Optimising your funnel with AI: from spotting the problem to the test that measures the result

The website brings in visits, the campaigns bring in leads, and yet contracts aren’t growing. The temptation is to add something: a chatbot, a new landing page, an automation that sends three more emails. Before adding anything, it’s worth working out where the journey breaks down. A funnel usually jams at a few specific points, and often not where you look first.

The short answer is this: a funnel is optimised one bottleneck at a time, through a six-step cycle. You observe where people stop, diagnose the likely cause, turn it into a hypothesis that can be disproved, test it, measure the result and start again from the next bottleneck. Artificial intelligence speeds up the laborious steps: it reads hundreds of recorded sessions, flags anomalies in the data, groups the reasons leads are rejected and suggests copy variants. Which problem matters, how much risk to accept and whether a test really worked are decided by a person who knows the business.

This article is about the system that comes after the click: website, forms, sales response and CRM. If you’re interested in how AI manages the campaigns further upstream, Gaetano covers it in what to automate in advertising campaigns. If you first want to check that the numbers you’re working with are reliable, start from Milica’s article on how to verify data from click to sale.

What is a funnel bottleneck?

A bottleneck is the point in the journey where you lose a disproportionate share of people compared with the other steps, or where the time it takes to get through stretches out for no reason. It’s the step that limits the result of everything else: if little gets through there, improving the other stages achieves little.

In B2B and services, the most common bottlenecks are in five areas:

• the landing page, which promises something different from the ad or is slow on mobile;
• the form, with too many fields, validation errors or a submission that gives no confirmation;
• the response, when the lead comes in but nobody calls back for two days;
• qualification, when marketing and sales don’t share the same definition of a good lead;
• the offer, when the quote goes out late or doesn’t answer the question the customer asked.

Only the first two are on the website. The other three are in the processes, and they’re the ones an analysis limited to Google Analytics can’t see.

Why shouldn’t you automate a bottleneck before you’ve understood it?

Because automation makes whatever is already there faster, including the mistakes. If leads are being lost because the form asks for information the customer doesn’t have, an automated email sequence sends messages to people who never finished filling it in. If leads go cold because the salesperson doesn’t know which campaign they came from, a chatbot that collects more of them just lengthens the queue.

Automation comes after diagnosis. First you work out where and why the journey breaks down, then you decide whether the right fix is an automation, a change to the page, a rule in the CRM or a conversation between marketing and sales. How to set up automations when they’re needed is explained in our guide to business process automation with n8n.

The six-step optimisation cycle

The method is always the same, whether it’s an e-commerce site or an engineering company collecting quote requests. The tools and timescales change.

Step	Question	Typical tools	Where AI helps	Who decides
1. Observation	Where do people stop?	GA4 funnel exploration, recordings and heatmaps, CRM reports	Flags anomalies, summarises recorded sessions	Those with the full picture: marketing and sales together
2. Diagnosis	Why do they stop right there?	Sessions, sales notes, rejection reasons	Groups free text and recurring patterns	Those who know the product and the customers
3. Hypothesis	What do we change and what do we expect?	A shared document	Suggests wordings and variants	The person responsible for the funnel
4. Test	How do we verify without fooling ourselves?	A/B tests, platform experiments, process trials	Prepares copy and layout variants	Whoever approves risks and budget
5. Measurement	Did it really work?	Test data, CRM, revenue	Calculations, summaries	Whoever reads the data critically
6. Iteration	What’s the next bottleneck?	Hypothesis backlog	Reorders the backlog	The person responsible for the funnel

The cycle works if it stays small: one bottleneck at a time, one hypothesis per test, a date by which you look at the result.

Observing: how do you combine quantitative data and qualitative signals?

Quantitative data tells you where people stop; qualitative data suggests why. You need both, and on their own each will lead you astray.

The numbers: funnels, paths and anomalies in GA4

In Google Analytics 4, funnel exploration shows the steps people take to complete a task and how many reach each one. You can choose an open funnel (people enter at any step) or a closed one (they start from the first) and break it down by device, source or landing page. Path exploration does the reverse: it starts from an end point, such as a form submission, and shows where people came from to get there.

Analytics insights use machine learning to flag unusual changes. You can create up to 50 custom insights per property, including with the “has anomaly” condition, which leaves the choice of threshold to Google, and receive an alert by email. This is the first practical use of AI in observation: a system that looks at the data every day and raises its hand when something changes.

