Finding friction is a search problem, and most people run the search backwards. They start with a list of best practices, check their site against it, and end up with forty things that could be improved and no idea which one is costing them anything. That's an inventory, not a diagnosis.
Run it forward instead. Friction has a location. Your job is to narrow the search space until you're standing on it — and you can do most of that narrowing with instruments you already have.
Instrument one: walk your own site as a stranger
The cheapest, most-skipped diagnostic. On a phone, on cellular data, not logged in, not on your office network, ideally after a night's sleep so the page is unfamiliar.
Do the thing you want visitors to do. Buy something. Sign up. Book the call. And keep a note of every moment you had to think, wait, hunt, or guess.
The reason this works is that you cannot see your own site. You know the shipping is calculated at checkout. You know the second button goes somewhere different. You know what the form field means. Every one of those pieces of knowledge is invisible scaffolding a first-time visitor doesn't have, and every gap between what you know and what the page shows is a candidate leak.
This produces hypotheses, not answers. Write them down. You'll test them with instrument two.
Instrument two: read the shape of the drop, not the size
A site-wide conversion rate is a summary statistic, and summary statistics hide locations. Two percent tells you nothing about where the other ninety-eight went.
What you want is the step-to-step view. For each transition in your flow — landed → viewed product, viewed product → added to cart, added → started checkout, started → completed — what fraction continues?
Then look for the cliff. Most transitions lose a modest, unremarkable share. One usually loses far more than its neighbours. That's not a rounding difference; that's your leak, and it will be obvious once you stop averaging over it.
If you don't have step data yet, you can approximate: compare the count of people who reach a page against the count who complete its action. Rough numbers find cliffs fine. Precision matters later, when you're measuring a fix.
Instrument three: compare the same flow across devices
Run the identical funnel for mobile and desktop separately and put the completion rates side by side.
A gap is normal. A large gap is diagnostic, and it almost never means mobile visitors want your product less. It means your interface is physically harder on a phone — a keyboard covering the field, a tap target too small, autofill that doesn't fire because inputs lack the right attributes, a validation error scrolled off-screen. Same intent, same funnel, different friction.
This instrument is powerful because it controls for everything else. Offer, price, traffic source, copy — all held constant. The delta is interface.
Where the three instruments converge
Now cross-reference. If your walk-through flagged the address step as tedious, your funnel shows a cliff between "started checkout" and "payment," and your mobile rate at that step is half your desktop rate — you have not guessed. Three independent instruments agree, and you've located the friction without installing anything.
That convergence is the goal. One instrument produces a suspicion; three produce a diagnosis.
What these instruments cannot see
Be clear-eyed about the ceiling, because it's where most people quietly start guessing again.
You can find the step. You cannot find the field. Your funnel shows people abandoning inside a form. It does not show you which input they focused, hesitated over, errored on, and quit at. Two forms with identical abandonment can be failing at completely different fields, and the fix for one does nothing for the other.
You can see that they left. You cannot see what they tried. Rage clicks on a non-clickable element, a dropdown that never opened, a submit button that silently failed on their browser — none of this appears in a funnel. It appears as an unexplained drop.
You cannot see the counterfactual. After you change something, "the number went up" and "the number went up because of the change" are different claims. Separating them requires watching that specific step over time with enough traffic to distinguish a signal from a Tuesday.
Everything above the ceiling is inference. Good inference — it will find your biggest leak more often than a checklist will — but inference.
Ordering what you find
You will find several things. Fix one.
Rank candidates by how many people meet them multiplied by how badly they stall. A friction point on a page that ten people a month see is not your problem, however ugly. A half-second of hesitation on a step every buyer crosses is.
Then change one thing, and watch that one step. Changing five things and watching your total is how teams learn nothing over a whole quarter — the number moves, and no one can say which change moved it, so the next decision is a guess again.
Defrixa's free scan runs the structural half of this for you: it reads a public URL, scores its friction deterministically, and names the single biggest structural issue rather than handing you a checklist. The behavioral half — which field, which step, and whether your fix actually worked — is what the tracking snippet is for.
Common questions
The walk-through works at any volume. The funnel-shape and device-comparison instruments need enough visitors that a step's drop isn't a handful of people. Below that, trust direction over time rather than any single week's number.
They show attention and clicks, which is useful for the "unseen" class of problems. They don't tell you which step of a flow costs the most, and they won't tell you whether a change worked. They're an instrument, not the method.
Then you haven't found it, and you shouldn't ship a fix yet. Disagreement is information: it usually means the leak is somewhere you haven't instrumented, often inside a form or an embedded third-party widget.
No. It's frequently the checkout on ecommerce sites, and frequently not on lead-gen and SaaS sites, where the leak tends to sit at the moment of signup or at the page that was supposed to answer an objection and didn't.