Most "we need more sales" conversations start at the wrong end. The team debates a new homepage, a punchier headline, a discount, all before anyone can say where the customers are actually falling out. A funnel is the cheap fix for that: it forces you to find the leak before you reach for the paint.

The quick version

  • A funnel models the journey from first contact to a desired action as a series of steps, with fewer people surviving each one. Conversion optimisation is the work of widening the steps that leak the most.
  • The idea is over a century old, the AIDA sequence (Attention, Interest, Desire, Action) was described by Elias St. Elmo Lewis around 1898, and modern versions like Dave McClure's AARRR ("pirate metrics") just adapt it for digital products.
  • The single most useful move is to segment before you redesign: an overall conversion rate can look fine while one group (mobile, paid traffic, a particular country) quietly fails. Averages hide leaks.
  • Test changes honestly with an A/B test, but set the sample size first and don't "peek," or you will keep shipping changes that did nothing.

The idea in depth

The funnel is one of the oldest ideas in commerce still in daily use. Its spine is AIDA, Attention, Interest, Desire, Action, a sequence the advertising pioneer Elias St. Elmo Lewis sketched around 1898 to explain how a salesperson leads a stranger to a sale. Lewis originally wrote about it as principles rather than an acronym; the four-letter "AIDA" was coined later, by C.P. Russell in 1921. (The attribution isn't airtight, a 2023 paper has argued the underlying ideas trace to Frank Dukesmith and Arthur Sheldon, but Lewis is the name the literature has carried for a century, and the reference works still credit him.) The shape is the point: lots notice you, fewer get interested, fewer still want it, and a small remainder act. Draw that as widths and you get a funnel.

The shift that matters is to stop treating "conversion" as one number and start treating it as a chain. Each step has its own survival rate, and the chain is only as strong as its leakiest link. A 40% lift on a step that already converts well is worth less than a 5% lift on the step where you are haemorrhaging people. Name your steps, count who makes it through each, and the maths tells you where to spend your attention.

flowchart TD
  A(["Visitors, 10,000"]) --> B(["View product, 4,000"])
  B --> C(["Add to cart, 1,200"])
  C --> D{"Big drop here?"}
  D -->|"Only 300 reach checkout"| E(["Checkout, 300"])
  E --> F(["Purchase, 240"])
  D -.->|"75% leak at cart→checkout"| G(["Fix THIS step first
(illustrative figures)"])
A funnel turns "why aren't people buying?" into "why do three in four abandon the cart?", a question you can answer. Leaders Loop

Digital products gave the funnel a refresh. In a 2007 talk, Startup Metrics for Pirates, investor Dave McClure proposed AARRR, Acquisition, Activation, Retention, Referral, Revenue, nicknamed "pirate metrics" for the way it reads aloud. His argument, in his own framing, was that founders obsess over eyeballs, raw traffic, instead of whether those visitors activate, come back, pay, and bring friends. It is the same funnel logic, but it adds two stages a classic purchase funnel ignores: keeping the customer (retention) and turning them into a channel (referral). Pick the model that fits your business before you instrument anything. A one-off transaction lives or dies on the purchase funnel; a subscription product lives or dies on the retention step further down.

Why segmenting beats the average

Here is where most funnel work goes wrong. A team looks at a single blended conversion rate, declares it "fine," and moves on, or declares it "bad" and redesigns the most visible page. Both are guesses, because an aggregate number is an average of very different journeys. Practitioners call this the flaw of averages: a 3% overall conversion rate can hide the fact that mobile converts at 1% while desktop sits at 5%, or that paid traffic converts at half the rate of organic. The funnel looks healthy in total while one whole segment is failing badly.

Never optimise the average. The leak is almost always living inside one segment, and the average is the thing hiding it.

So: segment before you redesign. Before you touch a pixel, break the funnel down by the cuts that plausibly behave differently, device, traffic source, new versus returning, geography, product category. The diagnostic question is not "is our conversion good?" but "where is the same step dramatically worse for one group?" That is your leak, and it usually points at a specific, cheap cause: a checkout that breaks on small screens, a paid campaign sending the wrong audience, a shipping cost that only bites in one country. Fixing one bad segment often moves the headline number more than a redesign would, you are repairing a hole, not repainting a wall.

Testing a fix without fooling yourself

Once you have a leak and a hunch, the disciplined way to check whether your change helped is an A/B test: split traffic between the current version and the new one, and compare conversion. The logic is sound, but the practice is full of traps that produce confident, wrong conclusions, which is worse than not testing at all, because it manufactures false certainty.

