Before you scale acquisition on a declining curve
Retention Curves for Startups: does your cohort curve flatten? A pre-written demand tripwire
A slide that prints “D30 is 22%” without the cohort, the core action, or what you will do if the curve still slides toward zero is not validation. A retention curve asks a smaller, honester question: of the people who started together, what share still do the product’s core action over time — and did that line flatten at a level you wrote down before you looked? Matching the gate is a decision signal. It is not a forecast that you will succeed.
Last updated: October 21, 2026
Direct answer
What is a retention curve — and why write the tripwire first?
A cohort retention curve plots, for a group of users who started in the same period, the percent still doing the product’s core action at day, week, or month N. Sequoia’s data-science team describes three shapes: flattening (some users found lasting value), declining toward zero (no product-market fit yet), and rare “smiling” curves. Write the tripwire before you look: “by period N the curve must flatten at ≥X% for my category, measured on <core action>, or we change the product, not the funnel.” Hitting it is a decision signal — not a prediction of success.
Key takeaways
What to remember
- A retention curve is user/logo usage over time, not revenue. State the temporal cohort (everyone who signed up in one week or month), the denominator, the core action, and a time frame that matches how often the product is naturally used. Andrew Chen’s line: “Revenue retention expands, while usage retention shrinks.” Dollar retention lives on the NRR page.
- Sequoia’s three archetypes: flattening (some users found lasting value — “The higher the level at which the curve flattens… the healthier the product”), declining toward zero (not yet product-market fit; growth-hacking the top of the funnel without fixing retention is a leaky bucket — their BranchOut example), and rare smiling curves. Casual puzzle games D30 7–10% versus strategy games >30%. Poncho revised sign-up questions and 7-day retention moved 60% → 80%. Total retention is a weighted average of cohorts — a 10× larger cohort dominates. D1 and the D7/D1 ratio are early indicators.
- Lenny Rachitsky with Casey Winters (9 June 2020) asked 20 growth practitioners plus public comps for 6-month user retention. Consumer social ~25% GOOD / ~45% GREAT; consumer transactional ~30% / ~50%; consumer SaaS ~40% / ~70%; SMB/mid-market SaaS ~60% / ~80%; enterprise SaaS ~70% / ~90% in the summary list (the detailed section header says ~75% GOOD — present here as ~70–75%, not a silent pick). Required caveat: lower retention can be OK if you are just starting out, have low CAC and marginal costs, or are not building a venture-scale business — “a flat retention curve that drives a scalable acquisition strategy is enough to keep your business alive.” Also: “startups rarely increase retention significantly.”
- Andrew Chen (8 September 2025): a common decay is “Whatever the D1, it drops by 50% on D7. Whatever the D7, it drops by another 50% at D30.” If early retention is not incredibly good, late retention probably is not either. D1 40% → 50% can be workable; D1 10% → “you probably just have built something that nobody wants,” and many products are better off pivoting. Most products churn 90% or more in the first 30 days. Early users retain best (“The Golden Cohort”); later, paid, and expansion cohorts retain worse. Seeing 50% D30 is rare — “once every couple years.” Viral growth with bad retention fails. Seasonality makes the number hard to read. Chen & Quettra (2015 Android, 125M+ devices): average app loses 77% of DAUs within 3 days of install, 90% within 30 days, over 95% within 90 days. That table is labeled 2015 Android, not a current benchmark.
- Write the tripwire before you measure: cohort + denominator + core action + time frame + checkpoint + threshold + named action. With tiny cohorts, show counts (n retained / n started), not only percents. Do not read a curve from one cohort of 12 users. Separate cohorts by channel. A Yibud score is not retention and not a success prediction.
Why it matters
Why a blended “retention is fine” slide can still hide no demand
Founders treat “people keep coming back” as a feeling. Feelings do not name whether this week’s sign-ups still do the core action at week 8, or whether a huge paid cohort is dragging the blended number down. Three sittings are useful here as instruments, not as personal odds: Sequoia’s shapes, Lenny’s 6-month category bands, and Chen’s decay pattern.
