A sales funnel behaves like a lota with holes because both obey the same conservation law: the level rises only while inflow exceeds outflow, and it settles where the two are equal. For a customer base with new customers \( \lambda \) a month and monthly churn \( c \), that level is \( N^* = \lambda / c \): 16 new customers a month at 2% churn tops out at about 780 customers, no matter how long you wait, and halving churn doubles the ceiling just as surely as doubling leads.
- \( N^* = \lambda / c \): a customer base stops growing where new customers equal churned ones.
- 780 customers is the ceiling for 15.6 new customers a month at 2% monthly churn.
- Halving churn raises the ceiling exactly as much as doubling leads, with no extra media.
- 0.78% of leads become customers on Implisit's B2B averages (13% then 6%).
- 16%+ higher value from referred customers in Schmitt et al.'s bank study.
The analogy breaks in one instructive place: a lota's leak slows as the water drops (Torricelli's law), while churn scales with the base, which is why fixing retention beats buying more leads.
Every Indian household knows the lota, the small, rounded, usually brass or copper water vessel used in homes and rituals since at least the second millennium BCE. Put a few holes in it and pour water in from a tap, and you have built a working model of a business.
- The tap is your marketing.
- The holes are every point where a prospect drops out or a customer leaves.
- The water level is your customer base.
This post solves the lota first, then the funnel, and then shows where the two equations part ways.
How fast does water leak out of a lota?
Water leaves a hole at the speed it would reach by falling from the water surface, \( v = \sqrt{2gh} \), so the leak is fast when the lota is full and slows as it empties. That is Torricelli's law, a special case of Bernoulli's principle.
Real holes discharge less than the ideal, because the jet narrows just outside the opening (the vena contracta); for water through a round hole the discharge coefficient is about \( C_d = 0.65 \). So the outflow rate through a hole of area \( a \) is
Assumptions for a household lota (my estimates): a roughly cylindrical body 12 cm across, so cross-section \( A = \pi (0.06)^2 = 0.0113 \) m²; one 3 mm hole in the base, \( a = 7.07\times10^{-6} \) m²; water 10 cm deep.
- Mass balance. The level falls as water leaves: \( A\,\frac{dh}{dt} = -C_d a \sqrt{2gh} \).
- Separate and integrate. \( \int_{H}^{0} \frac{dh}{\sqrt{h}} = -\frac{C_d a \sqrt{2g}}{A}\int_0^T dt \Rightarrow 2\sqrt{H} = \frac{C_d a \sqrt{2g}}{A}\,T \).
- Drain time. \( T = \frac{A}{C_d a}\sqrt{\frac{2H}{g}} \).
- Plug in. \( T = \frac{0.0113}{0.65 \times 7.07\times10^{-6}} \times \sqrt{\frac{0.2}{9.81}} = 2{,}459 \times 0.143 \approx 351 \) s, just under 6 minutes. (my assumed dimensions and Torricelli's law)
Calculated in this post from the assumed lota size above (my estimates), a discharge coefficient of about 0.65 and Torricelli's law (Wikipedia).
Now open a tap into it at a steady \( Q_{\text{in}} \). The level climbs until outflow matches inflow, \( Q_{\text{in}} = C_d a\sqrt{2gh^*} \), so
At 5 mL per second, \( h^* = \frac{1}{19.62}\left(\frac{5\times10^{-6}}{4.6\times10^{-6}}\right)^2 \approx 6 \) cm. Double the tap to 10 mL/s and the level quadruples to about 24 cm, so the lota overflows.
Calculated in this post from the same assumed hole size, the 0.65 discharge coefficient and Torricelli's law (Wikipedia), with a tap rate of 5 mL per second as a premise.
In a lota, more inflow is rewarded more than proportionally, because the leak only grows with the square root of the level. Hold on to that; it is exactly where a business differs.
Quibi, the short-video streaming app led by Jeffrey Katzenberg and Meg Whitman, raised $1.75 billion before it launched on 6 April 2020 at $4.99 a month. The tap opened wide: in the first three days about 9.1 lakh (910,000) people signed up for its 90-day free trial, according to the app-analytics firm Sensor Tower.
When those trials ran out in early July, Sensor Tower estimated that at most about 72,000 of them, roughly 8%, had become paying subscribers. Quibi said the estimate was wrong "by an order of magnitude" but did not publish its own figure. On 21 October 2020 the founders confirmed the service would close, and it shut down around 1 December. The money filled the lota; the base drained it.
Next TV: Quibi netted only 72K paying subscribers from initial wave of signups (July 2020)What is the lota equation for a customer base?
The same mass balance, with one change: customers leave at a rate proportional to how many you have, \( cN \), not to the square root. The base therefore settles at \( N^* = \lambda/c \).
Variables: \( L \) leads a month; stage conversion rates \( p_1, p_2, \dots \); new customers \( \lambda = L\,p_1 p_2 \cdots \); monthly churn rate \( c \), the share of customers who leave each month.
