AI Product Recommendations for Shopify: Conversational vs. Static Widgets
AI product recommendations for Shopify come in two forms: static widgets and conversational chat. What each does, current app pricing, and when each wins.
AI product recommendations for Shopify split into two genuinely different things wearing the same label. One is a widget: a "customers also bought" or "related products" grid, generated by an algorithm from purchase and browsing data, sitting in a fixed spot on the page. The other is conversational: a shopper describes what they need in a chat, and the agent names a specific product from the live catalog because it fits what was just said. Both get called "AI recommendations." They solve different problems, and most stores that install one assume it also does the other's job.
This post covers what each type actually does, what's built into Shopify for free versus what the paid apps add, current pricing in the category, and how to tell which one your store needs, which is often both.
Two categories, one label
Shopify's own Search & Discovery app is the free baseline. It generates Related and Complementary product recommendations automatically for every product, using purchase-pattern and description-similarity signals, and as of early 2026 runs on an updated sequence-prediction model that also powers Trending and Recently Viewed blocks. It's a real algorithm, not a manual "customers also bought" list you curate by hand, and it costs nothing.
It also has fixed limits merchants run into fast: four recommendation types, no per-visitor personalization exposed to the merchant, no A/B testing between recommendation strategies, and no visibility into which recommended products actually get clicked or bought. You can't edit what it generates beyond adding manual overrides, and it only shows up in the placements Shopify decided on, product pages and a few others, not checkout or post-purchase.
That gap is what paid recommendation apps sell. Rebuy is the category leader for algorithmic merchandising: smart cart bumps, post-purchase upsell offers, and bundle logic layered across product pages, cart, checkout, and post-purchase, with A/B testing and revenue attribution per placement. As of publication, its plans run in tiers by monthly order volume, roughly $99/month up to 1,000 orders, scaling to $249, $499, and $749/month as order volume climbs toward 7,500, plus an Enterprise tier above that, or an à la carte "build your own" option starting at $25/month per module. LimeSpot plays a similar role, with tiers that as of publication scale by order count or revenue starting under $20/month at the low end.
What none of these apps do, including Shopify's own version, is answer a question. A widget can surface products statistically correlated with what's in front of a shopper. It can't parse "I'm 5'10 and 170 lbs, will the medium fit" or "will this work with a 2019 model" and narrow the catalog to the one product that answers that specific constraint. That's a language problem, not a merchandising problem, and it needs a different kind of AI to solve it.
What a conversational recommendation actually does
A chat-based recommendation starts from what the shopper typed, not from what similar shoppers bought. "Which one should I get for a gas grill" filters a catalog down to the one or two products that fit that stated constraint and names them, with a link and a reason. "I'm between a medium and large" gets an answer grounded in the store's actual size chart, not a shrug or a generic sizing tip.
For this to work, the agent needs two things a static widget doesn't: live access to the catalog (current stock, current prices, current variants, not a stale export) and grounding in the store's actual content, so answers come from the real size chart, the real shipping policy, and the real product description instead of the model's general knowledge. An agent answering from general knowledge instead of store content will make things up, fluently and confidently, which is the most common failure mode in this category. Grounding is the entire difference between a useful answer and a plausible-sounding wrong one.
The two approaches aren't really competing for the same shopper. A widget works when someone already knows roughly what they want and a nudge toward a related item helps. Chat works when someone doesn't know what they want yet, and the block is a specific unanswered question standing between them and checkout. A well-run store benefits from both: the widget doing ambient merchandising across every page, and chat catching the shopper who's stuck.
What the numbers say, and how much to trust them
The most repeated figure in this space is that roughly 35% of Amazon's revenue comes from its recommendation engine, a number that traces back to McKinsey research and has been cited across the industry for over a decade. Amazon itself has never published the figure, and some researchers have questioned how precisely it was ever measured. Treat it as a directional claim about how much recommendation-driven revenue can matter at scale, not a number your own store should expect to replicate.