One limit to be aware of: with Consent Mode, for people who decline cookies, GA4 can estimate behaviour through modelling, provided the property exceeds minimum traffic thresholds. Part of the funnel numbers, therefore, is estimated. That’s fine for comparing steps, but not for counting people one by one.

The signals: recordings, dead clicks and sales notes

Microsoft Clarity is free and records sessions, creates heatmaps and calculates some frustration signals. Two are useful for finding bottlenecks on the website (official definitions): rage clicks, meaning repeated clicks in quick succession in the same area, and dead clicks, clicks on an element that produces no response within a reasonable time. A form that collects rage clicks on the submit button almost always has a validation or loading problem.

Watching five hundred recordings isn’t a sustainable task. This is where AI makes the difference: Copilot in Clarity summarises a single session or a group of sessions and the heatmaps in a few lines. Microsoft itself warns that generative AI can misinterpret things or produce incorrect information, so the summaries point you to the sessions to watch; they don’t replace watching them.

The most valuable qualitative signals, however, aren’t on the website: they’re the sales team’s notes, the rejection reasons in the CRM, the questions customers ask on the phone. An export of the free-text fields from the last six months, passed to a language model that groups them by theme, brings out in an hour patterns nobody had lined up before: “they always ask about delivery times”, “the price comes up too late in the conversation”, “the form doesn’t ask for the sector, so the salesperson has to call back”.

Diagnosing: how do you get from an anomaly to a probable cause?

An anomaly is a symptom. A form submission rate that drops by 30% in a week could be due to an error introduced with the latest website update, a change in the traffic mix (a new campaign bringing in less interested people), a tracking problem or seasonality. Diagnosis means narrowing down the possible explanations.

Three questions help:

1. Is it a data problem or a funnel problem? Before looking for commercial causes, check that the tag works, that the event is still being recorded and that consent hasn’t changed. Kohavi, Tang and Xu’s book on online experiments has a whole chapter on Twyman’s law: any figure that looks interesting or unusual is probably wrong. That applies to drops too.
2. Does it affect everyone or one segment? Break it down by device, source, landing page and new or returning visitors. If the drop is all on mobile from a certain date, the cause is technical; if it’s all on one campaign, it’s down to the audience or the message.
3. What do the people who stopped tell you? Recordings from that segment and the sales team’s notes confirm or rule out the lead.

AI helps you work through the combinations and bring the sources together, and here it pays to use it like a colleague who asks questions: “list five alternative explanations for this drop and the data we’d need to rule them out”. Choosing which lead to follow remains with whoever knows the context: a trade fair, a changed price list, a salesperson on holiday.

How do you write a testable hypothesis?

A testable hypothesis states what you’re changing, what you expect, why, how you’ll measure it and by when. If a result can’t disprove it, it isn’t a hypothesis.

The format we use:

If [change], then [metric] will go from [current value] to [expected value] because [observed cause], we’ll measure it with [source] by [date or volume].

A hypothetical example, with figures invented for illustration: “If we remove the VAT number and address fields from the quote request form, which the salesperson collects on the first call anyway, then the submission rate on mobile will go from 18% to at least 24%, because the recordings show people abandoning at the VAT number field. We’ll measure it with the submission event in GA4 and the number of requests in the CRM, after 400 visits per variant.”

Two details make the difference. The first: the control metric in the CRM, because a shorter form can bring more leads but of worse quality. The second: the minimum threshold set in advance, so nobody is tempted to stop the test when they like the result.

Here AI is useful for widening the range: you can ask it for ten different hypotheses for the same symptom, or five variants of the button text. Choosing which one to test and what value to expect is a decision for whoever takes responsibility for the result.

What do you fix without testing?

Not everything needs testing. A test is useful when there are two reasonable options and you don’t know which works better. When something is broken, you fix it.

This category includes:

• forms that don’t submit in certain browsers, required fields that shouldn’t be required, incomprehensible error messages;
• slow pages: Core Web Vitals have public, measurable thresholds, namely LCP within 2.5 seconds, INP within 200 milliseconds and CLS within 0.1, assessed at the 75th percentile of visits (web.dev);
• broken links, phone numbers that can’t be tapped on mobile, submission confirmations that never arrive;
• leads left without an owner in the CRM.