Sequoia: flattening is the demand signal; a declining curve is not yet fit
Sequoia’s data-science team: a flattening curve means some users found lasting value. A curve that declines toward zero means the product has not achieved product-market fit yet — change the product and the value proposition, do not just pour more people in the top. Their BranchOut example is the leaky-bucket warning: growth-hacking acquisition without fixing retention. The height of the flatten matters — higher is healthier — but only against similar products in your genre.
Lenny: category bands, and startups rarely raise retention later
Lenny and Casey Winters (9 June 2020) collected 6-month user-retention bands from 20 growth practitioners plus public comps. Those bands are vocabulary for a tripwire you write first, not a law and not a success certificate. The same post is blunt: “startups rarely increase retention significantly.” If the curve is still sliding toward zero, hoping next quarter’s onboarding A/B test will save it is usually hope editing the bar.
Chen: early retention predicts late retention; the Golden Cohort lies
Chen (8 September 2025): if early retention is not incredibly good, late retention probably is not either. The common pattern halves from D1 to D7 and again from D7 to D30. Early organic users — “The Golden Cohort” — retain best; later, paid, and expansion cohorts retain worse. Scaling paid acquisition on the first happy cohort is how a declining curve gets a marketing budget.
The six terms
Cohort, core action, time frame, curve shape, triangle chart, user vs revenue retention
These six fields are the instrument. If any one is fuzzy, the percent is a slide, not a tripwire. This page does not publish a homemade “good D30 for SaaS” grade table beyond the four sources named below.
Cohort — a temporal group, with the denominator written down
A temporal cohort is everyone who started in the same period — all sign-ups in one week or one month. Lenny’s denominators, used as written: registered users for consumer social; users with at least one transaction for consumer transactional; paid subscribers for consumer SaaS; companies that have started a paid subscription for SMB/mid-market SaaS. Lenny’s post does not specify a denominator for enterprise SaaS — do not quietly reuse the SMB/mid-market company definition there. If you change the denominator after the report, you changed the sitting.
Core action — Sequoia’s “meaningful activity,” not an app open
Sequoia: define what counts as meaningful activity. Uber or Lyft = a completed ride, not an app open. Spotify or Netflix = starting content. A messaging app = writing or reading a message. Sequoia also notes Facebook’s attributed example: connecting to seven friends in 10 days. This page adds no extra numbers to that example. If you count sessions, you are measuring curiosity.
Time frame — match the product’s natural frequency
Sequoia: pick daily, weekly, or monthly to match how the product is actually used. An e-commerce or travel product might expect a visit once a quarter; a social app or game expects daily use. Measuring a weekly product on a monthly clock hides the leak. Infrequent products (travel, tax) need a longer checkpoint — do not force a D7 sitting on a once-a-year job.
Flattening, declining, smiling — Sequoia’s three archetypes
Flattening: some users found lasting value and keep coming back. Declining toward zero: not yet product-market fit. Smiling: rare; churned users return as the product and network effects improve. Sequoia: for most products, including those with smiling curves, retention eventually trends toward zero as competition and behavior shift. Write which shape you will treat as a miss before you plot.
Triangle (cohort) retention chart — horizontal, diagonal, vertical
Sequoia’s usual view: each row is one temporal cohort; column 0 is cohort size; later columns are the percent still active. Horizontal features are cohort-specific (a campaign month). Diagonal features hit every cohort at once (a launch, an outage, news). Vertical features show up in annual subscriptions or trials. Total retention is a weighted average — a 10× larger cohort dominates the blend.
User/logo retention vs revenue retention
This page is the share of people or companies still doing the core action. Chen: “Revenue retention expands, while usage retention shrinks” — the remaining users sometimes spend more. Dollar retention of a starting ARR base is NRR/GRR, on a different Yibud page. Do not paste a 110% NRR onto a declining usage curve and call demand proven.
The three shapes
Flattening, declining toward zero, and the rare smiling curve
Sequoia’s data-science team named three archetypes. Your job is to write, before the first plot, which shape fires the action — flatten and keep going, or still declining toward zero and change the product. Do not invent a fourth “it’s complicated” shape after the chart looks ugly.
Flattening — some users found lasting value
A share of the cohort keeps doing the core action, and the line stops falling. Sequoia: “The higher the level at which the curve flattens… the healthier the product.” Height is relative to similar products in your genre, not to an enterprise SaaS band pasted onto a consumer social app. Flattening is a decision signal. It is not proof the startup will succeed.