The deepest trap is peeking. The industry norm is 95% statistical significance, meaning roughly a 1-in-20 chance the result is a fluke. But if you keep checking an in-progress test and stop the moment it crosses that line, you quietly inflate that false-positive rate well past 5%, every extra look is another roll of the dice, and a handful of them pushes the real error rate up sharply. That is one of the classic A/B-testing pitfalls. The fix is unglamorous: decide the sample size and duration before you start, then leave the test alone until it finishes. Two companion disciplines matter as much, change one thing at a time, so you actually know what caused the lift, and run long enough to cover a full business cycle, so a Tuesday-versus-Sunday quirk or a one-off promotion doesn't masquerade as a real effect.

flowchart TD
  A(["Spot the leakiest step
(segmented funnel)"]) --> B(["Form one specific
hypothesis about why"])
  B --> C(["Set sample size &
end date up front"])
  C --> D(["Run A/B test,
change one thing"])
  D --> E{"Significant at the
pre-set end?"}
  E -->|"Yes"| F(["Ship it, then
re-segment the funnel"])
  E -->|"No"| G(["Keep the original,
try the next hypothesis"])
The loop: diagnose, hypothesise, pre-commit, test, decide, then start again on the next-leakiest step. Leaders Loop

An honest limitation. The funnel is a model, and like any model it lies a little. Real journeys are not a tidy one-way slide: people loop back, leave and return weeks later, research on a phone and buy on a laptop, or arrive already convinced by a friend. A strict linear funnel undercounts all of that, which is one reason retention-and-referral models like AARRR exist. The funnel is also silent on why a step leaks, it tells you where, not the cause; the cause comes from session recordings, support tickets, and talking to the people who dropped out. And a test only proves a change moved a number in a window of time; it can't promise the lift will last. Treat the funnel as a map that shows you where to dig, not as the thing being measured.

A worked example

Take a small online course business, call the founder Priya. (Illustrative figures throughout; this is a teaching example, not real accounts.) Sales have stalled, and the team's instinct is to rebuild the sales page. Before spending three weeks on it, Priya draws the funnel: of 10,000 monthly visitors, 4,000 view a course, 1,200 add it to the cart, but only 300 reach checkout and 240 buy. The biggest drop by far is cart to checkout, three in four people who wanted the course abandon it before paying. The sales page she was about to rebuild sits earlier and is converting fine.

Then she segments that one leaky step. On desktop, cart-to-checkout holds at a respectable rate; on mobile, it collapses. A quick look on her own phone finds the cause in thirty seconds: the discount-code field is broken on small screens, so mobile buyers get stuck and give up. The fix is an afternoon's work, not a redesign. She ships it as a proper A/B test, mobile traffic split between old and new checkout, sample size and a two-week end date set in advance, and no peeking, and at the end the new version wins clearly. The headline conversion rate moves more from that one repair than the sales-page rebuild would have, and it cost a fraction of the time. The funnel didn't tell Priya the answer; it told her where to look, and the segmentation told her where to look harder.

Frequently asked questions

What's the difference between a funnel and conversion optimisation?

The funnel is the map, the named steps from first contact to the action you want, with a survival rate at each one. Conversion optimisation is the work you do with that map: finding the leakiest step, forming a hypothesis about why, and testing a change. The funnel diagnoses; optimisation treats. Doing the second without the first is how teams end up redesigning pages that were never the problem.

How many steps should my funnel have?

As few as cleanly describe the real journey to your one key action, typically four to six. Too few and you can't locate the leak (one giant "consideration" step tells you nothing); too many and every step looks fine because the drops are spread thin. Pick the moments where a person makes a real decision or hits real friction, and measure those.

Should I use AIDA, AARRR, or something else?

Match the model to the business. A one-off purchase fits the classic purchase funnel (roughly AIDA). A subscription or product-led business should use something like AARRR, because its money is made in the retention and referral stages a purchase funnel ignores. The labels matter far less than choosing steps that reflect how your customers actually decide and stay.

How much traffic do I need to A/B test?

Enough that a real difference would show up above the noise, which depends on your current conversion rate and how big a change you're hoping to detect. The honest answer for low-traffic businesses is that proper A/B testing may be impractical; a test that never reaches significance is just a slow guess. In that case, lean on qualitative evidence (session recordings, talking to people who dropped out) and obvious-fix changes rather than pretending a tiny test is conclusive.

Isn't this just for e-commerce websites?

No. Any process where people progress through stages and some fall out is a funnel: a B2B sales pipeline, a hiring process, a free-trial-to-paid path, a charity's donor journey. The discipline, name the steps, count the drop-offs, segment before you act, test changes honestly, transfers directly. The dashboards differ; the thinking doesn't.

Related in the Toolkit

A funnel is one view of the wider go-to-market machine: how you choose to reach buyers at all is GTM strategy & motions, and the retention end of an AARRR funnel is really the territory of engagement and retention.

Where to go next