--- title: "How to Analyse a Funnel Using AI Tools" description: "Read a funnel drop-off without an analyst: the three questions to ask before you trust a step, the instrumentation gaps AI cannot see, and the anchor check." canonical_url: "https://builderscamp.com/guides/tools/ai-for-funnel-analysis" date_published: "2026-09-18" date_modified: "2026-09-18" author: "Andre Albuquerque, Mário Araújo" publisher: "Builders Camp" guide_class: "tools" --- # How to use AI for funnel analysis without trusting a broken step **TL;DR:** A funnel drop-off tells you where people leave, never why, and roughly half the dramatic drops in a first analysis are instrumentation rather than behaviour. Before you believe any step, run the anchor check: reconcile the top of the funnel against a count from a different system, such as billing or confirmation emails. AI is fast at rebuilding funnels that were never configured and at comparing them across segments, and blind to every event that was never fired. ## What is a funnel drop-off actually evidence of? It is evidence about where, and it is silent about why. Every step that loses people loses two different populations: the ones who tried and could not, and the ones who looked and decided not to. A payment step that drops 60 percent might be a broken card form on one browser or a price people will not pay, and the chart is identical in both cases. That is not an argument against funnels. It is an argument for treating the chart as a map of where to go looking rather than a finding in itself. The definitions underneath the shape are in [product funnel](https://builderscamp.com/guides/glossary/product-funnel). ## Ask three questions before you believe any step Is this step instrumented everywhere the step can happen? A checkout completed inside a native app and a checkout completed on web are frequently two different events, and a funnel built on one of them reports the other platform's users as drop-offs. Is the conversion window longer than the real decision time? A B2B trial where the buyer needs a procurement conversation will not convert inside a seven-day window, and a funnel configured for seven days will tell you the product failed when the calendar did. Does this step identify the same person as the previous one? This is the question almost nobody asks, and it is the one that quietly destroys most cross-session funnels. ## The gap AI cannot see: identity resets between steps Safari's Intelligent Tracking Prevention caps the expiry of client-side cookies at seven days, a change WebKit announced in February 2019, and it now deletes all of a site's script-writable storage after seven days of Safari use without user interaction on that site. Local storage, session storage and IndexedDB are covered, not just cookies. The consequence for a funnel is direct. A user who signs up on Monday, thinks about it, and comes back the following Wednesday arrives as a brand new identity. Step one has them, step four does not, and the export a model reads shows a clean drop-off with no marker of any kind saying the row was severed. Ask the model why conversion is low and it will produce a plausible answer about friction in step three, because plausible answers are what it has. The fix is not analytical. It is either server-side identity resolution keyed on a login, or an honest note on the chart that anything spanning more than a week is a floor and not a measurement. ## Two more places the export lies quietly Google Analytics condenses dimension values into a row labelled (other) once a table exceeds its row limit, and its documentation treats any dimension with more than 500 values as high cardinality for that reason. The worked example it gives is a property with 150,000 unique pages against a 100,000-row table limit: the least common 50,000 pages get folded into one row. Separately, Google applies data thresholds that withhold data from a report altogether when showing it could let someone infer an individual's identity from demographic or interest signals. Both mechanisms are documented, sensible and invisible to anything downstream. A model reading that export counts the rows that are there. ## Where AI is genuinely faster than the funnel report Two places. The first is rebuilding a funnel that was never configured. Testing a different step order in a dashboard usually means asking someone to build a new report; rebuilding it against a raw event export takes one prompt and the answer arrives while you still remember the question. That speed matters most when you are still arguing about what the funnel even is. The second is segment sweep. A funnel report gives you one funnel and a segment picker you click through by hand. A model given the same export will run the funnel across every segment in the file, rank the segments by how far each one's step-to-step conversion sits from the overall rate, and hand you the three worth opening. That is the analysis most teams intend to do and never finish. ## The verification step: anchor the top of the funnel to another system Before reading any step, reconcile step one against a count that comes from somewhere else entirely. Billing records, signup confirmation emails sent, contacts created in the CRM, rows in the accounts table. Pick whichever is closest to an independent ledger. If the funnel's first step is within a couple of percent of that number, the instrumentation is probably sound and you can read the rest. If it is 15 percent short, you have already found something more important than any drop-off further down, and every conversion rate below it is computed on a denominator that is missing people. Do this before the analysis, not after somebody challenges it in a review. Then check the last period. Funnels are cumulative over time, so the most recent cohort has had the least time to convert, and its conversion rate will look terrible for no reason other than the calendar. ## What do you do once the drop-off is real? Go and watch it. The funnel