Closer to something you can check: McKinsey's retail research frames a simpler bar, a 2 to 4 percentage point lift in average order value is enough to justify the cost of a recommendation tool. That's a threshold you can actually test against your own AOV before and after, which the bigger headline numbers usually aren't.
The honest version of the pitch, for either kind of recommendation, is narrow: a shopper who gets pointed at the right product, instead of guessing or leaving, converts better than one who doesn't. How much better depends on your catalog, your traffic, and how many of your product pages currently leave a real question unanswered.
Where Ernest fits
Ernest is one AI agent that plays three roles for a Shopify store, and the recommendation angle here is its product-agent and sales-agent sides. It has live access to your product catalog, so when a shopper describes a constraint in chat, sizing, compatibility, which of two similar products fits their situation, it names the actual product and links it, grounded in your real catalog data and whatever your site says about it: descriptions, specs, size charts, FAQs, all ingested automatically when Ernest connects to your store.
That's a different job from what Shopify's Search & Discovery or an app like Rebuy does, and Ernest doesn't replace the widget. It doesn't run cart-page merchandising, bundle logic, or post-purchase upsell offers; those are Rebuy's and LimeSpot's core turf, and a store running Ernest for pre-sale chat can still run Shopify's free recommendations block or a merchandising app alongside it for the algorithmic surface. What Ernest covers is the conversation: the shopper who asks a specific question a static grid can't answer, and either buys because the question got answered or leaves because it didn't.
The same agent handles the two roles most stores pair with product guidance: sales-agent duty for pre-purchase questions like shipping cost and stock, and support-agent duty for order status, cancellations, and returns after the sale, all in the same chat instead of separate tools. Read more on the sales-agent side specifically, including the mechanics of how answered pre-sale questions convert into orders.
Honest limits, since a recommendation you can't check isn't worth much:
- No merchandising or bundling logic. Ernest doesn't build cart bumps, curated bundles, or "frequently bought together" grids. It answers questions in chat; it doesn't run the algorithmic widget layer.
- It's reactive. Ernest answers when a shopper opens the chat rather than proactively popping the widget open based on browsing behavior.
- Channels are the storefront widget and email, with escalation to a human by email. No SMS, Instagram, or WhatsApp.
- No proactive outreach. No cart-abandonment emails or follow-up campaigns tied to the recommendation engine.
Pricing is conversation-based, not order-volume-based like Rebuy's tiers: a free plan with 100 conversations that never expire, a real free plan rather than a trial, then $49/month for 500 conversations, $149 for 2,000, and $299 for 10,000. AI is included at every tier, with no per-seat or per-resolution charges layered on top, so the bill scales with how much the agent actually gets used rather than with order count or site traffic.
Deciding what your store actually needs
Start by reading your own product pages' worth of chat and email history for pre-sale questions, filtered to messages before an order was placed. If most of them are sizing, compatibility, or "which one," that's chat-shaped demand a widget structurally can't meet, no matter how good its algorithm is. If your product pages rarely generate a question and the opportunity is closer to cross-sell and bundling, an algorithmic app is the better first dollar spent, and the free Shopify option is worth exhausting before paying for one.
Before trusting either category's vendor numbers, run your own test. Pick your ten hardest real pre-sale questions, the between-sizes case, the compatibility edge case, the "will this work with my model" question, and ask a candidate tool each one against your actual catalog. A widget can't answer any of them; that's not a flaw, it's outside its job. A conversational tool that gets them wrong is disqualifying regardless of what its pricing page promises. Then track the number that matters more than any lift statistic: how many of those chat conversations turned into an order, checked against your baseline conversion rate before you installed anything.
If you run a Shopify store and the gap is pre-sale questions specifically, our breakdown of what an AI chatbot does for Shopify conversion goes deeper into the three mechanisms and how to measure them against your own numbers before deciding.