What the assistant can answer for shoppers

The Sales Assistant is more than a keyword box. It reads your catalogue and your knowledge content and can hold a real conversation about them — matching on meaning, narrowing by the things shoppers actually ask about, and always keeping prices, stock and colours honest against your live store. This page is a tour of what shoppers can ask, so you know what it does well and how to feed it (see Feeding your Assistant) to make the most of it.

Everything here works only from your own store's data. The assistant never invents a product, price, colour or fact — if it doesn't have something, it says so.

How it finds products: keyword search, or keyword + meaning

Shoppers rarely type the exact words in your product names, so the assistant never searches with their raw question. It first rewrites the question into search terms and filters — "something warm for a winter hike under $100" becomes terms like jacket, fleece, insulated, winter, hiking plus a price filter — and then searches your catalogue with those. It searches your catalogue and your knowledge content on every question, so "do you ship to Ireland and do you have a rain jacket?" is answered from both at once.

What an embedding is, and what it's for

Keyword search matches words: a shopper's "rain jacket" finds products whose text contains rain or jacket. It's fast, exact and free — but it can't tell that "something to keep me dry on a hike" is asking for the same thing.

An embedding captures what a piece of text means instead of which words it uses. The assistant turns each product description or page into a long list of numbers — a kind of fingerprint of its meaning — and does the same with the shopper's question. Texts about similar things get similar fingerprints, even when they share no words, so the assistant can find "waterproof shell, taped seams" for "something to keep me dry on a hike".

Creating an embedding is done by an AI model, once per piece of text, and that's what your plan's embedding allowance pays for — measured in tokens, roughly three-quarters of a word each. After that, an embedding is only made again when the text changes, and the fingerprints are stored in your store's own search index. Each shopper question is turned into an embedding too, so it can be compared — that's only a few tokens per question, a tiny share of the allowance next to embedding your content.

How it searches your catalogue is your choice (Stores → Configuration → Sales Assistant → Catalogue search → Meaning search for products):

	Keyword search (default)	Keyword + meaning search
How products are matched	By the words in the product's name, description and attributes, using the search terms the assistant derived from the question	The same, plus by meaning — a product can match without sharing any words with the question
Works well for	Product names and SKUs, prices, stock, colours, sizes and every other filterable attribute, and descriptive questions whose wording overlaps your product text	All of that, plus vague or creative requests: "a gift for someone who runs", "something cosy for the sofa"
Embedding allowance used	None for your catalogue	Every product, once — typically 500–1,000 tokens per product (more with long descriptions or many attributes); a 10,000-product catalogue is roughly 5–10 million tokens
Best for	Large catalogues, catalogues with clear product names and attributes, and getting started	Smaller catalogues, and stores where shoppers describe needs rather than products

Your knowledge always uses meaning search, whichever you choose — the documents, pages and crawled sites you give the assistant are usually small, and they're where questions are most loosely worded ("how long do I have to send something back?").

You don't have to work this out yourself. Under the setting, the admin shows your own store's figure — your number of products, the tokens per product measured from a sample of your actual catalogue, the total, and how that compares with your plan's embedding allowance and what's left this month — with a plain verdict: enough now, enough after your allowance renews, or more than your monthly allowance. The same estimate is printed by bin/magento salesassistant:status.

You can switch meaning search on at any time without a reindex: the assistant adds the embeddings for your existing products in the background, a batch at a time, and answers improve as it goes. That background work is deliberately paced so your store doesn't slow down — on a very large catalogue it can take a day or more to finish (see Indexing never slows your store). Before you do, check that your embedding allowance covers your catalogue (see Installing → Your plan's allowances) — if the allowance runs out part-way, the products that already have embeddings get an edge in results over the ones still waiting, so it's better to stay on keyword search than to run half-embedded. Switching back off stops using the product embeddings straight away.

Prices and stock are always checked live against your store when the assistant answers, in both modes.

Narrowing by product attributes

Shoppers can filter the way they would in a shop — and not just by colour and size. The assistant narrows by any attribute you've made filterable in your catalogue, so the same behaviour that works for clothing works for electronics, hardware, books or parts:

• "Do you have the Mithra pant in green?" — checks that product's colours and answers for green specifically.
• "Show me shorts in size M." — narrows to that size.
• "Cotton shirts." / "USB-C cables." / "A matte black finish." — narrows by whatever attribute you sell on (material, connector, finish, style, …).

Attribute matching is not case-sensitive, so "green", "Green" and "GREEN" all work.

It also understands numeric attributes as numbers, so shoppers can give a value or a range — "laptops with at least 16 GB of RAM", "under 2 kg", "over 200 hp" — as long as the attribute carries its unit (see Feeding your Assistant for the one small setup step that enables this).

