Before a percent counts as a result
How Many Visitors Does a Smoke Test Need — Read the Rate as an Interval
A single percent is not a verdict. “How many visitors does a smoke test need?” is a precision question: a small sample gives a wide range of plausible true rates. Before you launch, write who counts as a qualified visitor, the conversion event, the pass line, the fail line, a fixed N, and the named action for pass, miss, or inconclusive. Then read k signups out of n as an interval — not as “we’re at 5%.”
Last updated: October 6, 2026
Direct answer
How many visitors does a smoke test need?
Enough to read the signup rate as an interval against a line you wrote in advance — not enough to “validate” the idea. Read k signups out of n qualified visitors as a 95% Wilson interval, not just k/n. The same observed 5% is 2.2%–11.2% at 100 visitors, 2.7%–9.0% at 200, and 3.3%–7.6% at 400. Zero conversions are information: 0 of n means the true rate is at most about 3/n at 95% confidence (0/100 → about 3%; 0/300 → about 1%). Fix N before launch; do not stop the moment the graph looks good.
Key takeaways
What to remember
- A smoke test is not an A/B test. You are checking one rate against a pass line and a fail line you wrote before anyone arrived.
- Read k signups out of n qualified visitors as a 95% Wilson interval, not as a single percent.
- Write a three-way rule before launch — clear miss, clear pass, or inconclusive — and name the action for each. That rule is Yibud’s suggested template, not a statistical law.
- Zero events are information (rule of three). Unqualified clicks shrink the interval around the wrong number. Traffic quality beats a bigger N.
- A passing signup rate is interest evidence, not money. After a pass, the next cheapest test is usually willingness to pay.
Why it matters
A lone percent is how founders keep the test open instead of deciding
A common pattern is treating “we’re at 5%” as a verdict after mixed clicks, then sliding the fail line when a later day looks sad. The number was never going to tell them the idea would succeed. It could only tell them how wide a range of true rates still fit what they saw. Brown, Cai and DasGupta (2001) showed that the textbook Wald interval — p ± 1.96 · sqrt(p(1−p)/n) — has erratic coverage, worse than textbooks suggest. They recommend the Wilson interval or the Jeffreys interval for small n (40 or less), and the Agresti–Coull interval for larger n. Wilson (1927) is where the score interval comes from. Decide the numbers before you look. A rate invented after a warm afternoon is a recap.
Vocabulary
What the seven fields mean before anyone visits
The page is not “wait until it feels like enough traffic.” It is a dated contract: who counts, what counts as a conversion, where pass and fail sit, how wide an interval you will accept, the fixed N, and what you will not do while the test is running.
1. Qualified visitor — the person the rate is about
A qualified visitor matches the role and the source you named before launch. A random click from the wrong job title is discarded, not averaged in. Unqualified traffic can make a tight interval around a rate that does not belong to your customer.
2. Conversion event — the one action you will count
Waitlist, reply, or a deposit you can actually refund. Clicks on the logo do not count. Two buttons split the signal. Write the event in one sentence a stranger could score.
3. Pass line — the rate that would let you continue
A number you can live with if it holds, written before the first visitor. “Interesting” is not a line. This page will not invent a typical startup conversion rate for you.
4. Fail line — the rate that would make you stop or rewrite
A second number, also written before launch. The fail line is usually below the pass line. If you only have one line, every messy week becomes a story.
5. Confidence interval — the range of rates still plausible
Here, a 95% Wilson score interval for k events out of n trials (z = 1.96). It is a range of plausible true rates, not a promise about the next hundred visitors. It is not a Yibud engine score.
6. Fixed N — how many qualified visitors you will wait for
Write N before launch. Optional: one pre-declared checkpoint and a cap (extend once, then stop). Do not pick N after you like the graph. Peeking-and-stopping is a different experiment than the one you wrote.
7. Peeking — looking, then stopping because it looks good
Evan Miller (2010), writing about A/B tests, showed that significance assumes the sample size is fixed in advance. Stopping as soon as a result looks significant inflates false positives. His worst-case example — checking after every observation up to 150, at a 50% base rate — gave 26.1% false positives instead of 5%. Peeking 10 times means a reported 1% significance is actually about 5%. The lesson transfers to a one-rate smoke test: fix N, then read.