Declining toward zero — not yet product-market fit
Sequoia: when the product has not achieved product-market fit, the curve continuously declines toward very few or zero users. The work is to change the product and find a value proposition for a core set of users. Growth-hacking the top of the funnel without fixing retention is their leaky-bucket outcome. BranchOut is the named example on that page. Do not scale paid acquisition on this shape.
Smiling — rare, and still not a success certificate
Sequoia: when a product is truly exceptional, the curve can rise as product development and network effects bring churned users back. They call this rare. Chen (2025) is blunter about what you almost never see: a curve that starts high, goes low, then becomes high again. Do not wait for a smile. Write a flatten-or-change tripwire you can honor this quarter.
Read more than one cohort. Sequoia’s triangle chart is how you see whether last month’s paid burst is a horizontal scar, whether an outage cut every row on the diagonal, or whether a 10× January cohort is the only number in the blend. One row of 12 users is not a curve.
Worked example
Two weekly cohorts you can copy (illustrative — fictional)
The counts below are illustrative. Names, product, and every headcount are invented so you can see a flattening row next to a declining row. This is not a study, not an industry benchmark, and not a Yibud report. Copy the method, not the numbers.
The promise, written before week 1 closes
Illustrative names: founder Priya Shah at Fieldnote, a weekly field-log tool for boutique survey crews. The core action is a completed weekly log, not an app open. Denominator: everyone in the weekly sign-up cohort who completed their first weekly log in that week — same definition on every row. The tempting trophy is a blended “retention is fine” slide that mixes an early organic week with a later paid week. The tripwire is written first, on 3 January: by week 8, each weekly cohort must flatten (the line stops falling) at ≥35% week-8 retention on completed weekly logs among users in that cohort, or we change the product, not the ad set. That 35% floor is a pre-PMF bar Priya justified for this weekly crew workflow before she looked — not Lenny’s 6-month consumer-SaaS band pasted onto week 8. If the sample is still tiny at week 8, we report n retained / n started and do not scale paid.
Week-of-6-Jan organic cohort — counts, not only percents
Started: 40 users in the cohort (each completed their first weekly log that week — the same denominator as the tripwire). Still completing a weekly log: week 1 = 24 / 40; week 4 = 16 / 40; week 6 = 14 / 40; week 8 = 14 / 40. The line moved 60% → 40% → 35% and then held at 14 / 40 through week 8. At week 8 that is a flatten at 14 / 40, which meets the ≥35% bar written before the data. A slide that printed only “35%” without 14 / 40 would have hidden how small the sitting is.
Week-of-3-Mar paid cohort — still declining toward zero
Started: 40 users from a paid burst (each completed their first weekly log that week — same denominator as the tripwire). Week 1 = 18 / 40; week 4 = 8 / 40; week 8 = 3 / 40. The line is still sliding — a miss on shape and on the ≥35% bar. Sequoia would read this shape as not-yet-fit, not as a funnel problem. Chen’s Golden Cohort warning is why the January organic row must not be averaged into March. A 10× paid week would have dominated any blended percent.
What the tripwire does — and what must not happen
January’s flatten at 14 / 40 meets the ≥35% bar they set before looking — one cohort, not proof the startup will succeed. March’s 3 / 40 still declining is a miss: change the product or the value proposition; do not 3× the paid set. Do not blend 14 + 3 and call the company “21%.” Do not switch the core action to app opens after week 8 so the percent looks kinder. Keep the next cohort on the same definitions.
Write the tripwire before the first plot. If a later cohort is still declining toward zero, honor the action. If a cohort flattened at the bar you wrote, continue to the next cheapest test — usually the next cohort on the same definitions, or a kill/pivot contract that now has a curve to read. The method is the dated page, not the 35% slide. Neither a flatten nor a Yibud score predicts success.