- Inflow. In Implisit's analysis of B2B CRM data, leads became opportunities at 13% on average and opportunities became deals at 6%. With \( L = 2{,}000 \): \( \lambda = 2{,}000 \times 0.13 \times 0.06 = 15.6 \) new customers a month.
- Balance. \( \frac{dN}{dt} = \lambda - cN \).
- Steady state. Set the change to zero: \( N^* = \lambda/c = 15.6/0.02 = 780 \) customers at 2% monthly churn. (with Implisit's 13% and 6% and 2% churn as an example input)
- Path over time. Solving the linear equation from zero gives \( N(t) = N^*\left(1 - e^{-ct}\right) \). The time constant is \( 1/c = 50 \) months; reaching 90% of the ceiling takes \( \ln 10 / c \approx 115 \) months. (calculated from the 2% example churn input; see Implisit's rates above)
- The two levers. Doubling leads to 4,000 gives \( N^* = 1{,}560 \). Halving churn to 1% also gives \( N^* = 1{,}560 \), and costs no extra media. (same Implisit rates, example inputs)
Calculated in this post from Implisit's 13% and 6% averages (Salesforce blog, 2014) and these example inputs: 2,000 leads a month and 2% monthly churn, which are illustrative, not measured.
See the data as a table
| Churn 4%/month (customers) | Churn 2%/month (customers) | Churn 1%/month (customers) | |
|---|---|---|---|
| 0 | 0 | 0 | 0 |
| 6 | 83 | 88 | 91 |
| 12 | 149 | 166 | 176 |
| 24 | 241 | 297 | 333 |
| 36 | 298 | 400 | 472 |
| 48 | 333 | 481 | 595 |
| 60 | 355 | 545 | 704 |
| 84 | 376 | 635 | 887 |
| 120 | 387 | 709 | 1090 |
Source: My calculation from N(t) = (λ/c)(1 − e^(−ct)), with λ from Implisit's 13% and 6% stage conversion averages · Months on the horizontal axis. Same inflow, three leak sizes: the ceilings are 390, 780 and 1,560.
Note what the steady state does to marketing maths. At 4% churn, the business plateaus at 390 customers within about five years, and from then on every rupee of acquisition only replaces customers who leaked out.
A founder looking at a flat customer count often blames lead volume. The equation says to look at the base of the lota first.
Size your leaky lota
Where does the lota analogy break?
In two places: the leak law and the holes themselves. Both breaks make retention more valuable than the lota suggests.
| Property | Lota | Customer base |
|---|---|---|
| Outflow law | \( \propto \sqrt{h} \) (Torricelli) | \( \propto N \) (constant churn) |
| Level vs inflow | Grows with the square of inflow | Grows in proportion to inflow |
| Hole size over time | Fixed | Shrinks as a cohort ages |
| Bigger tap | Wins disproportionately | Moves the ceiling only in proportion |
The leak law
A lota leaks in proportion to \( \sqrt{h} \), so its level grows with the square of inflow. A customer base leaks in proportion to \( N \), so its level grows only linearly with inflow.
In a lota, a bigger tap wins disproportionately; in a business it does not. You cannot outrun churn with leads; you can only move the ceiling in proportion.
The holes shrink
A lota's hole is the same size whether the water is old or new. Customer churn is not. Within any group, the customers most likely to leave do so first, so the survivors are increasingly loyal and the observed retention rate rises over time.
Peter Fader and Bruce Hardie built a simple model of exactly that, the shifted-beta-geometric, and showed that ignoring it produces large forecasting errors.
Using survival data for two segments of a subscription business over seven years, the authors first fit the regressions an analyst would usually try. Projected to year 12, the linear model underestimated survival by 81%, the exponential by 30%, and the quadratic overestimated it by 92%.
They then model each customer as having their own constant churn probability, drawn from a beta distribution across customers. Aggregated, this produces retention rates that rise with tenure as high-churn customers drop out, and it forecasts accurately in a spreadsheet. The limit: it is for contractual settings where you can see a customer leave; for D2C repeat buying you need the non-contractual cousins of the model.
Read the paper (free)For a lota, that is like holes that seal themselves as the water ages. Practically, it means two things:
- The widest hole is early. The most expensive leak is in the first weeks, when the least committed customers leave; onboarding is where the hole is widest.
- Averages hide the leak. A churn rate averaged across the whole base hides it: measure churn by cohort, as the D2C maths in CAC payback and LTV requires.
Which holes should you patch first?
The ones with the most water pressure behind them: the stages where the most people are lost and the churn of your largest cohort. Rank them with the flow equation, not by instinct.
A funnel is a chain of holes, and losses multiply. If a prospect passes each stage with probability \( p_i \), overall conversion is \( \prod p_i \). Improving any stage by the same relative amount raises the output by the same relative amount, so the choice between stages is about cost.
With Implisit's averages, overall lead-to-deal conversion is \( 0.13 \times 0.06 = 0.78\% \): 99.2% of the water entering the top never reaches the base.