The evidence
What the papers and the essay actually measured — and what they did not
These are published findings, labeled as such. They are not a personal forecast, not a typical signup rate, and not a Yibud rule. Do not import conversion benchmarks this page does not publish.
Wilson (1927) — the score interval this page uses
Edwin B. Wilson, “Probable Inference, the Law of Succession, and Statistical Inference,” Journal of the American Statistical Association 22(158):209–212 (1927). Origin of the Wilson score interval. The table below uses that interval at 95% (z = 1.96). The numbers are computed examples, not field results from Yibud.
Brown, Cai & DasGupta (2001) — why not the Wald interval at small n
“Interval Estimation for a Binomial Proportion,” Statistical Science 16(2):101–133. The standard Wald interval (p ± 1.96 · sqrt(p(1−p)/n)) has erratic coverage, worse than textbooks suggest. The authors recommend the Wilson interval or the Jeffreys interval for small n (40 or less), and the Agresti–Coull interval for larger n. That is why this page reads smoke-test rates with Wilson, not the plus-or-minus formula most founders remember from a class.
Hanley & Lippman-Hand (1983) — rule of three when nothing happens
“If Nothing Goes Wrong, Is Everything All Right? Interpreting Zero Numerators,” JAMA 249(13):1743–1745. If none of n shows the event, the 95% upper limit is about 3/n. The approximation agrees with the exact calculation to the nearest percentage point when n > 30. Their setting is medical (zero adverse events). The same math applies to zero signups: 0/100 → about 3%; 0/300 → about 1%. Exact (binomial) zero-event upper bounds on this page, 1 − 0.05^(1/n): n = 100 → 2.95%; n = 300 → 0.99%. Those are not Wilson bounds — the Wilson upper bound for 0/100 is 3.7% (see table). Zero is information, not “no data.”
Evan Miller (2010) — peeking inflates false positives (A/B context)
“How Not To Run an A/B Test” (2010-04-18). Significance assumes sample size is fixed in advance. Stopping as soon as the result is significant inflates false positives. Worst-case example in that essay: checking after every observation up to 150, 50% base rate, 26.1% false positives instead of 5%. Peeking 10 times means a reported 1% significance is actually about 5%. His advice: decide sample size in advance. Label the source as A/B-testing context. The transfer to a smoke test is the same habit: do not stop the moment the signup graph looks kind.
A three-way rule you may adopt — written before launch, not a law
Yibud’s suggested template, not a statistical theorem: clear miss if the whole interval sits below your pass line (even the optimistic reading misses). Clear pass if the observed rate is at or above your pass line and the interval’s lower end is above your fail line. Anything else is inconclusive — do the one action you named before launch (extend once to the cap, or treat as a miss). Do not invent a kinder band after you see 6/200.
Not the same page
Sample size vs how-to-run, interview count, money, PMF surveys, picker pages, contracts, and scores
Several Yibud pages sit next to this one. Mixing them up turns a landing-page build, a saturation count, or a mid-band engine score into “we had enough visitors.”
Smoke test, landing page, fake door — how to run, not how to size
Those pages are the experiment: the page, the one CTA, the channel, the honest waitlist. This page is how many qualified visitors to wait for and how to read k/n as an interval. A live page is not a sample-size contract. The 200–500 visitor rule of thumb on the landing-page guide is not being overturned here — this page shows what 200 can and cannot tell you.
How many customer interviews — coverage, not a conversion interval
That page’s signal is whether a batch in one segment still adds new need codes. This page’s signal is the precision of one visitor conversion rate. Saturation is not a Wilson interval. An interval is not interview coverage.
How many customer interviews is enough →Willingness to pay — money moving, not a signup rate
A passing smoke-test rate is interest. It is not a deposit, a checkout, or a priced concierge. After a pass, move to the willingness-to-pay page. Do not treat an email as a price.
Willingness-to-pay test →Product-market fit survey — usage among current users, not visitor conversion
That page’s signal is the percent of recent real users who would be very disappointed if the product disappeared. This page’s signal is k signups out of n qualified visitors. A Sean Ellis percent is not a landing-page conversion interval.
Product-market fit survey →Cheapest validation test — which test, not how large
That page picks the cheapest instrument for the unknown you actually have. This page sizes a smoke test once that is the instrument. A method picker is not an N.