Published bands
Lenny’s 6-month user retention, plus 2015 Android from Quettra — labeled, not a grade table
The first table is what Lenny Rachitsky and Casey Winters published on 9 June 2020 for 6-month user retention (20 growth practitioners plus public comps). The second table is Chen & Quettra’s 2015 Android extract. Neither is a Yibud letter-grade chart, a personal-odds table, or a current “good D30 for SaaS” law. Use them as vocabulary, then write your own tripwire.
| Category | GOOD (6-month user retention) | GREAT (6-month user retention) |
|---|---|---|
| Consumer social | ~25% | ~45% |
| Consumer transactional | ~30% | ~50% |
| Consumer SaaS | ~40% | ~70% |
| SMB / mid-market SaaS | ~60% | ~80% |
| Enterprise SaaS | ~70–75% | ~90% |
Source: Lenny Rachitsky, “What is good retention?” (with Casey Winters; Lenny’s Newsletter, 9 June 2020). User retention = percent of users who signed up and are still active six months later. Denominators in that post: registered users (consumer social); users with ≥1 transaction (consumer transactional); paid subscribers (consumer SaaS); companies that have started a paid subscription (SMB/mid-market SaaS only — the post does not specify an enterprise denominator). This page does not reproduce Lenny’s NRR table — that sitting belongs on the SaaS NRR page.
Enterprise note, required: the summary list at the top of Lenny’s post says enterprise SaaS ~70% GOOD / ~90% GREAT. The detailed section header for the same category says ~75% GOOD / ~90% GREAT. This page prints ~70–75% so the two lines are not silently collapsed.
Required caveat from the same post: lower retention can be OK if (1) you are just starting out, (2) you have low CAC and marginal costs, or (3) you are not building a venture-scale business. Lenny: “a flat retention curve that drives a scalable acquisition strategy is enough to keep your business alive.” Also quote Lenny: “startups rarely increase retention significantly.” Do not treat a miss against the GOOD band as automatic death, and do not treat a GREAT print as proof you will succeed.
Quettra 2015 Android — labeled historical, not today’s benchmark
Andrew Chen and Ankit Jain (Quettra) published install-to-active rates from more than 125 million Android devices. Average app: loses 77% of DAUs within 3 days of install, 90% within 30 days, over 95% within 90 days. Top apps show a higher D1 but a similar D1→D30 falloff. Chen: bending the curve happens in the first days of activation, not by notification spam. Label this 2015 Android data. It is not a 2026 SaaS law.
| Segment | D1 | D7 | D30 | D90 |
|---|---|---|---|---|
| Average app | 29.17% | 17.28% | 9.55% | 3.97% |
| Top 10 apps | 74.67% | 67.39% | 59.80% | 50.87% |
Source: Andrew Chen & Ankit Jain (Quettra), “New data shows losing 80% of mobile users is normal, and why the best apps do better.” Methodology on that page: Android, five months from 1 January 2015, apps with >10,000 installs, Google and pre-installed apps excluded. Only the average and top-10 rows are reproduced here.
Chen 2025 — the D1 → D7 → D30 halving pattern
Chen (8 September 2025): “Whatever the D1, it drops by 50% on D7. Whatever the D7, it drops by another 50% at D30.” “If your early retention isn’t incredibly good… your late retention also probably isn’t any good.” D1 40% → 50% is workable; D1 10% → “you probably just have built something that nobody wants,” and many products are better off pivoting. Most products churn 90% or more in the first 30 days. Seeing 50% D30 is rare — “once every couple years.”
Sequoia — genre matters more than a universal D30
Sequoia: benchmark against similar products in your genre. Casual puzzle games D30 7–10%; strategy games >30%. The same D30 can be ahead of one genre and behind another. Do not compare your consumer-social flatten to Lenny’s enterprise band.
Sequoia — Poncho, and the 10× cohort that owns the blend
Poncho, a weather chatbot, revised the number of sign-up questions; 7-day retention moved from 60% to 80%. Separately: total retention is a weighted average of cohorts. If January is 10× December, January is the number. D1 and the D7/D1 ratio are Sequoia’s early indicators — if D1 falls and D7/D1 holds, work the first day; if D1 holds and D7/D1 falls, work the first week.
Chen — Golden Cohort, seasonality, viral plus bad retention
Early organic users retain best (“The Golden Cohort”). Later, paid, and expansion cohorts retain worse. Seasonality (commerce, travel, wellness, dating — and even business-software calendars) makes retention hard to measure. Viral growth with bad retention fails; Chen says that experiment has already been run many times. None of that is a Yibud grade.
