Can you automate conversion optimisation end to end?

The six steps, and which of them a machine can hold

Conversion work breaks into six steps: watch sessions, diagnose what went wrong, estimate what it costs, write the change, test it, publish it. Five are mechanical enough to automate well, and the fact that they are mechanical is exactly why nobody does them consistently by hand.

Watching is pure volume, and a machine wins on the first day. Diagnosis is pattern recognition against a known catalogue of friction. Estimating is arithmetic against your own numbers. Writing the change is a constrained edit to a template. Running and reading a test is statistics.

Publishing is different in kind. It is not a harder computation; it is a decision with your storefront as the stake, and it is the step where the cost of being wrong stops being an opportunity cost and becomes a real one.

Why publishing is the step that stays supervised

Every other step is cheap to get wrong. A wrong diagnosis wastes a test. A wrong estimate ranks an opportunity too high. A bad draft gets rejected on reading. A losing test is what tests are for.

A wrong publish is different: it is live to every shopper at once, it is attributable to you rather than to a tool, and the failure that matters most is the one nobody thought to detect. A change can be a statistically clean win and still be wrong: off-brand, legally unwise, or in conflict with a promotion that starts on Friday. No detector catches that, because it is not a fact about shopper behaviour.

So the argument for supervision is not that the machine is bad at conversion. It is that the machine is not the party that carries the consequence.

What earned autonomy means

The useful middle ground is granting autonomy to a class of change rather than to a product. “Ship winning shipping-clarity fixes without asking” is a decision you can reason about: you have seen that class win on your store several times, you know what it touches, and you know what it looks like when it goes wrong.

“Ship winners automatically” as a single global switch is not the same thing with a different label. It folds every future class of change, including ones that do not exist yet, into one consent you gave in a settings panel months earlier.

The practical test: can the tool tell you, per class of fix, how many times that class has been tested on your store and how many times it won? If it cannot, it has no basis on which to earn anything.

What stays human permanently

Some of this work cannot be automated because it is not an optimisation problem. Your offer, your price, your product range, your returns policy and your brand voice are decisions about what business you are running. A tool that starts testing those has misunderstood its remit.

Merchandising sits on the boundary. Which products lead a collection is partly a conversion question and mostly a commercial one: margin, stock, seasonality, a supplier relationship. A machine sees the conversion half and none of the rest.

And anything with a legal or regulatory surface (delivery-time claims, price displays, subscription terms) needs a person who is accountable for the sentence. That is not a limit of the technology. It is what accountability means.

The questions worth putting to a vendor

Ask where the default sits. If “ship automatically” is on out of the box, the product’s risk model was designed around the demo rather than around your store.

Ask what the rollback is. A rollback that means “we edit the theme back” is not a rollback; one that means switching a block off or ending the test is.

Ask whether a change can reach every shopper without a test and without your approval. If the answer is yes under any circumstance, the autonomy ladder is decoration.

And ask what happens when the tool is wrong: how it is detected, how long the exposure lasts, and who finds out first, you or a customer.

Where Liftable draws the line

Liftable is built around this ladder: an autonomy ladder per surface, guardrails, an approval modal, six reject reasons that become standing rules, and a decisions log, all built and covered by tests. The AI in it can read your store and write down a proposal; it cannot start a test or approve its own work. These rules live in code, not in a prompt.

Drafts come from 20 fix templates for you to read and approve. Liftable never writes theme files: a page change can only be shown through its theme app extension, and shipping and discount tests run as Shopify Functions.

Be clear about where this stands. A shipping test has run on our own development store, and no test has run on a merchant’s store. Today the automation is coming rather than live: Liftable finds the opportunity, and our founding-team programme implements and measures the first fix with you.

The autonomy ladder, in order

1. 01

   Propose

   The tool names a problem, shows the evidence it saw it in, and estimates what it costs against your own traffic. Nothing has changed on your store. This rung should always be on.

2. 02

   Draft the change

   The fix is written down as a concrete change you can read, edit or reject before it exists anywhere a shopper could see. A change you cannot read before it runs is not a proposal; it is a deployment.

3. 03

   Test it on a share of traffic

   The change runs for part of your visitors through a route that leaves your theme’s code alone, and the rest see the original. A bad idea costs a few weeks of one surface for part of your traffic, not your storefront.

4. 04

   Ship on approval

   A won test is presented with its evidence and you decide. This is the rung most stores should sit on for a long time, and there is nothing embarrassing about staying here.

5. 05

   Earned autonomy, per class of change

   Once a specific class of fix has won repeatedly on your store, you can let that class go ahead without asking, and only that class. Autonomy is granted to a track record, never to a vendor.

Sources

1. Liftable: what is built today. Liftable’s own account of what the product does, checked against the code on 2 October 2026 and shown on /how-it-works and /compare. Every figure this page gives about Liftable comes from it. No test has run on a merchant’s store and there are no published merchant results yet.

Questions