Cheapest validation test →Kill, pivot, assumption, decision — contracts that can later read this rule
Those pages write Stop, rewrite a dimension, name one claim, or read a week of evidence you already have. This page supplies one input those contracts can later read: a dated pass / fail / inconclusive reading of one rate. It is not a Stop / Pivot / Continue sitting.
Yibud scores — not a conversion rate or a confidence interval
The validation-score and score-calculator pages explain Startup MRI numbers from a deterministic engine on a one-line idea plus five questions. Those scores are flashlights on assumptions. They are not a signup rate and not a Wilson interval. Do not read a mid-band score as “enough visitors,” and do not read 10/200 as a Yibud GO.
The sitting
How to write and honor the sample-size contract in one sitting
Do this before the page is public. Set a timer. Leave the ads off until the fields are filled. If you have a co-founder, each of you writes a draft alone, then you keep the tighter visitor definition.
- 1
Write the contract before anyone can click
Qualified visitor, conversion event, pass line, fail line, fixed N, optional single checkpoint and cap, and the named action for pass, miss, and inconclusive. If any field is blank, you are not ready to buy traffic.
- 2
Define the qualified visitor so a stranger could score the log
Role plus source. “US independent bookkeepers who arrived from one named community post” is a visitor definition. “Traffic” is a conference. Discard the rest. A tight interval on mixed junk is still junk.
- 3
Write pass, fail, and the three-way actions on the same page
Yibud’s template, not a law: miss if the whole interval is below the pass line; pass if the observed rate is at or above the pass line and the lower end is above the fail line; otherwise inconclusive → the one action you already named. Do not add a fourth outcome called “promising.”
- 4
Fix N — and decide the optional checkpoint before launch
Pick N for the precision you need, not for the number that would make a blog post. The table below shows what 100, 200, and 400 do to the same 5%. If you will allow one extension, write the checkpoint and the cap now. A second extension invented on Friday is peeking.
- 5
Run the test without stopping when it looks good
Miller’s A/B-testing essay is the warning: peeking-and-stopping inflates false positives (26.1% vs 5% in his worst-case example; ten peeks turns a reported 1% into about 5%). The transfer: wait for the N you wrote, then read. Track source. One realistic channel beats three you cannot name.
- 6
Read k/n as the interval, then do the named action the same day
Look up the row, or compute Wilson in a spreadsheet. Honor miss, pass, or inconclusive. A pass is interest, not revenue — the usual next page is willingness to pay. Do not announce that the idea is validated.
What you leave with
The smoke-test sample-size contract (copy this)
If the sitting produced a mood, you did not finish. The page should have one filled rule a stranger could apply. Fill the brackets. Leave nothing as “users,” “enough traffic,” or “we’ll know it when we see it.”
The sample-size rule
Qualified visitor: [role + source — not “traffic”]. Conversion event: [the one action]. Pass line: [e.g. 5%]. Fail line: [e.g. 1%]. Fixed N: [e.g. 200 qualified visitors]. Optional checkpoint + cap: [one pre-declared extension / none]. Action on clear miss: [stop / rewrite the offer / kill the assumption]. Action on clear pass: [named next test — usually willingness to pay]. Action on inconclusive: [extend once to the cap / treat as a miss].
If any bracket still says “visitors,” “interesting,” or “later,” the contract is not written. Clicks from two jobs do not share an N. Compliments are not conversions.
What does not count as the signal
Not this signal: a Yibud score, a compliment, “I’d use that,” a Sean Ellis very-disappointed percent, interview saturation, a deposit or checkout (that is the WTP page), matching Blank’s five, a signed design partner, Nielsen’s 5 usability users. Optional later: those instruments have their own pages.
You may still take notes on praise or money. You may not let them fire this sample-size rule. k conversions out of n qualified visitors — read as an interval — fire this rule.
Example starting rule (template — the lines are yours, the interval math is not)
Qualified visitor: [named role + one channel]. Event: email waitlist. Pass line: 5%. Fail line: 1%. Fixed N: 200. No peeking. Clear miss: whole 95% Wilson interval below 5%. Clear pass: observed rate ≥ 5% and lower end > 1%. Inconclusive: extend once to 400, then treat as a miss if still straddling. Next test after a pass: a named deposit test.
5% and 1% in this template are blanks you fill, not industry benchmarks. This page invents no typical conversion rate. The interval numbers in the table are computed Wilson examples (z = 1.96).
