AI conversion tools vs manual CRO: what actually differs

What manual work is genuinely better at

Novelty. A change nobody in the category has tried, a hypothesis that depends on knowing your customer, a diagnosis that requires understanding why this audience distrusts this particular claim: these come from a person, and current tools are not close.

Commercial context is the second. An analyst knows the margin on the product being pushed up the collection, knows a supplier is unreliable in November, and knows the founder will not approve the copy however well it tests. A detector sees a conversion rate.

Third, and most underrated: a person can decide a finding is not worth acting on. Judgement is largely the ability to discard, and discarding well is a skill that does not show up in any feature comparison.

What automated tools are genuinely better at

Coverage, which is not a small advantage dressed up. A person auditing a store looks at the pages they suspect, on the device they use, in the week they were paid to look. A detector runs on every session on every template continuously, which means it finds the problems that are unglamorous, persistent and expensive.

Consistency is the other half. A manual audit varies with who did it, how tired they were and how the last one went. An automated pass is the same in July and December, which makes its output comparable over time: you can see a problem appear the week a theme changed.

And speed of estimating. Working out what a specific friction costs each month needs cohort-comparison arithmetic that is tedious by hand and quick by machine. That is why manual programmes so often rank by gut feel: not because analysts do not want the number, but because getting it for twenty candidate problems is a week of work.

The failure mode of each

Manual programmes fail by running too few tests. The pipeline is a person’s attention, so it produces few tests, and some of those are inconclusive. Progress is real and slow, and the problems that were never on anyone’s list stay put.

Automated tools fail differently and more insidiously: they generate plausible findings faster than anyone can evaluate them. Forty suggestions arrive, each defensible, none costed, and the store owner picks the three that sound most interesting. That is not optimisation; it is a longer list of opinions.

The specific pathology to watch for is a detector conditioned on the outcome it is measured against. Fire only on sessions that left, and every finding reports a catastrophic abandonment rate: impressive, and a restatement of the filter rather than a fact about your store. Ask any vendor whether their detection can see conversion outcome.

Why estimating cost matters more than detection

Finding friction is the easy half and has been for years. Dead clicks, abandoned form fields, rage taps and slow paints are all straightforward to detect, which is why so many tools surface them and why surfacing them has stopped being worth much.

The expensive question is which of them cost you money. That needs comparing shoppers who hit the friction with a matched group who did not, controlling for the obvious confounds (traffic source, device, whether the shopper was already price-sensitive), and converting the difference into monthly revenue at your own order value.

Done properly, this usually shortens the list a great deal. Some detected friction turns out to be free: shoppers hit it, shrug and carry on. A tool that will not tell you which findings it discarded, and why, has skipped the step that turns a list into a ranking.

How to evaluate an AI conversion tool

Ask what happens when it is uncertain. Systems with no visible uncertainty are not more accurate; they are less instrumented. A confidence figure that never drops is decoration.

Ask to see the arithmetic behind one dollar figure, in full: the cohort, the comparison group, the order value, the sessions. If the answer is a paragraph rather than a calculation, the number is an assertion.

Ask what share of its findings get tested, and what share of those win. A tool that has run tests on stores like yours knows this; a tool that only advises does not.

Finally, ask what it refuses to do. A product with no stated limits has not been thought about hard enough to have found any.

The arrangement that works

Automate the search and the arithmetic; keep the person on the judgement and the commercial context. In practice a machine says “here is the ranked list of what your store is losing, with the working shown”, and a person decides which items are worth acting on and which are a business decision in disguise.

That also fixes the manual programme’s real bottleneck, which was never idea quality. Analysts spend much of their time on discovery (funnel digging, watching recordings, building spreadsheets) and comparatively little on the judgement they are valuable for. Removing the first is the only way to get more of the second.

Liftable is built for that split. Each opportunity is shown on your own pages with its evidence; once sessions and orders are connected it gets an estimated monthly range with its inputs shown; the fix arrives as a draft you read; and nothing changes on your store until you approve it. It is early: there are no published merchant results yet, and our founding-team programme implements and measures the first fix with you.

Questions