Scope notes, still the sources: Sequoia “Retention” (data-science essay). Lenny, 9 June 2020 (6-month user retention). Chen, 8 September 2025 (decay pattern). Chen & Quettra, 2015 Android. Do not invent a “good D30 for SaaS” table. Do not paste Lenny’s NRR bands onto this page. Do not treat a flatten as proof you will succeed.
How to run it
How to run a retention-curve tripwire without faking scale
This is Yibud’s suggested sequence, not a Sequoia checklist and not a Lenny audit. Honor the tripwire you wrote before the first plot. If you cannot name the decision the curve would change, you are building a slide.
- 1
Write the tripwire before you look
Name the temporal cohort, the denominator, the core action, the time frame, the checkpoint (week 8 / month 6), the threshold (flatten at ≥ the category GOOD band, or a smaller pre-PMF floor you justify in writing), and the action (flatten → keep going / elevate; still declining toward zero → change the product or value proposition, do not scale acquisition). Date the page. Do not rewrite the bar after the chart.
- 2
Lock cohort, denominator, core action, and clock
Same definitions on every row. Sequoia: do not measure a weekly product monthly. Lenny: do not switch from registered users to “people who opened the email.” If you only have four weeks, report four weeks and say so — do not multiply a D7 by four and call it a month.
- 3
Plot counts, then percents. Read more than one cohort.
With tiny n, write 14 / 40 next to 35%. Do not read a curve from one cohort of 12 users. Sequoia’s triangle is the honest view: several rows, sizes in column 0, so a 10× burst cannot hide inside a blend. Optional early indicators from Sequoia: D1, and whether D7/D1 is holding.
- 4
Separate cohorts by channel. Do not average the Golden Cohort into paid.
Chen: early organic users retain best; later and paid cohorts retain worse. Sequoia: one huge cohort dominates the blended number. Put each channel on its own row. A pretty blend that hides a declining paid row is vanity with extra steps.
- 5
Honor the action. If you have no users yet, use a different instrument.
If the checkpoint arrives and the curve is still declining toward zero, do the thing you named — change the product, not the funnel. Pre-launch, there is no honest cohort: use a smoke test, a willingness-to-pay sitting, or interviews first. A Yibud score is not a retention curve.
What you leave with
The retention-curve tripwire (copy this)
If the sitting produced a percent and no dated action, you did not finish. Fill the brackets. Leave nothing as “retention is fine,” “the Golden Cohort looks good,” or “we’ll know it when D30 arrives.”
The fill-in contract
Product: [name]. Cohort: [all sign-ups in week/month starting ____], written before the first plot. Hypothesis this curve can falsify: [these users found lasting value / we can scale this channel without a leaky bucket]. Denominator: [registered users / ≥1 transaction / paid subscribers / paying companies — pick Lenny’s matching category]. Core action: [completed ride / started content / completed weekly log — not app open]. Time frame: [daily / weekly / monthly — matches natural frequency]. Checkpoint: [week 8 / month 6]. Signal: [does the curve flatten, and at what % and n retained / n started]. Threshold: [e.g. flatten at ≥ Lenny’s category GOOD band for 6-month user retention, or a smaller pre-PMF floor we justify here — not a homemade D30 SaaS grade]. Action if still declining toward zero: [change the product / value proposition; do not scale acquisition]. Action if it flattens at the bar: [keep going / elevate long-term retention]. Channel split: [organic vs paid on separate rows]. A Yibud score does not fire this line.
The “≥ category GOOD band” line is Yibud’s suggested template from Lenny’s 6-month user-retention sitting — not a census, not permission to skip counts, and not permission to treat a flatten as success. Rewrite the threshold before you measure if you have a reason (Lenny’s own “lower retention can be OK” cases). Do not rewrite it after.
Example tripwire (Yibud’s suggested template, not a law)
If, for each weekly sign-up cohort, the share still completing a weekly log has not flattened by week 8 — still declining toward zero — we do not 3× paid acquisition; we change the product or the value proposition. If the curve flattens, we compare the flatten level to Lenny’s consumer-SaaS 6-month GOOD band (~40% user retention among paid subscribers) and write whether we are using that band or a smaller pre-PMF floor. Report n retained / n started. Do not blend organic and paid. Do not switch the core action to app opens after the close.