The numbers
What k out of n can actually tell you (Wilson 95%, z = 1.96)
Computed examples. Not field results, not Yibud scores, and not a claim about what usually happens on a startup landing page. Use these values as printed.
| Observed (k/n) | Observed rate | 95% Wilson interval |
|---|---|---|
| 0/50 | 0% | 0%–7.1% |
| 0/100 | 0% | 0%–3.7% |
| 0/200 | 0% | 0%–1.9% |
| 1/50 | 2% | 0.4%–10.5% |
| 2/100 | 2% | 0.6%–7.0% |
| 3/100 | 3% | 1.0%–8.5% |
| 5/100 | 5% | 2.2%–11.2% |
| 2/200 | 1% | 0.3%–3.6% |
| 6/200 | 3% | 1.4%–6.4% |
| 10/200 | 5% | 2.7%–9.0% |
| 20/400 | 5% | 3.3%–7.6% |
| 5/50 | 10% | 4.3%–21.4% |
| 10/100 | 10% | 5.5%–17.4% |
The same observed 5% is 2.2%–11.2% at 100 visitors, 2.7%–9.0% at 200, and 3.3%–7.6% at 400 — quadrupling N roughly halves the width. Exact (binomial) zero-event upper bounds, 1 − 0.05^(1/n), for comparison with the rule of three: n = 100 → 2.95%; n = 300 → 0.99%. Those are not Wilson bounds — the Wilson upper bound for 0/100 is 3.7% (see table). Rule of three (Hanley & Lippman-Hand): 0 of n is at most about 3/n at 95% (0/100 → about 3%; 0/300 → about 1%).
The Wilson formula, once, in plain text
For k signups in n qualified visitors, let p̂ = k/n and z = 1.96. The 95% Wilson interval is [p̂ + z²/(2n) ± z · sqrt(p̂(1 − p̂)/n + z²/(4n²))] / (1 + z²/n).
Any spreadsheet can compute it. This page does not ship an interactive calculator. If you would rather not compute, use the table.
The fork
When sample size is the expensive unknown — and when it is not
This page decides only whether “what can this N tell me about one rate?” is the unknown you should buy evidence for now. How you build the page, interview coverage, money, usage surveys, which test to run, and decision contracts have their own pages.
Write the interval contract first
You already know how to put a page up. The expensive unknown is what N can tell you. Write the seven fields. Then buy or ask for traffic.
Build or pick the test first
You do not yet have a page, a CTA, or a reason this instrument beats an interview. Use the smoke, landing-page, fake-door, or cheapest-test page. An N without a test is a spreadsheet.
Move to money first
A pass already fired, or the unknown is already whether they will pay. Use the willingness-to-pay page. More uncharged visitors will not move money.
When a smoke-test sample-size rule is the expensive unknown
You can name a qualified visitor and a conversion event. You have not written N, the pass line, the fail line, or the inconclusive action. You are about to read a percent from a small n. That is this page. A weak distribution or demand line on a Yibud report sits near this fork — as a flashlight on which assumption the smoke test should target, not as an N.
When a smoke-test sample-size rule is the wrong next test
You cannot name who counts. You are still asking “would you use this?” You only needed a script, a usage survey of people who already did the real thing, a scarce-commitment partner, or a charge. You are mixing three channels into one n. You are running an A/B test of two headlines and calling it a smoke test. Run those pages instead. Counting unqualified clicks across mixed sources is theater.
Worked example
One interval contract you can copy (illustrative)
The week below is illustrative — names, product, and counts are invented so you can see the setup. It is not a study, not a conversion benchmark, and not a Yibud report. Copy the method, not the story. The interval numbers are the computed Wilson examples from the table.
The rate bet, written before the first click
Illustrative names: founder Mei Chen at Ledgerping, a B2B waitlist page for independent bookkeepers. Qualified visitor, written before launch: US independent bookkeepers who arrived from one named community post. Conversion event: email waitlist. Pass line: 5%. Fail line: 1%. Fixed N: 200 qualified visitors. Optional checkpoint: none. The tempting next step is to “see how the first fifty look.”
2/200 — clear miss
Observed 1%, interval 0.3%–3.6%. The whole interval sits below the 5% pass line. Even the optimistic reading misses. That is a miss you already named — stop or rewrite the offer. It is not a reason to buy another two hundred “because the copy was almost there.”
