finally, find the account that contradicts the preferred explanation. If one qualified account succeeds through the supposedly broken path, learn what is different before authorizing a large fix.

Complete the workbook with a short decision record:

RETENTION READ — [REVIEW DATE]
Cohorts mature enough to interpret:
Largest trustworthy loss:
Strongest retained pocket:
Continuous unassisted accounts:
Founder-assisted accounts and minutes:
Resurrections and likely cause:
Evidence failure or definition change:
Leading explanation:
Account that challenges it:
One intervention or investigation:
Accounts exposed next:
Expected movement by the next eligible cycle:
Result that would reject the explanation:

Protect the History

Do not overwrite old cohorts when the product, event definition, segment, or clock changes. Date the new definition and begin a comparable series. Do not fill immature cells with zero, combine assisted and independent returns, or smooth tiny cohorts until their account identities disappear.

The workbook is ready for a decision when every summary can be reconstructed, accounts not yet due remain out of the denominator, and one explanation is specific enough to challenge in the next natural cycle. Its purpose is not to produce a pleasing curve. It is to discover whether the product carries the same promise twice.