How recommendations work

A recommendation widget sits at a placement — a fixed spot on a page. Each widget runs one model, and the model picks the products. This page explains the placements, the models, and the fallbacks.

The placements

These are the 13 placements where recommendation widgets can show. The placement names are the names the app uses on the Models tab and in reports. Each placement has a suggested model. The "Also good here" column lists the other models that suit the placement. See Add widgets to your theme for how each widget reaches your storefront.

How the app picks the model for a widget

Most widgets have a model setting. You set it in these places:

• The Recommendations and Cart recommendations blocks — the Recommendation model setting on the block, in the Shopify theme editor. A new block starts on Automatic (best for this spot). This option shows the suggested model for the block's Placement. With Placement on Automatic — match the page, it shows the suggested model for the page. To always show one model, select that model.
• Cart drawer recommendations — the Recommendation model setting inside the PersonalizerAI app embed, in the same theme editor.
• The checkout block — the Which products to suggest setting under Upsell carousel on the Checkout screen of the PersonalizerAI dashboard. The default is Complete the Look. The Recommendation model setting on the placed block, in the Shopify checkout editor, overrides it.

The post-purchase, thank-you, and order status widgets have no model setting. See the note above. The post-purchase offer has its own settings on the Post-purchase offer tab of the Checkout page: up to three offers in a row, a discount, a price limit, excluded products, a holdout, and rules (Growth plan and up) that choose the offer for orders containing certain products. See Checkout & after purchase.

Two things can take the model choice out of your hands. Auto-pick (Growth plan and up) tests the models at a placement and shifts traffic to the winner. It chooses only for blocks set to Automatic; a block where you chose a model keeps that model. You turn auto-pick on per placement on the Auto-pick tab of Recommendations. See Auto-pick. A running A/B test can also set the model for its test group. See A/B testing.

The Recommendation model setting on the Recommendations block in the theme editor.

What happens on each request

When a shopper opens a page, the widget asks the app for products. The app checks the widget's model, applies your controls, and returns the products.

"Your rules" means your merchandising rules (Growth plan and up) and your "never recommend" tags. Hand-picked curations replace the model's output for their anchor product. See Merchandising rules, Model settings that matter, and Hand-picked & custom models.

Fallbacks: when a model is not ready

Each model needs a minimum amount of data before it can work. The minimums cover three counts:

• Products in your synced catalog.
• Orders in the last 30 days.
• Daily visitors, averaged over 7 days.

Below its minimums, a model returns nothing. When a model is short on data, the widget does not go empty. Each placement has a fallback chain — an ordered list of simpler models. The app serves the first model in the chain that has results. A model that returns too few products gets the same help: the app fills the empty slots from the chain. Turning a model off uses this same chain. See Model settings that matter.

The full minimum-data table for all 16 models is on Model settings that matter.

Edit a fallback chain (Growth plan and up)

Every placement starts with a built-in chain. You can change that chain for each placement. The control is the Fallback Chains tab of the Recommendations screen. An expanded model card on the Models tab also links to it.

The tab shows one card for each placement. Each card holds:

• A badge. Default means the placement uses the built-in chain. Customized means you saved your own chain.
• The chain, as a numbered list. Move a model with the up and down arrows. Remove a model with the × button.
• An Add a model… list. A chain holds 1 to 6 models. It cannot hold the same model twice.
• Fall back when fewer than N products come back. Set N from 1 to 20. This is the shortfall that starts the chain.
• When it falls back. Select one of two behaviors.

1. Open the PersonalizerAI dashboard.
2. Go to Recommendations.
3. Open the Fallback Chains tab.
4. Find the card for the placement.
5. Change the order, the models, the number, or the behavior.
6. Click Save chain.

To undo your changes before you save, click Cancel. To remove a saved chain, click Reset to default. The placement then uses the built-in chain again, and it also follows any later improvement we make to that chain.

Sold-out products in widgets

One store-wide setting decides what a widget does with a sold-out product: Show it with a Sold out badge, Demote it to the end of the widget, or Hide it. The default is Show. The setting works on all plans.

The control is on the Basics tab of the Search screen, because the same setting also controls your search results. See Ranking, curations & facets for the control and the full steps.

New or small stores

A new store starts with widgets that work from the catalog alone: New Arrivals and On Sale. Trending Now joins after about 50 daily visitors. Best Sellers joins after 20 orders in 30 days. Until the smarter models are ready, their placements serve these popularity models through the fallback chains. Nothing on your storefront waits or breaks — the results get smarter as your data grows.

Shopper privacy consent

Some shoppers decline analytics tracking in your cookie banner. The app respects that choice: it stops collecting their events. Recommendation widgets still show for these shoppers, with these differences:

• Trending Now, Best Sellers, New Arrivals, On Sale, and Back in Stock work fully. They read store-wide signals, not personal history.
• Similar Products, Complete the Look, Frequently Bought Together, and Premium Picks work fully. They read the product on the page, not the shopper.
• Complete Your Cart works fully. It reads the live cart, not tracked events.
• Recently Viewed and Up Next degrade. The shopper's session history does not build, so these widgets show their fallback models.
• Smart Mix works fully. It reads the product on the page or the cart, not tracked events.
• Replenishment does not change with consent. It always needs a signed-in customer with past orders.
• Search works fully.

For the full privacy posture, see Privacy & shopper consent.

A recommendation widget at the product-page placement below the add-to-cart button.