Every algorithm ends by ordering what it found. The Recommendation engine screen holds the account-wide settings underneath every recommendation widget: which item pairs are trusted, how the “Recommended” ordering weighs its signals, what kind of catalog the engine is tuned for, and what the AI reads and checks. They apply to every widget on the site at once; each widget adds its own filters and boosts on top, in the campaign editor.
On personyze.com: AI product and content recommendations →
Video: Trending products and recommendation sets: what shoppers buy together (3:09) — it plays at the top of Recommendation Algorithms — What Each One Picks.
Settings → Recommendation setup → Recommendation engine. Each card says what it is set to; Change opens its control, and every control saves the moment you change it — there is no Save button. Changes to the blend reach live widgets within a couple of minutes, with nothing to republish.
| Card | What it decides |
|---|---|
| Which recommendations are trusted | How many users must back a computed pair before it is used. |
| The “Recommended” ordering | The Score blend — a preset or a recommendation profile — and the four weights under it. |
| Similar items (AI) | The nightly AI similarity pairs, for products and articles alike. |
| Read products (AI) | Whether the AI reads each item, and at which level. Its name follows the profile. |
| Product matches (AI) | How strict the “goes with” lists are, and your own instructions for the AI that checks them. |
The safety floor: minimum users behind a recommendation
The first card, Which recommendations are trusted, decides what the engine is allowed to use at all rather than how it is ordered. Minimum users behind a recommendation is how many distinct users (or transactions) must back an item-to-item pair before the engine will serve it.
It is a straight trade: higher is safer and gives fewer recommendations. One person who viewed two unrelated products is not evidence that they go together; twenty people are. Raise it when a widget is showing pairings that look like coincidence, and lower it when a smaller catalog or a quieter site leaves widgets falling back too often.
It applies to the computed item-to-item pairs — bought-together, viewed-together and their relatives. It has nothing to say about popularity, the catalog itself, or pairs you uploaded by hand as cross-sell or up-sell sets.
The “Recommended” ordering: Score blend
Every algorithm ends by ordering what it found, and the Score blend dropdown sets how that ordering weighs its signals, account-wide. It holds two kinds of choice:
- Presets — editorial starting points that fill the four weights below in one pick. See Presets.
- Recommendation profiles — the kind of catalog the engine is tuned for. See Recommendation profiles.
While the weights match no preset or profile, the dropdown says Custom.
Recommendation profiles
A profile tells the engine what kind of catalog it is ranking. It fills the four weights with its own, decides what the AI reads in each item, and names the lists after what they are on your kind of site — the same card is Complete the look in a fashion store, Accessories that fit in an electronics store and Goes well with in Beauty & grocery and Other stores (every name by store).
For stores: Fashion, Electronics, Home & furniture, Beauty & grocery, Sports & outdoor and Other stores. For content sites: News & publishing, B2B blog, Learning & docs, Recipes & how-to and Video & entertainment.
- Detected until you pick. Personyze looks at your catalog and shows the profile it found, such as Fashion (detected). Nothing is saved until somebody picks.
- Picking a profile saves it with its weights. Change a weight afterwards and the dropdown says Fashion · custom weights.
- Picking a preset while a profile shows means “this is not that kind of site”: the account goes on General, with the preset’s weights, and General reads nothing.
- No profile fits? A store that fits none, and whose shoppers have not bought enough together yet, is read as nothing, and an AI reading card says so. Once they have, the engine reads it as Other stores by itself; to read it now, pick Other stores.
The four weights
| Weight | What it does | Default |
|---|---|---|
| Purchases vs views | How much a purchase counts next to a view when weighing the evidence behind a recommendation. Above ×1, what people buy together outweighs what they view together; Off ranks by views alone. | ×1 |
| Freshness | How strongly new arrivals (added in the last 30 days) and fresh price drops (the last 14) are nudged up the ordering. Off removes the nudges. | ×1 |
| Click & purchase feedback | How much what visitors actually clicked and bought from the widgets feeds back into the ranking. The engine scales each recommendation by its measured click-through, by up to ±50%; this amplifies or mutes that. | ×1 |
| Visitor’s price range | Per visitor, not global — see below. | Off |
Each is a dropdown: Off, ×0.5 — reduced, ×1 — standard, ×1.5, ×2 and ×3 — strong (the price range stops at ×2 — strong).
placeholder. It is where to start, and where to go back to if a change makes things worse.
Presets, by kind of site
A travel site is not a shop, and a publisher is neither. The presets fill the weights with editorial starting points in one pick, and you can tweak from there:
Balanced · Content & publishing · Travel & booking · B2B · Services
The preset for the kind of site set in your Site profile is marked (suggested).