Calculated in this post from Implisit's 13% and 6% (Salesforce blog, 2014).
| Hole | Lota version | Funnel version | Typical fix |
|---|---|---|---|
| Wrong water in the tap | Pouring in sand with the water | Low-intent leads from bought lists (under 0.1% lead-to-deal in Implisit's data) | Change the source, not the volume |
| Hole at qualification | Side hole near the top | Leads that never become opportunities (87% lost on average) | Faster response, clearer qualification |
| Hole at the deal | Side hole lower down | Opportunities lost (94% on average) | Better proof: case studies that sell |
| Hole in the base | The constant drip | Monthly churn \( c \) | Onboarding, product fit, service |
| Lining that seals holes | Waxing the inside | Customers who arrive pre-committed, such as referrals | Referral programmes |
That last row has strong evidence behind it. Referred customers leak less, which is the closest thing marketing has to a better lota.
The authors tracked about 10,000 customers of a leading German bank for almost three years, comparing those acquired through a referral programme with similar non-referred customers acquired at the same time. Referred customers had a higher contribution margin, though the gap narrowed over time; a higher retention rate, and that gap persisted; and an average value at least 16% higher. The size of the advantage varied by segment, so referral programmes should be targeted. The limit: one bank in one country, and referral rewards themselves have a cost that must come out of that 16%.
Read the paper (free)Referrals also sit near the top of the conversion tables in B2B lead sources ranked by conversion, so they fix two holes at once: better water into the tap, and less leaking out of the base.

So, does a funnel really leak like a lota?
Both are mass balances, so the core result transfers exactly: the level stops rising where inflow equals outflow, \( N^* = \lambda/c \), which is about 780 customers for 15.6 new customers a month at 2% monthly churn. The corrections: a lota's leak follows Torricelli's \( \sqrt{h} \), which rewards a bigger tap, while churn is proportional to the base, so leads raise the ceiling only linearly; and churn holes shrink as cohorts age, as Fader and Hardie's model shows. Both point the same way. Patch the base before you buy a bigger tap.
The working rule that follows: whenever a team asks for more lead budget, compute \( N^* \) at the current churn first. If the base is already near its ceiling, the extra budget will mostly refill the leaks, and the cheapest growth sits in onboarding, service and referrals. It is a cheap check to run on any pipeline, including B2B ones like those on my work page.
What's new
- AppsFlyer: India shopping-app installs up 85%, but day-30 retention fell 41% in two quarters CXOToday, Sept 2026
- Fader and Hardie's retention model revisited with the beta-discrete-Weibull, letting churn rise or fall with tenure Bruce Hardie, latest research
- India's ad market grew 8.3% to ₹1,21,339 crore in 2025, making every acquired customer costlier to replace BestMediaInfo, Feb 2026
What if next?
- The chai-wala has better retention than your app His base never drains, and habit research explains why.
- Bollywood trailers are A/B tests How studios find the leak before release day.
- The wedding card is the perfect piece of direct mail Personal, physical and impossible to ignore.
Questions people ask
What is a leaky bucket in marketing?
It is the idea that a business loses customers while it acquires new ones, so growth depends on the balance. Mathematically, the base levels off at new customers per period divided by the churn rate.
Is it cheaper to reduce churn or to get more leads?
In the steady-state equation, halving churn and doubling new customers raise the ceiling equally. Which is cheaper depends on your costs, but halving churn needs no extra media, and retained customers usually earn more each year they stay.
What is a good B2B funnel conversion rate?
In Implisit's aggregated CRM data, leads became opportunities at about 13% and opportunities became deals at about 6%, with referrals converting best and bought lists worst. Benchmarks vary widely by deal size and channel.
What is Torricelli's law?
Water flows out of a hole at the speed it would gain falling from the surface, \( v = \sqrt{2gh} \), reduced by a discharge coefficient of about 0.65 for a round hole. Evangelista Torricelli found it in 1643.
Sources
- Wikipedia: Torricelli's law (source for \( v = \sqrt{2gh} \), the 1643 date, the 0.65 discharge coefficient and the drain-time formula)
- Wikipedia: Lota (vessel) (source for the lota's history, materials and uses)
- Salesforce blog, Implisit data: B2B sales benchmark research (2014; source for the 13% and 6% stage conversion rates and lead-source differences)
- Fader and Hardie, Journal of Interactive Marketing: How to project customer retention (2007; source for the regression errors and the shifted-beta-geometric model)
- Schmitt, Skiera and Van den Bulte, Journal of Marketing: Referral programs and customer value (2011; source for the 16% value difference and retention result)
- Hardie and colleagues: How to project customer retention revisited (source for the beta-discrete-Weibull extension)
- Next TV: Quibi netted only 72K paying subscribers from initial wave of signups (Sensor Tower) (July 2020; source for the 910,000 trial sign-ups, about 8% conversion, $4.99 price, $1.75 billion raised and Quibi's response)
- Wikipedia: Quibi (source for the 21 October 2020 shutdown announcement and closure around 1 December 2020)
- CXOToday: AppsFlyer report on Indian shopping apps (Sept 2026; source for install growth and retention decline)
- BestMediaInfo: dentsu-e4m India AdEx report (Feb 2026; source for 2025 ad spend)