For these points AI is useful for finding them quickly (an automated check of forms, a read-through of error logs), and the fix is verified with a before-and-after check, with no need for control groups.

Testing: what kinds of test can you run?

It depends on what you’re changing and how much traffic you have. There are three families.

A/B tests on the website

Two versions of a page are shown to two randomly selected groups of visitors and the target metric is compared. Google Optimize, the free tool many people used, has not been available since 30 September 2023; Google suggests third-party tools integrated with Analytics as alternatives, such as AB Tasty, Optimizely and VWO.

The real limit is statistical. To detect a small difference you need a lot of people: Evan Miller’s rule of thumb estimates the sample per variant as 16 × p(1−p) / δ², where p is the current conversion rate and δ the minimum difference you want to detect (How Not To Run an A/B Test). With a rate of 5% and a difference of one percentage point to detect, you need around 7,600 visits per variant. Many B2B websites don’t collect that many in a reasonable time.

The same article explains the most common mistake: checking the test every day and stopping it as soon as it becomes significant. Doing that at every new observation, a test set up for a 5% false positive rate produces 26.1%. The sample size is decided in advance and respected.

Ad platform experiments

When the bottleneck is the message or the landing page of your campaigns, you can use the native tools. Google Ads custom experiments create a trial campaign from an existing one and split traffic and budget, using a cookie-based split in which each person sees only one version. Meta’s A/B test divides the audience into non-overlapping groups and compares one variable at a time.

Process trials

Many bottlenecks can’t be tested with software, because they lie in the way people work. Lead response time is the classic example. A study published in Harvard Business Review in 2011, The Short Life of Online Sales Leads, analysed 1.25 million leads from 42 US companies: those who contacted the potential customer within an hour were nearly seven times more likely to qualify them than those who waited even one hour longer. It’s a figure from fifteen years ago and from a different market, so it should be taken as an indication, not a prediction.

The trial works like this: for a month, every request that arrives during office hours gets a call within an hour, with an automatic alert to the salesperson and a reassignment rule if nobody responds. You compare the opportunity rate with that of previous months, keeping an eye on seasonality and the campaign mix. It isn’t a controlled experiment, and that should be stated: it’s a before-and-after trial that gives an indication, to be repeated if the result is uncertain.

The right automations, at this point, are those that make the fix stick: the alert to the salesperson, the reassignment, the score that ranks leads by priority.

Measuring: how do you know whether the test worked?

A test has succeeded when the target metric has improved, the control metrics haven’t got worse and the result holds beyond the test period.

To assess it you need three readings:

• the target metric, read only once the agreed sample has been reached;
• the downstream control metrics: qualified leads, opportunities, contracts in the CRM. An increase in forms submitted that doesn’t reach the pipeline isn’t a success;
• duration: new variants sometimes attract attention simply because they’re new. Checking again after a few weeks avoids mistaking novelty for improvement.

If the control metrics depend on the CRM, you first need to know that the CRM data is reliable and linked to the campaigns: we explain how in from lead to contract.

What can AI do, and what needs a human decision?

AI pays off where there’s a lot of material to read and little judgement to exercise. Decisions that involve a risk for the business stay with a person.

Activity	AI can	You need a person to
Monitoring data	Flag anomalies and changes in trend every day	Work out whether the anomaly is real or a tracking problem
Reading recorded sessions	Summarise groups of sessions and point out which to watch	Watch the flagged sessions and confirm the problem
Analysing the CRM	Group rejection reasons and free-text notes by theme	Interpret the themes in the light of product, prices and customers
Forming hypotheses	Suggest alternative explanations and variants	Choose which hypothesis to test and what effect to expect
Preparing the test	Write copy variants, calculate the sample	Approve messages, commercial promises and risks
Reading the result	Do the calculations, prepare a summary	Decide whether to apply, repeat or drop it
Making the fix stick	Manage alerts, reassignments, sequences	Set rules, exceptions and responsibilities

A simple rule: anything the customer sees or hears, meaning copy, prices, promises and timescales, goes through human approval before it goes live.

How do you choose where to start?