10/200 — clear pass
Observed 5%, interval 2.7%–9.0%. Observed rate is at the pass line, and the lower end (2.7%) is above the 1% fail line. That is a pass you already named — interest, not money. The action is the named willingness-to-pay test, not a factory.
6/200 — inconclusive
Observed 3%, interval 1.4%–6.4%. The range straddles the 5% pass line. That is inconclusive. Do the one action written before launch — extend once to the cap, or treat as a miss. Do not invent “promising” after you see six emails.
Why 5/100 is a fragile pass
A team that only had 100 qualified visitors and saw 5/100 (5%, interval 2.2%–11.2%) technically has a pass under the template: the observed rate sits at the 5% pass line, and the lower end (2.2%) is above the 1% fail line. It is a fragile pass. The interval runs from under half the pass line up to 11.2%, so the true rate could easily sit well below 5%. That is far less precise than 10/200 (2.7%–9.0%) or 20/400 (3.3%–7.6%) — which is why you fix N for the precision you need, not for the number that fits a calendar.
Write the contract before the first click. Read k/n as the interval. Honor miss, pass, or inconclusive. After a pass, Continue to the next cheapest test — usually money, not a factory. The method is the dated page, not the story you tell after.
Where Yibud fits
Use the free validator as a flashlight, then write the sample-size rule yourself
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 dimension, then a Day-7 Continue / Refine / Re-test / Stop signal. Sample-size work sits next to that assumption — use the weakest dimension to decide which claim the smoke test should target. Do not hope a mid-band score means “we already have enough visitors.” If you already have a report, start the contract with the named assumption — not with the overall score. Then run the test yourself. The method stands alone if you never open the analyzer.
Analyze my idea →Common mistakes
What usually wastes the sample-size rule
Stopping when the graph looks good
Miller’s A/B-testing essay: significance assumes N is fixed in advance; peeking-and-stopping inflates false positives (26.1% vs 5% in his worst-case example; ten peeks turns a reported 1% into about 5%). The transfer: do not close a smoke test the afternoon the waitlist “looks fine.” Wait for the N you wrote.
Reading 3/100 as “3%”
3/100 is 3% observed, and the 95% Wilson interval is 1.0%–8.5%. A lone percent hides the width. The same habit at 5/100 (2.2%–11.2%) is how a team calls a pass they cannot defend.
Counting unqualified traffic
A tight interval around the wrong people is still the wrong people. Track source. One realistic channel. Friends and newsletter readers who already like you inflate signups — debug the page with them, then read strangers separately.
Treating zero as “no data”
Hanley and Lippman-Hand: if none of n shows the event, the 95% upper limit is about 3/n (and agrees with the exact calculation to the nearest percentage point when n > 30). 0/100 → about 3%; 0/300 → about 1%. Exact (binomial) zero-event upper bounds here, 1 − 0.05^(1/n): 2.95% at n = 100, 0.99% at n = 300. Those are not Wilson bounds — the Wilson upper bound for 0/100 is 3.7% (see table). Zero is a result.
Treating a pass as willingness to pay
An email is interest. A deposit, a checkout, or a priced concierge is money. After a pass, open the willingness-to-pay page. Do not announce product-market fit.
Using the Wald interval at a tiny n
Brown, Cai and DasGupta: the Wald interval has erratic coverage, worse than textbooks suggest. They recommend Wilson or Jeffreys for n of 40 or less, and Agresti–Coull for larger n. A plus-or-minus from a class slide is a poor smoke-test instrument at fifty visitors.
Sources
Where these ideas come from
- Brown, Cai & DasGupta, “Interval Estimation for a Binomial Proportion,” Statistical Science 16(2):101–133 (2001) — Used for the finding that the standard Wald interval (p ± 1.96 · sqrt(p(1−p)/n)) has erratic coverage, worse than textbooks suggest; and for the authors’ recommendation of the Wilson interval or the Jeffreys interval for small n (40 or less) and the Agresti–Coull interval for larger n. This page does not invent other coverage claims. Author PDF
- Wilson, E. B., “Probable Inference, the Law of Succession, and Statistical Inference,” Journal of the American Statistical Association 22(158):209–212 (1927) — Used as the origin of the Wilson score interval. The table on this page is computed from that interval at 95% (z = 1.96).