A flatten on one cohort of 40 is still one row. A Yibud report score does not fire this tripwire. If you have no users yet, this template is the wrong first instrument — use a smoke test, willingness to pay, or interviews instead.
The fork
When to use a retention-curve tripwire — and when not to
A retention curve is the wrong first instrument when there are no users yet, or when the natural job happens once a year and you are forcing a D7 sitting. It is the right instrument when people can start in the same week, a core action can be named, and you are deciding whether to change the product or scale the funnel.
Use it when you have a start date, a core action, and a scale-or-change decision
You can name who started together, what “still here” means, and a clock that matches use. You are about to 3× a channel, tell yourself “retention is 40% so we are fine,” or average the Golden Cohort into a paid burst. That is when a dated flatten-or-not earns its keep. Optional: run NRR, Quick Ratio, Ellis, or NPS as separate sittings — do not average them into a homemade “fit score.”
Do not start here when you have no users, or when the job is very infrequent
Pre-launch? There is no honest cohort. Do not invent a forecast curve. Use a smoke test, a willingness-to-pay sitting, or interviews first. Travel, tax, or other once-in-a-long-while products: Sequoia says pick a time frame that matches natural frequency — a longer checkpoint, not a fake D7. Still asking whether current users would be very disappointed? Ellis PMF survey. Still writing Stop / Pivot / Continue? Those are the decision contracts. This page supplies one usage-curve instrument those contracts can later read. A Yibud score is not a retention curve.
Not the same page
Retention curves vs NRR, Quick Ratio, Ellis PMF survey, NPS, LTV:CAC, vanity metrics, Yibud scores, and decision contracts
Several Yibud pages sit next to this one. Mixing them up turns dollar retention, a same-period inflow÷outflow, a disappointment survey, a recommend score, a lifetime ratio, or a waitlist into “the curve flattened, so scale.”
NRR — dollar retention of a starting ARR base, not this usage curve
That page asks whether customers live at period start paid you more or less after churn, downsell, and expansion (NRR% / GRR%). Chen: revenue retention can expand while usage retention shrinks. A 110% NRR does not tell you the sign-up cohort still does the core action. A flatten at 35% does not tell you the starting ARR base retained 110%.
SaaS NRR →Quick Ratio — same-period inflow ÷ outflow, not a cohort curve
That page asks whether this period’s new + expansion outran churned + contraction (a multiple). This page is one cohort’s usage over time. A 4× Quick Ratio does not tell you week-8 flattened. A flatten does not tell you this month’s bucket is not leaking in dollars.
SaaS Quick Ratio →Sean Ellis PMF survey — stated disappointment, not observed behavior
That page’s signal is the percent of recent real users who would be very disappointed if the product disappeared. A flatten is what people did. Do not treat “very disappointed” as a retention curve, and do not treat a flatten as Ellis’s bar.
Product-market fit survey →NPS — recommend intent, not core-action retention
That page is % promoters minus % detractors on a 0–10 recommend question. Stated intent among people who already used you is not the share still doing the core action at week 8. A +42 NPS does not flatten a declining curve.
NPS for early-stage startups →LTV:CAC — lifetime value ÷ CAC, which this curve feeds
That page asks whether a cohort’s contribution cleared a multiple of CAC. Retention is an input to LTV, not the ratio itself. A 3.2 LTV:CAC does not tell you the usage curve flattened. A flatten does not tell you paid CAC pays back.
LTV:CAC for startups →Vanity metrics — hygiene class, not this cohort tripwire
That page is the sorting hat: hits, cumulative sign-ups, and waitlist size feel like progress and do not tell you what to do next. A blended retention percent with no cohort, no core action, and no tripwire is vanity in a nicer font. Cumulative registered users are not a flatten.
Vanity metrics for startups →Yibud scores — not a retention curve, and not a success prediction
The validation-score and score-calculator pages explain Startup MRI numbers from a deterministic rule engine on a one-line idea plus five questions. Those scores point at the weakest assumption to test next. They are not a cohort flatten, not NRR, not Ellis’s percent, and not a forecast. Do not read a mid-band score as “D30 is fine,” and do not read a flatten as a Yibud GO.
Kill and pivot — contracts that can read this curve
Those pages write Stop, or rewrite one dimension, before hope edits the bar. This page can supply the usage instrument those contracts later read: a dated flatten-or-not against a bar you wrote first. A 35% slide with app opens in the numerator cannot fire a kill line honestly.
