What a marketing funnel is and where it leaks
A marketing funnel is the road one person walks from first seeing you to paying you, split into a few steps that can each be counted. Its use is not prediction, it is finding the step where people go missing the most.
- Lesson 2 of 8
- Beginner
- Free, no signup
The four steps one person walks before paying
The fewest steps that are still worth counting. Cutting it finer is worth it only when you have a number for each step.
-
1
Being seen
They have not heard of you and have no intention of buying. Your only job here is to exist where they already are.
-
2
Interest and comparison
Usually they are not comparing you with a competitor, they are comparing buying with not buying.
-
3
Contact or cart
They took a step: messaged, called, or put the item in the cart. No money has moved yet.
-
4
The purchase
The most expensive place to leak, because the whole cost of bringing that person this far has already been spent.
The narrowing shows the order of the steps only; there is no numeric ratio in it. Nor is the purchase the end of the story, only the end of this drawing.
Last checked: Facts and tool names in this lesson are re-checked against their sources on this date.
What does marketing funnel mean
A marketing funnel is that road from being seen to being paid, once it is split into a few countable steps. It is called a funnel because the number shrinks at each step: not everyone who saw you enters the site, not everyone who entered sends a message, and not everyone who messaged buys. The narrowing is unavoidable; what can be changed is how much it narrows at each step.
Four steps are enough for most small businesses: being seen, interest and comparison, contact or cart, and the purchase. It can be cut finer and in most places it is, but a nine step funnel where you cannot count a single step is a drawing rather than a tool. The simple rule: a step whose number you do not have is not a step.
And one point the definitions tend to drop: the purchase is not the end of the funnel, only the end of this drawing. For a business whose customer can buy again, the cheapest next sale comes from that same person and not from the top of the funnel.
What the funnel of a small coffee roastery looks like in practice
Picture a small roastery that sells both in person and through a simple online shop. You got its Instagram from a cafe and saw a post about the difference between freshly roasted beans and six month old ones. That is step one: being seen. You are not a customer yet and you have no intention of buying.
Three days later you see another post from the same account, this time about why coffee should not be kept in the freezer. Step two has now happened: interest. You are comparing, but not between this roastery and a competitor; between buying coffee at all and the instant jar you already have. Most real comparisons look like that, and not like the ones written into a strategy document.
Step three is the day you tap the link and put a 250 gram bag in the cart. Notice that no money has moved and you are still not a customer. Step four, the payment, may happen that moment or never; and if it does not, it is usually not because of the price of the coffee but because of something you only saw on the payment page itself.
Now look at the same road from the roastery owner side. They have four numbers and none of them come from the same place: post impressions from the Instagram panel, product page entries from the site analytics, carts created by the shop itself, and paid orders from the payment gateway. Four steps, four different sources. That scattering is what makes this hard, not the arithmetic.
Where does the funnel leak and how would you notice
A leak means an abnormal drop between two consecutive steps. The word abnormal matters: dropping is normal and every step loses people. The signal is a step compared with itself last month, not with a number you read in a foreign report.
Every step has its own sign and the signs do not resemble each other. If impressions are low, the message is not the problem; nobody has even seen it to judge it yet. If impressions are high and entries low, the first sentence and the first image did not do their job. If entries are high and carts or contacts low, the problem is elsewhere: the page, the price, how fast it opens, or the absence of anything that builds trust. And if carts get built but payment does not follow, look at the last step: a shipping cost that only appears here, a form that misbehaves on a phone, or a gateway throwing an error.
There are two traps inside this measurement and you need both. First, numbers from four different systems do not weld together: the Instagram panel does not know which of those impressions later entered the site, so any conversion rate joining the two is an estimate and has to be treated as one. Second, counting steps is not counting people: one person who opened the page three times is counted three times in most tools.
The depth of this, meaning where each number comes from and which of them can be leaned on, is the analytics lesson of this path. Keep only this from here: before believing any drop, make sure the tool reporting it is still running.
Four steps and the sign each one gives when it leaks
Each cell is one step, with what you see in the numbers when the leak is there written beside it.
-
Impressions are low
When the number at the top of the funnel is small, the rest of the numbers are beside the point. The problem here is not the message; nobody has seen it yet to reject it.
-
Seen but not clicked
High impressions, low entries. The first sentence and the first image did not do their job, or this person was never the audience for this message.