has told you which step and which segment; the reason lives in session recordings, support tickets from users who reached that step, and five conversations. A model can help you find the tickets that mention the step and cluster them, which is real work saved. It cannot watch someone fail at the form. If the step you are fixing is worth an experiment, the honest constraint is sample size, not instrumentation. That is a different discipline, covered in [A/B testing](https://builderscamp.com/guides/glossary/ab-testing). ## Is any of this worth doing while your instrumentation is bad? The strongest objection to running an AI funnel analysis is that it makes a weak measurement faster to misread. That objection is correct, and it is still not a reason to wait. A team with thin instrumentation has two options: spend a quarter fixing tracking with no idea which steps matter, or spend an afternoon reconstructing the funnel badly, find out which three steps carry the argument, and instrument those properly first. The second order is cheaper and the reconstruction pays for itself even when the numbers are wrong, because the anchor check tells you exactly how wrong. What you must not do is skip from the reconstruction to a roadmap. An analysis that failed its anchor check is a list of instrumentation bugs wearing the costume of a product insight. ## The instrumentation plan is the actual deliverable Most funnel analyses end with a slide. The ones that change anything end with an event plan: which events you now need, what properties attach to each, and, unusually, where you will track failure and not only success. Tracking only successes is why so many funnels can tell you that 40 percent of users did not finish and nothing at all about what they hit on the way. ## Start with the funnel you cannot currently rebuild Pick the flow your team argues about, reconstruct it from raw events rather than the saved report, and compare the two. The gap between them is usually where the last six months of conclusions came from. [See the Product Analytics bootcamp](https://builderscamp.com/bootcamps/product-analytics?utm_source=guide&utm_medium=organic&utm_campaign=ai-for-funnel-analysis) Product Analytics is built around exactly this problem: instrumenting events and properties so tracking reflects how value is actually created, defining activation instead of inheriting it, and reading retention without fooling yourself. Growth for Product Managers puts the funnel inside the wider growth loop, and both sit in the Growth Specialist Track. For the analysis mechanics on a single export, see [analysing product metrics with Claude Code](https://builderscamp.com/guides/tools/claude-code-metric-analysis-for-pms) and [how to use AI for product analytics](https://builderscamp.com/guides/tools/ai-for-product-analytics). ## Frequently asked questions ### What can AI do with a funnel that the funnel report cannot? Two things. It can rebuild a funnel that was never configured, straight from a raw event export, which means you can test a step order in minutes instead of filing a request. And it can run the same funnel across twenty segments at once and tell you which segment differs most from the rest, which is the question you actually had. ### Why does my AI-computed funnel disagree with the analytics tool? Usually the conversion window or the ordering rule. Most funnel reports count a conversion only if the steps happen within a set window and in a set order; a model rebuilding the funnel from raw events will count any user who did all the steps, in any order, ever. Neither is wrong, but they answer different questions. ### What is the anchor check? Reconcile the top of the funnel against a number that comes from a different system: billing records, signup confirmation emails sent, CRM-created contacts. If step one disagrees with that independent count by more than a few percent, the instrumentation is the finding and nothing downstream is worth reading yet. ### Why do returning users show up as new users? Safari's Intelligent Tracking Prevention caps client-side cookies at seven days and deletes script-writable storage after seven days of Safari use without interaction on the site. A user who comes back on day eight is a new identity. Any funnel whose steps span more than a week under-counts conversion, and the export gives the model no way to see it. ### Should a funnel be strictly ordered or any-order? Strict when the product forces the order, any-order when it does not. Most onboarding flows are less linear than the funnel chart implies, and a strict funnel applied to a non-linear flow reports a drop-off that is really users doing things in their own sequence. Run both and compare the gap. ### What does a drop-off actually tell you? Where, never why. A step can lose people because they tried and failed, or because they looked and decided not to. Those two need opposite fixes, and no funnel chart distinguishes them. Session recordings, support tickets and five user interviews do. ### Which Builders Camp bootcamp covers funnel work? Product Analytics covers instrumentation, activation definitions and account-level analysis across one week, 2 live sessions and 9 microlessons. Growth for Product Managers puts the funnel inside the acquisition, retention and monetization playbook across two weeks and 4 live sessions. ## Sources - [WebKit: Full Third-Party Cookie Blocking and More](https://webkit.org/blog/10218/full-third-party-cookie-blocking-and-more/) - [Google Analytics Help: About the (other) row](https://support.google.com/analytics/answer/13331684) - [Google Analytics Help: About data thresholds](https://support.google.com/analytics/answer/9383630) - [Builders Camp: Product Analytics bootcamp](https://builderscamp.com/bootcamps/product-analytics) ## How this guide was made Researched from Builders Camp's bootcamp, track and masterclass material and the sources listed on this page, drafted with AI, and fact-checked against every source cited.