When there are ten hypotheses and only time for two, you need explicit criteria. The best-known frameworks are ICE (impact, confidence, ease), popularised by Sean Ellis, and PIE (potential, importance, ease), proposed by Chris Goward (a summary of the PIE method). We add a fourth item that carries a lot of weight on B2B websites: risk.

Criterion	Question	Example of a high score
Impact	How many people and how much value pass through that point?	The form that all quote requests come through
Confidence	How many independent sources confirm the problem?	GA4, recordings and the sales team all say the same thing
Effort	How much does the change and the test cost?	A copy change that can go live in an hour
Risk	What happens if the variant makes things worse?	Low: you can roll back in one click, no commitment to customers

The scores are there to start a discussion, not to calculate. Two people who give different scores to the same hypothesis have different information, and that’s exactly the conversation to have.

How should you handle personal data when you give it to an AI?

Session recordings, CRM exports and sales notes contain personal data. The GDPR requires data to be “adequate, relevant and limited to what is necessary” in relation to its purposes (Article 5 of Regulation (EU) 2016/679). In practice: before uploading an export to an AI tool, remove names, email addresses, phone numbers and company names, and use tools with contracts that exclude training on your data. Recording tools should be configured to mask form fields. If you want to understand how to set up the use of AI in your company with clear rules, we discuss it on our page about AI Act training.

Mistakes to avoid

Starting from the tool. “Let’s add a chatbot” isn’t a diagnosis. First the bottleneck, then the solution.

Optimising the wrong step. Improving a landing page’s click-through rate when the problem is that nobody calls leads back just makes the queue longer.

Stopping tests when you like the result. The sample size is decided in advance. Checking every day multiplies false positives.

Measuring only the first step. A shorter form that brings in worse leads isn’t an improvement. The control metrics are in the CRM.

Taking the AI’s summary at face value. A summary tells you where to look. Confirmation comes from looking.

Testing everything. Errors, slow pages and broken forms get fixed without testing.

Checklist for the first cycle

• Map the complete funnel, from click to contract, with someone responsible for each step.
• Check that the main events are recorded in GA4 and that requests reach the CRM with the source filled in.
• Build a funnel exploration in GA4 and break it down by device and source.
• Switch on recordings and heatmaps with field masking, and watch the sessions with rage clicks and dead clicks on key pages.
• Export the CRM’s rejection reasons and notes, remove personal data and have the themes grouped.
• Choose one bottleneck and write the hypothesis in the “if, then, because, we’ll measure it with, by” format.
• Fix whatever is broken straight away; for everything else, decide the type of test, the minimum sample and the control metrics.
• Set the date on which you’ll read the result and who decides what to do next.

Want to find out where your funnel is getting stuck?

We analyse the complete journey, from campaigns to contract: GA4 data, behaviour on the website, forms, response times and CRM. At the end you receive a map of the bottlenecks, the first three hypotheses to verify and a test plan proportionate to your traffic. We work on analytics and data, on websites and on the marketing automation that makes the fixes stick.

You can book a call with Federico or write to us via the contact form. If the bottleneck isn’t where you thought, we’ll tell you.

Frequently asked questions

How much traffic do you need for an A/B test?

It depends on your current conversion rate and the difference you want to detect. With a rate of 5% and a difference of one percentage point, you need around 7,600 visits per variant. With less traffic, it’s better to test bigger changes, use before-and-after trials or work on bottlenecks in the sales process.

Which tools do you need to get started?

To start, all you need is Google Analytics 4 with the main events configured, a free session recording tool such as Microsoft Clarity and a CRM in which every request has a source, a status and a rejection reason. An A/B testing tool is only needed when your traffic allows it.

Can AI optimise the funnel on its own?

No. It can flag anomalies, summarise sessions, group feedback and suggest variants. Choosing which problem to tackle, approving what the customer sees and deciding whether a test worked takes a person who knows the business and takes responsibility for it.

How often do you repeat the optimisation cycle?

The cycle ends when a test reaches the agreed sample. With the traffic many B2B websites get, each cycle takes two to six weeks. Monitoring for anomalies, on the other hand, is continuous.

What’s the difference between CRO and funnel optimisation?

CRO (conversion rate optimisation) usually focuses on the website’s pages. Funnel optimisation also covers what happens after the form is submitted: response, qualification, offer and closing. In B2B, the most costly bottlenecks often lie in this second part.