-
Enters but builds no cart
High entries, few contacts. The page, the price, how fast it opens, or the absence of anything that builds trust. This step is usually the most repairable one.
-
Builds a cart but does not pay
The thing that only appears on the payment page is usually the shipping cost. After that, a form that misbehaves on a phone and a gateway throwing an error.
These signs point at a direction, not a cause. One sign can have several causes and telling them apart is still your own work.
Which leak should you fix first
Our position here goes against what most sources say: rank leaks by the number of people lost, not by percentage. The two are not the same and the difference changes the decision. A drop with the worst percentage, if it sits low in the funnel where few people are left, may be a handful of people in total; a gentler percentage at the top of the funnel can take away several times that many.
The arithmetic is simple and done on paper: for each step, the number entering minus the number leaving. That number sets the order of work. Keep the percentages, but for seeing which step is sick, not for choosing which one to fix first.
One important constraint on that rule: the biggest leak is not always the most repairable one. If the main drop is at the top of the funnel because nobody knows you exist, fixing it is months of work; if the drop is at the payment step because a page opens slowly on a phone, it can be solved the same week. The real order is those two multiplied: how many people are lost, and how long it takes to stop it.
And one family of leaks lands almost always in that second group: the page itself. If you have visitors and none of them reach the cart, look at the speed and structure of that page before changing the message or the budget; if you have no time for that yourself, a site speed service targets exactly this one step.
Why a real buyer does not walk straight down
Now the honest word about the drawing itself: nobody actually walks this road in a straight line. A person sees you, forgets, sees you again somewhere else two months later, asks a friend, abandons a cart, and buys on the day their coffee runs out. The real road is a loop with several stops, not a pipe pointing downward.
So why keep the model? Because what we want from it is not predicting one person behaviour, it is comparing one business against itself across two months. For that job an imperfect fixed model beats a perfect fluid one: when you count with the same four steps every month, you see the change even if the steps themselves simplified reality.
There are two places where that simplification genuinely hurts and you should know both. First, repeat purchase: the classic funnel ends at the payment, and for a business like that roastery most of the revenue comes from the second and third purchases, which have no place in this drawing. Second, a purchase decided by several people: in selling to a company, the person who searches is usually not the person who pays, and a single person funnel does not show that.
The road that is actually walked
The same four steps, but this time in the shape they actually happen: with stops, forgetting and coming back.
- 1 Sees you
- 2 Forgets
- 3 Sees you again elsewhere
- 4 Asks around and compares
- 5 Buys the day they need it
This loop is a model too, not reality. Its only use is explaining why someone who left last month came back this month with no new advertising; for counting, those four steps are still better.
The fast path, with AI
What most people do with their funnel numbers is compare each step against a standard figure they heard somewhere. No such figure has been published for the Iranian market, and the comparison you actually need is with your own previous month. A language model is good at exactly this job, provided you put two constraints on it: bring in no outside number, and rank by people lost rather than by percentage.
- Collect the number for four or five steps for last month, from whatever panels you have. Write beside each number where it came from, because in the last step that note is what stops a wrong conclusion.
- Take the same steps for the month before that as well. Without a second month the output of this recipe is a story rather than a comparison; if you do not have one, record this month and come back next month.
- A cheap fast model of the Flash class is enough for this pass, because its work is subtraction and ordering. The current pick is listed in <a class="text-link" href="/en/ai/">the AI section</a> of this site. Keep the expensive model for the last step, where you have to decide which leak to take on first.
- Run the recipe below. If you see a phrase like industry average or standard rate in the output, reject it there; that number did not come from your data and it bends every decision after it.
- Pick one step and change only that. If you touch the page, the message and the budget together, next month the number has moved and you do not know which one moved it.
Copy-ready recipe
Funnel numbers for a {type of business}, two months:
Last month:
{step 1}: {number} (source: {where from})
{step 2}: {number} (source: {where from})
{step 3}: {number} (source: {where from})
{step 4}: {number} (source: {where from})
The month before: {the same four numbers}
Rule: bring in no outside number. Industry averages, standard rates and global benchmarks are forbidden. Work only with these numbers.
1) For each pair of consecutive steps, write how many people were lost (entering minus leaving) and what percentage.
2) Rank the steps by people lost, not by percentage. Write both rankings separately so I can see the difference.
