Shopify AI app or custom integration

Install the app when the job is a standard one - recommendations, search, review summaries - and your data lives in Shopify. Build custom when the decision needs data Shopify does not hold, when you need to control the logic, or when app subscriptions across your stack have quietly passed the cost of owning it. Most merchants should install first and build the exception.

The deciding question is not features. It is whether the thing that makes your merchandising good is expressible inside someone else’s product.

Key takeaways

• If the answer depends only on Shopify data, an app almost certainly does it.
• Apps are priced per store per month, which compounds across a stack and across markets.
• The usual ceiling is not capability, it is that your margin, stock or supplier data lives elsewhere.
• Rebuilding a recommendations widget that an app does for $50 a month is poor capital allocation.
• Check the app’s data-sharing terms before assuming it is the low-risk option.

Who this applies to

You are running a Shopify store somewhere between roughly $1M and $50M in annual revenue and deciding whether to install an app or commission a build for recommendations, search, forecasting, personalisation or fraud scoring.

Three questions that settle it

1. Does the decision need data Shopify does not hold?

The single most useful question.

Apps see what Shopify sees: products, variants, orders, customers, and whatever the theme sends them. That covers a great deal.

It does not cover margin per SKU, supplier lead times, warehouse stock in a system Shopify does not know about, offline sales, or a loyalty tier held elsewhere. If the recommendation you actually want is “surface the high-margin item that is in stock and not on backorder from a supplier who is late”, no app can express it, because two of those facts are not available to it.

Data outside Shopify is the most common legitimate reason to build.

2. Is the logic your differentiator?

If your merchandising rules are the thing that makes your store work - a curation philosophy, a bundling strategy, a pricing model - you want control over them. An app implements its vendor’s model of what good merchandising is, tuned to their average customer.

If your merchandising is ordinary and fine, that average is probably better than what a first custom implementation produces.

3. What is the app stack actually costing?

Individually these are $30-300 a month. Collectively, across six apps and three regional stores, they compound into a real line item, and app pricing typically scales with orders or store count.

Worth doing the arithmetic annually across the whole stack rather than per app at the moment of installing.

What each path gives you

The maintenance row is the one merchants under-weight. Shopify changes - theme architecture, checkout extensibility, API versions - and an app vendor absorbs that. A custom integration is yours to keep working through those changes, indefinitely.

Where apps genuinely hit a ceiling

Cross-system decisions. Covered above. The dominant real case.

Sub-second personalisation at scale. Some apps add render-blocking calls that cost you page speed, which costs conversion. If measured latency is material, owning the path may be worth it.

Unusual catalogue structures. Configurable products, made-to-order, rental, subscription-with-swaps. Apps assume a fairly standard product model.

Measurement you can trust. Apps report their own uplift, and the methodology is rarely inspectable. If you want a holdout group and an honest read on incremental revenue, you generally need to control the experiment. This matters more than it sounds - a lot of reported app uplift is measured without a control.

Data residency or contractual limits. If customer data cannot go to a third-party processor, that rules out most apps regardless of merit.

The hybrid that usually wins

As with support tooling, the binary framing suits vendors more than merchants.

The pattern that works: keep apps for the standard surfaces, build the one decision that depends on data they cannot see. Use an app for review summaries and on-site search; build the merchandising rule that needs margin and stock data, and expose it through Shopify’s own APIs and metafields so the storefront consumes it like any other data.

That gets you live quickly on the easy surfaces and spends engineering only where reach genuinely requires it. It also produces a much better second decision, because after two quarters you know which surface actually moves revenue.

Before you install: check the terms

Apps request scopes. Some request more than they need, and some vendors reserve the right to use aggregated merchant data.

Read what a given app can read, and what it may do with it. This is not paranoia - it is the same due diligence you would apply to any processor touching customer records, and the install flow makes it easy to skip. It is also occasionally the finding that makes the build case, when an otherwise ideal app wants scopes you cannot grant.

What we tell merchants, including against ourselves

We build custom ecommerce AI and our published starting figure covers one component. It is worth being direct that we talk a meaningful share of enquiries out of it.

The test we run first is question one: list the inputs the ideal decision needs, and check which live in Shopify. When they all do, an app almost certainly does the job for a fraction of the cost, sooner, with the vendor absorbing maintenance. Building a recommendations widget that an app does for $50 a month is a technically successful project and a commercially poor one, and it produces a merchant who quietly regrets the spend.

Where we think building is genuinely right: when margin, stock or supplier data has to be in the decision. That is not an edge case - it is common in merchants with real supply-chain complexity - and it is invisible to every app in the store.

The other thing we push on: ask for a holdout. Whichever path you take, insist on measuring against a control group rather than accepting a vendor dashboard or our own claim. Uplift measured without a control is not a measurement, and that applies to work we deliver as much as to an app.

When to do neither

Under roughly $1M in revenue. Improving product photography, page speed and the returns policy will move conversion more than a personalisation layer.

When your catalogue is small. Recommendations need something to recommend. Under a few hundred SKUs, curated collections outperform an algorithm.

When the data is poor. Inconsistent product taxonomy and missing attributes defeat both paths. Fix the catalogue first; it improves search, filtering and ads at the same time.

When you have not measured the baseline. Without current conversion and AOV by segment, you cannot tell whether anything worked.

Frequently asked questions

Can a custom integration coexist with apps?

Yes, and that is usually the right architecture. Write your outputs into metafields or your own endpoint and let the storefront consume them alongside app-provided data.

What does a custom build cost to run?

Hosting, model or compute costs if there is inference, and maintenance through Shopify’s platform changes. Budget an annual figure rather than treating it as one-off.

How long before we know if it worked?

Enough traffic through a holdout to reach significance - typically two to six weeks depending on volume. Agree the metric and the test design before building.

Do apps slow the store down?

Some do, especially those injecting scripts into the theme. Measure before and after with your own tooling rather than trusting the app listing.

Is Shopify’s own AI functionality enough?

For some jobs, increasingly yes, and it is worth checking what the platform now includes before buying either an app or a build. Platform-native features carry no extra subscription and no integration risk.

Next step

The three questions at the top take an afternoon and settle most of these decisions without spending anything. The AI for ecommerce engagement starts there - including the answer that an app does this and you should install it.

If the plan includes generated outbound email, that is a separate problem with its own failure mode - see deliverability for AI-generated outreach.

Related: Build or buy AI customer support · How to evaluate an AI agency proposal · How to measure whether an AI system works · AI for ecommerce