Where Yibud fits
Neither a flatten nor a Yibud score predicts success
Yibud is a free, no-signup startup idea validator. Scores come from a deterministic rule engine, not from a language model guessing success. Optional AI text, when it is used, only polishes prose. It does not invent the number. A Startup MRI report can name a weak demand or distribution assumption, then a Day-7 Continue / Refine / Re-test / Stop signal. The score is a flashlight on the weakest assumption to test next — not a retention curve, not NRR, not Ellis, not a forecast. Retention-curve work sits next to that assumption: write the tripwire, do not hope a 35% slide means demand is real. If you already have a report, start with the named assumption, not the overall score. Then run the cohorts yourself. The method on this page stands alone if you never open the analyzer.
Analyze my idea →Common mistakes
What usually wastes a retention-curve week
Counting app opens instead of the core action
Sequoia: Uber/Lyft = completed ride, not app open; Spotify/Netflix = starting content. Sessions measure curiosity. Put the named action on the dated page before the first plot.
Blended retention hiding a decaying cohort
Sequoia: total retention is a weighted average; a 10× cohort owns the blend. Chen: the Golden Cohort retains best; later and paid cohorts retain worse. If you only publish one company percent, you hid the row that is still sliding toward zero.
Measuring a weekly product on a monthly clock
Sequoia: match the time frame to natural frequency. A weekly tool reported monthly can look “retained” while the week-over-week line is already dying. Travel and tax need a longer clock, not a fake D7.
Scaling paid acquisition on a declining curve
Sequoia’s leaky-bucket warning (BranchOut): growth-hacking the top without fixing retention. Chen: viral growth with bad retention fails; early retention that is not incredibly good usually means late retention is not either. Write flatten-or-change before you raise the ad budget.
Comparing your consumer-social curve to enterprise SaaS bands
Lenny’s 6-month GOOD band is ~25% for consumer social and ~70–75% for enterprise SaaS. Sequoia’s games split is D30 7–10% versus >30%. Genre first. A 28% flatten can be GOOD in one category and a miss in another.
Treating a flat curve as proof you will succeed
Sequoia’s flatten is a healthier-product signal, not a guarantee. Lenny: a flat curve that supports a scalable acquisition strategy can keep a business alive — and startups rarely raise retention later. Chen: 50% D30 is rare. Clearing the bar you wrote means that cohort cleared a gate. It is not Ellis’s question, not NRR, and not a Yibud GO. Neither a flatten nor a Yibud report score predicts success.
Sources
Where these ideas come from
- Sequoia Capital Data Science, “Retention” — Used for the three archetypes (flattening, declining toward zero, smiling); declining = not yet PMF; leaky-bucket / BranchOut; “The higher the level at which the curve flattens… the healthier the product”; core-action examples (completed ride, starting content); time frame matching natural frequency; triangle chart (horizontal / diagonal / vertical); weighted average and the 10× cohort; D1 and D7/D1 as early indicators; casual puzzle D30 7–10% vs strategy >30%; Poncho sign-up questions 7-day retention 60% → 80%; Facebook “seven friends in 10 days” as Sequoia’s attributed example only. No extra BranchOut or Angry Birds headcounts.
- Lenny Rachitsky, “What is good retention?” (with Casey Winters; Lenny’s Newsletter, 9 June 2020) — Used for 6-month user-retention GOOD / GREAT bands: consumer social ~25% / ~45%; consumer transactional ~30% / ~50%; consumer SaaS ~40% / ~70%; SMB/mid-market SaaS ~60% / ~80%; enterprise SaaS summary ~70% / ~90% and detailed header ~75% GOOD (printed here as ~70–75%); denominators; the three “lower retention can be OK” cases and “a flat retention curve that drives a scalable acquisition strategy is enough to keep your business alive”; “startups rarely increase retention significantly.” This page does not reproduce the NRR table.
- Andrew Chen, “Why retention is so hard for new tech products” (8 September 2025) — Used for “Whatever the D1, it drops by 50% on D7. Whatever the D7, it drops by another 50% at D30”; early retention predicts late; D1 40% → 50% workable vs D1 10% “nobody wants” / pivot; “most products churn 90% or more in the first 30 days”; “The Golden Cohort”; later/paid/expansion worse; seasonality; viral + bad retention fails; 50% D30 rare (“once every couple years”); “Revenue retention expands, while usage retention shrinks.”