3) Say which step got worse against the previous month and which got better, using only these numbers.
4) For the first step on that list, write three possible causes and beside each one what would rule it in or out.
5) If two steps come from different sources, warn me yourself that their ratio is an estimate.
6) Give no recommendation that is not attached to one of these four steps.Before you trust the output: Check three things yourself before believing the output. First, numbers from several systems do not weld together, so any ratio joining two steps from two sources is an estimate. Second, counting steps is not counting people; one person who visited three times has been counted three times. And third, the usual human constraint: the biggest leak is not necessarily the one you can fix this month, and the model does not know which job is within your reach.
AI in this kind of work
In this job AI does two things well and one thing badly, and the bad one is the more tempting. Well: computing and ordering your numbers, and pulling three hypotheses out of a pattern. Badly: telling you what a good number is. That number does not exist for your market, and instead of saying it does not know, the model produces one that looks like an answer.
Tools that actually help
- Claude It holds the no outside numbers constraint well and is accurate at arithmetic across a few lines of figures. Iran is on neither of the two Anthropic supported country lists; we read that on the Anthropic page itself.
- NotebookLM You supply the sources yourself and the answers stay tied to them, which is exactly what this job wants: when only your own numbers are in the notebook, an outside number has nowhere to come from. Google own help says it works in the same regions the Gemini app works in, and Iran is not on the supported countries list.
- Gemini It understands Persian itself, so a Persian table of numbers needs no translation first. Google own page says the app works in over 230 countries and territories and more than 70 languages, and Iran is not on that list.
Where it backfires
The main risk in this lesson is a number, not a text. Ask what a good conversion rate is for an online shop and you get an answer that has a figure, has a confident tone, and has no source. That this happens is not our guess: Anthropic keeps a page in its own documentation whose job is reducing exactly this production of plausible looking answers, and the existence of such a page means the behaviour is the default rather than the exception. The real danger is not the moment you see the number; it is two weeks later, when that same number is sitting in your budget decision and you no longer remember where it came from. Simple rule: any number in the output that was not in your input gets deleted. Alongside that, a payment gateway export and an order list are real customer data; strip the name, phone and address columns before pasting.
Sources: Anthropic docs: reduce hallucinations Anthropic: supported countries and regions Google: where Gemini Apps are available Google: NotebookLM availability
Where this advice stops
The funnel is a model and it falls short in three places. First, it assumes the purchase happens once; for a business whose customer returns, a large part of the revenue is outside this drawing. Second, it assumes one person decides; in selling to a company several people usually decide and the one who searched is not the one who pays. Third and more practical than both, the numbers in Iran come in pieces: the social panel, the site analytics and the payment gateway are three separate systems that do not join, so any conversion rate computed end to end across the funnel is an estimate. And last, the thing no funnel tells you: it shows the place, not the reason. Why people leave at the third step is a question to ask those people.
From our own work
We had a leak on this very site that appeared nowhere in the funnel, because the thing measuring it had switched itself off. We host the Google analytics script on our own domain, and until August the local file was named from its content hash; every time Google shipped a new version, the previous file was deleted that same moment. But the pages pointing at the old name stayed alive in the edge cache for up to thirty days. The result in the logs was this: not one of the historical names ever returned a 200, only 404s, which means analytics was effectively not running on cached pages. Since 2026-08-28 the file path is fixed and the hash sits in the query string, so an old page still reaches a working file. The lesson for this page: before believing any drop in your numbers, first make sure the thing counting them is still switched on.
Real follow-up questions
Is a marketing funnel different from a sales funnel?
They are one picture with two names and a border drawn differently in every place. Where the border is drawn is usually the moment someone turns from an audience into a lead: marketing before it, sales after. In a small business where the same person does both jobs, the split is more paperwork than practice.
How many steps should a funnel have?
As many as you can genuinely count. Four steps are enough for most small businesses, and adding a step only means something when you have that step number from somewhere. A seven step funnel where you know only two of the numbers is worse than a four step one, because five of its steps cannot be checked and it gives a false sense of precision.
Does a funnel mean anything for a business whose customers phone in?
It does, but the counting is manual and that is why most places never do it. The steps change: the call, the meeting or visit, the quote, the contract. A notebook beside the phone recording every call and its outcome is more accurate for this than any online tool, because you recorded the data yourself.