The visitor’s price range
The fourth weight is different in kind from the other three: it is per visitor. It nudges up items priced between roughly half and double that person’s own average — taken from what they have viewed, carted or bought.
- It is off by default, and no preset switches it on. Picking the Fashion, Electronics or Home & furniture profile sets it to ×1.
- Visitors with no priced history are unaffected — there is no average to work from, so nothing is nudged.
Similar items (AI)
A card of its own, with On or Off under its title and Configure on its title row. It is the nightly pass that reads your catalog’s titles, brands, descriptions and categories and computes each item’s nearest neighbours by meaning, so a brand-new item can be recommended before anyone has viewed it.
One switch covers the whole account, products and content alike, and it is the same setting as Similar items (AI) in each catalog’s ⋯ menu — two doors into one switch. Switched on, the first pass starts within about ten minutes; a large catalog can take an hour or two to fill in. With both a product and an article catalog, the same pass also pairs articles with the products they are about, for Content about this product and Products in this content. What each algorithm does with it is under Similar Items (AI) and the about cards.
AI reading
Where the profile reads your catalog, a reading card follows Similar items (AI). Its name says what is read: Read product pictures (AI) (fashion, home), Read products (AI) — the title and description — or Read articles (AI). The AI reads each item once — what it is, and how it looks or what it is about — and after that only new or changed items. What it read feeds the cards that need it: Complete the look (by its profile’s name), In your style and Time to restock, and for articles Go deeper and For you — see the AI cards.
| Row | What it does |
|---|---|
| Reading | The switch. Under it, what reading this catalog costs in AI actions — the first read and then each night. Switched on here, “Reading starts shortly – usually within an hour.” |
| Reading level | Best — the most detailed reading — or Standard, a lighter reading at a fraction of the cost. Each option shows what it costs this account. |
On by default. A new account’s reading is on before anybody chooses it, and the card says On by default – included with your plan. On a Free plan only your most popular items are read that way. Use my credits too (Read my whole catalogue when items are left out) makes it your own choice, charged to your account’s AI credit like every AI feature. Switched off, it is off until you switch it on again.
A profile that reads nothing shows an AI reading card instead, Off, with the reason and what to pick in Score blend to have your products read.
In the campaign editor’s algorithm picker, a card that needs reading while it is off says Needs product reading (picture or article), with a small Turn on product reading button that switches it on right there, at Standard.
Product matches (AI)
Shown with the reading card. After the rules choose which products go with a product — Complete the look, Accessories that fit, Goes well with — the AI checks each one against your product and leaves out what does not fit: the wrong device, size or kind. Items your shoppers bought together are never removed. On a content site it checks Go deeper the same way. It works while reading is on; with reading off it says Works when product reading is on. Only the account owner or a Manager can change it.
| Setting | What it does |
|---|---|
| How strict should matches be? |
Broad — suggest anything that plausibly goes with the product: more items, some looser. Balanced (recommended) — everything the rules suggest, except what the AI check finds doesn’t fit. Strict — only items shoppers bought together and what the AI check confirms: fewer items, all solid. A product the check has not answered yet is not listed. |
| Instructions for the AI | Your own words on what matters in your store, up to 2,000 characters — for example “Projectors: only ceiling mounts and HDMI cables.” or “Never recommend gift cards.” The AI follows them when it checks what goes with what. They can only remove suggestions, never add products. Saved when you leave the box. |
A new level takes effect at the next nightly update, not at once. Choosing Strict shows one line with your own numbers, such as “Strict removes about 1,200 picks the AI was unsure about. At least 150 products would have no list and show your widget’s fallback instead.” — so give those widgets a fallback. On an account the check has not run on yet, it says so instead of numbers. With Other stores, items no shopper bought together show only when the AI check confirms them, at every level.
To see what the check kept and why, open Analytics → Product analytics → Recommendation sets: each AI pick has a Why? with the reasons and what the AI read in the product.
Per-widget boosts, in context
A single widget can still be tuned on its own, under Fine-tune the algorithm in the recommendation
editor — Boost by profit margin, More in stock first, Boost new items, Reserve a slot for new items. To put what fits each visitor first instead — their brand, size, gender or price range — use Match the visitor.
That card shows the account blend live underneath it (“Account-wide blend now: Purchases ×1 ·
Freshness ×1 · Click & purchase feedback ×1 · Visitor’s price range
Off”), so a per-widget boost is never set blind to the account setting it stacks on.
Related
- Recommendations — Start Here — the whole path, from data to the widget.
- Recommendation algorithms — what each one picks before the blend orders it.
- AI Product Recommendations Setup — the guided setup, which turns product reading on for you.
- Recommendation sets — what the engine learned per product, with Why? on each AI pick.