When something genuinely isn't stocked, it stays helpful instead of dead-ending. Ask for a colour a product doesn't come in and it won't just say "no" — it tells you what the product does come in: "The Mithra Warmup Pant isn't available in purple, but it comes in Gray, Green and Orange." That turns a dead end into a choice.

Asking about a product's options and availability

• "What colours does the Mithra pant come in?" — lists the colours it's offered in.
• "What sizes do you have?" — lists the sizes.
• "Is it in stock in size M?" — the assistant checks live availability for that variant, not a stale flag, and treats a size or colour as available only if it's actually in stock right now.

Price questions

The assistant understands price as a number, not just a word, so shoppers can shop by budget:

• "Anything under €50?" — narrows to products within the budget.
• "What's the cheapest jacket you have?" / "the most expensive?" — returns the right one, ordered by price.

Prices are always the live price from your store at the moment it answers — even if a product's details haven't been re-indexed since a price change, the figure the assistant quotes is current. It states your prices exactly; it won't round or garble them.

Listing, counting, and price ranges

The assistant can answer questions about a whole set of products, not just show a few examples — and it works out the numbers from your catalogue rather than guessing from a sample:

• "Show me all the jackets under €60." — it lists them (cheapest first, up to a sensible on-screen limit) and, if there are more than it shows, tells you the true total: "showing the 24 cheapest of 60."
• "How many do you have in red?" — it answers with the actual count.
• "What's the price range for jackets?" — it gives the lowest and highest price.

Because these numbers come from a real query over your catalogue, they're consistent: "how many under €30" and "show me everything under €30" agree with each other, and with what the shopper sees on the page.

Finding a specific product by its code

When a shopper knows exactly what they want, they can type its code and get that one product — no guessing:

• "Do you have SKU ABC-123?" — a SKU always resolves to that exact product.
• "ISBN 978-3-16-148410-0" / "UPC 012345678905" / "part number XY-9987" — if you store those identifiers on your products (common for books and parts), the assistant looks the item up by code, dashes and spacing don't matter.

Finding parts that fit

If your catalogue lists which vehicles or models a part fits, shoppers can ask for parts for their own vehicle and get only the ones that actually fit:

• "Brake pads for my 2015 Honda Civic." — matches make, model and year together, so it won't offer a part that fits a different year or a different car.
• "A chain for a Trek FX 3." — works without a year for non-vehicle catalogues too.
• "An oil filter for 2015–2018 Civics." — understands a year range, not just a single year.

This depends on you listing compatibility on your products — Feeding your Assistant shows how.

Giving shoppers a direct link

When a shopper is ready to look at an item — "send me the link to the Mithra pant" — the assistant gives that product's own page URL so they can open it and buy in one tap.

Supported languages

The assistant can converse with shoppers in many languages — it understands the question and replies in the same language, always grounded in your store's data. It currently supports:

Arabic, Bengali, Chinese, Danish, Dutch, English, French, German, Hebrew, Hindi, Indonesian, Italian, Japanese, Korean, Polish, Portuguese, Russian, Spanish, Swedish, Thai and Vietnamese.

Your catalogue and knowledge content stay in whatever language you wrote them; the assistant bridges the shopper's language to your content when it answers.

Answering from your pages and documents

Alongside products, the assistant answers "how things work" questions — shipping, returns, sizing guides, plan comparisons — from the CMS pages, uploads and crawled pages you've given it. For these to be reliable, the content has to be fed in the right form; Feeding your Assistant explains which source to use for what, and why structured facts (price tables, size charts, plan comparisons) belong in a short curated upload rather than a scraped page.

When it can't help: honest answers and human handoff

Two things the assistant always does:

• It stays grounded. If your store doesn't have a product or the answer isn't in your content, it says so plainly and offers to help another way — it never makes something up to fill the gap.
• It hands off to a person when needed. When a shopper asks for a human, or reaches the limit of what the assistant can answer, it can pass the conversation to your team — or, if nobody's available, capture the shopper's email so the chat becomes a lead. See Talking to a human.

Getting the best answers

The assistant is only as good as what you feed it. Two habits make the biggest difference:

• Keep your catalogue accurate, and mark the attributes shoppers narrow by. Names, prices, stock and attributes are read straight from your catalogue, so the assistant is as right as your product data — and the more of your attributes you make filterable (not just colour and size), the more ways shoppers can narrow. Feeding your Assistant covers the small setup that unlocks attribute, numeric-range, code and fitment search.
• Curate structured facts. For anything it must state completely and exactly — plan tables, size charts, spec sheets — write one short document and upload it rather than relying on a scraped page. Feeding your Assistant shows how.