- Hanley & Lippman-Hand, “If Nothing Goes Wrong, Is Everything All Right? Interpreting Zero Numerators,” JAMA 249(13):1743–1745 (1983) — Used for the rule of three: if none of n shows the event, the 95% upper limit is about 3/n; the approximation agrees with the exact calculation to the nearest percentage point when n > 30. Medical context; this page applies the same math to zero signups. Author reprint PDF
- Evan Miller, “How Not To Run an A/B Test” (2010-04-18) — Used for the A/B-testing facts that significance assumes sample size is fixed in advance; stopping as soon as significant inflates false positives; the worst-case example (checking after every observation up to 150, 50% base rate) gives 26.1% false positives instead of 5%; peeking 10 times means a reported 1% significance is actually about 5%; and the advice to decide sample size in advance. Labeled as A/B-testing context; the peeking lesson transfers.
In one paragraph
Summary you can quote
How many visitors a smoke test needs is a precision question, not a demand trophy. Write qualified visitor, conversion event, pass line, fail line, fixed N, and the named actions before launch. Read k/n as a 95% Wilson interval (Wilson 1927; Brown, Cai & DasGupta 2001 warn that the Wald interval has erratic coverage and recommend Wilson or Jeffreys at n of 40 or less). The same observed 5% is 2.2%–11.2% at 100 visitors, 2.7%–9.0% at 200, and 3.3%–7.6% at 400. Zero events are information: rule of three ≈ 3/n at 95% (Hanley & Lippman-Hand 1983; 0/100 → about 3%). Fix N; peeking-and-stopping inflates false positives (Evan Miller 2010, A/B-testing context: 26.1% vs 5% in his worst-case example). Yibud’s three-way template — miss if the whole interval is below the pass line; pass if observed rate ≥ pass line and the lower end is above the fail line; otherwise the pre-declared inconclusive action — is a template, not a law. A pass is interest, not money. Yibud scores are not conversion intervals. Matching the rule you wrote first is the method.
FAQ
Questions founders actually ask
How many visitors does a smoke test need?
Enough that the 95% Wilson interval is narrow enough to read against the pass line and fail line you wrote in advance. The same observed 5% is 2.2%–11.2% at 100 visitors, 2.7%–9.0% at 200, and 3.3%–7.6% at 400. There is no N that “validates” an idea. These are computed examples, not a law.
Is the 200–500 visitor rule of thumb wrong?
This page does not overturn it. The landing-page guide states a 200–500 visitor rule of thumb. Here is what 200 can and cannot tell you: 10/200 at 5% is 2.7%–9.0%; 6/200 at 3% is 1.4%–6.4% and may be inconclusive against a 5% pass line. Write N for the precision you need.
What if I get zero signups?
Zero is information. Hanley and Lippman-Hand’s rule of three: 0 of n means the true rate is at most about 3/n at 95% confidence (0/100 → about 3%; 0/300 → about 1%), and the approximation agrees with the exact calculation to the nearest percentage point when n > 30. Exact (binomial) zero-event upper bounds on this page, 1 − 0.05^(1/n): 2.95% at n = 100, 0.99% at n = 300. Those are not Wilson bounds — the Wilson upper bound for 0/100 is 3.7% (see table). 0/200 is 0%–1.9%.
Can I stop early if the rate already looks good?
Not if you wrote a fixed N. Evan Miller’s A/B-testing essay: significance assumes sample size is fixed in advance; stopping as soon as significant inflates false positives. His worst-case example (checking after every observation up to 150, 50% base rate) gave 26.1% false positives instead of 5%. Peeking 10 times means a reported 1% significance is actually about 5%. Decide N in advance. If you want one extension, write the checkpoint and the cap before launch.
Is a Yibud score a conversion interval? Do I need an account?
No. Yibud scores come from a deterministic rule engine on your idea inputs. A Wilson interval comes from k conversions out of n qualified visitors. You do not need an account. The method on this page stands alone. Run the free analyzer if you want a flashlight on a weak demand or distribution assumption before you write the contract. Optional language-model text only polishes prose.
Write the interval contract before the first click
Yibud scores weak demand and distribution assumptions with a deterministic engine. You write qualified visitor, event, pass line, fail line, and N while you are still honest — then a percent is an interval, not a vibe.
Generate a Yibud report →