- Andrew Chen & Ankit Jain (Quettra), “New data shows losing 80% of mobile users is normal, and why the best apps do better” — Used for 2015 Android data from 125M+ devices: average app loses 77% of DAUs within 3 days, 90% within 30 days, over 95% within 90 days; table values Average D1 29.17 / D7 17.28 / D30 9.55 / D90 3.97 and Top 10 D1 74.67 / D7 67.39 / D30 59.80 / D90 50.87; higher D1 with similar D1→D30 falloff; bend the curve in the first days, not with notification spam. Labeled 2015 Android, not a current benchmark.
In one paragraph
Summary you can quote
A cohort retention curve plots, for users who started together, the share still doing the core action at day/week/month N. Sequoia: flattening (some users found lasting value; higher flatten, healthier product), declining toward zero (not yet PMF; leaky bucket if you growth-hack the funnel), rare smiling curves; genre benchmarks (casual puzzle D30 7–10% vs strategy >30%); Poncho 60% → 80% on 7-day after revising sign-up questions. Lenny (9 June 2020), 6-month user retention: social ~25/45, transactional ~30/50, consumer SaaS ~40/70, SMB/mid-market ~60/80, enterprise ~70–75/90 (summary ~70 vs detailed ~75); lower retention can be OK; startups rarely raise it later. Chen (2025): D1 halves by D7 and again by D30; D1 10% often means nobody wants it; Golden Cohort; 50% D30 is rare. Quettra 2015 Android: 77% / 90% / 95% DAU loss by day 3 / 30 / 90. Write the flatten-at-≥X% tripwire before you look. Matching the gate is a decision signal. Neither a flatten nor a Yibud score predicts success.
FAQ
Questions founders actually ask
What is a retention curve for a startup?
A plot, for one temporal cohort (everyone who started in the same week or month), of the percent still doing a named core action at day, week, or month N. Sequoia names three shapes: flattening, declining toward zero, and rare smiling curves. Write the flatten-or-change tripwire before you look. It is user/logo usage, not NRR, not Ellis’s survey, and not a success score.
What is good user retention?
This page does not publish a universal grade table beyond the sources. Lenny (9 June 2020), 6-month user retention: consumer social ~25% GOOD / ~45% GREAT; transactional ~30% / ~50%; consumer SaaS ~40% / ~70%; SMB/mid-market SaaS ~60% / ~80%; enterprise SaaS ~70–75% GOOD / ~90% GREAT (summary list ~70%, detailed header ~75%). Sequoia: casual puzzle D30 7–10% vs strategy >30%. Chen 2025: a common pattern halves D1→D7→D30; 50% D30 is rare. Lower retention can be OK if you are early, cheap to acquire, or not venture-scale. Write your own tripwire.
How is a retention curve different from NRR, Quick Ratio, the Ellis PMF survey, and NPS?
NRR/GRR is dollar retention of a starting ARR base. Quick Ratio is same-period (new+expansion)÷(churn+contraction). Ellis is the percent of recent real users who would be very disappointed. NPS is recommend intent. This page is whether a sign-up cohort still does the core action over time. They answer different decisions. Do not average them.
Can I read a retention curve from 12 users?
Not as a percent-only slide. Show counts (n retained / n started). Sequoia’s triangle needs more than one cohort; a 10× row will dominate any blend. Do not scale paid acquisition on one row of 12. If you have no users yet, use a smoke test, willingness to pay, or interviews first.
Is a Yibud score the same as retention? Does a flatten predict success?
No and no. Yibud scores come from a deterministic rule engine on your idea inputs. They point at the weakest assumption to test next. They are not a cohort flatten, not NRR, not Ellis, not NPS, and not a forecast. Neither a flatten nor a Yibud report score predicts success. You do not need an account to use this page.
Write the flatten-or-change tripwire before you scale the funnel
Yibud scores weak assumptions with a deterministic engine. You write cohort, core action, checkpoint, threshold, and action while you are still honest — then a week of n retained / n started is a tripwire, not a 35% slide.
Generate a Yibud report →