How does the Meta AI shopping assistant pick products?

What did Meta's early shopping test reveal about product picks?

Meta's AI shopping assistant chooses products by pulling candidates from Meta's own product catalog and from a quick live browse of a few retailer websites. It then filters them against what the shopper asked for. That's the clearest picture so far, and it comes from outside testing, not from Meta. The company first tried a 'Shopping research' button in the web version of Meta AI for some US users in early 2026. It returned a carousel of cards, each with an image, a brand, a price, a link to one retailer and a few bullets on why the item made the cut. In September 2026 Meta went further with Muse, a personal agent that can also complete checkout across Shopify stores and a growing list of large retailers.

The most revealing look under the hood came from a small test by an AI-visibility software company: 50 prompts in a single retail category, all run through one Muse account. For a typical prompt, the catalog search returned around 60 products from about 30 merchants. At the same time, Muse's browser visited roughly three websites and found just four products from two merchants. Yet of the six or so products the shopper finally saw, about four usually came from the browser and only two from the catalog. In other words, the catalog supplied nearly all the candidates, but the open web supplied most of the winners. The firm called its sample exploratory, and it sells tools in this market, so treat the numbers as a strong hint rather than a rule.

The checkout design matters too. Before Muse places an order, the shopper reviews it and taps Allow or Deny, so a person still signs off on every purchase. That sounds reassuring for retailers, but look at what it implies. Approval happens at the very end, after the agent has already decided which handful of products are worth showing. If your product never makes the shortlist, there's nothing for the shopper to approve. That's the point for every e-commerce owner reading this. Your store now needs product pages and buying-guide content that an AI system can understand, trust and recommend on its own merits, without a paid boost. That's exactly the kind of content Artiql produces for e-commerce brands, and the rest of this article covers what goes into it.

How does Meta AI choose products for a shopper?

Meta AI chooses products by matching the shopper's request and limits against its pool of candidates, then applying personalization and a few blunt ranking rules. Personalization showed up from the start. In early testing, a request for puffer jackets returned women's styles suited to New York, based on the user's location and a gender Meta inferred from their name. Muse also checks a shopper's saved preferences before it searches. Then come the rules. An instruction file reportedly pulled from Muse in late September tells it to rank well-known merchants first, drop listings priced outside a category's normal range, and check no more than three stores on the open web per request. Meta hasn't commented on the file, and it may have changed since.

Two more rules stand out. When a shopper names a brand, that brand's own store reportedly ranks first, and direct-to-consumer brands are favored by default. For a small brand, that's better news than it first sounds. The 'well-known merchants first' rule looks like a wall built for giant retailers, and the file doesn't define what 'well-known' means. I think it's closer to an opportunity. The brands that win here are the ones people already ask for by name, and that kind of demand comes from search visibility, reviews, press mentions and content that gets cited across the web. Organic marketing has always built that kind of compounding brand signal. Now it feeds a shopping agent as well as a results page.

The price rule deserves a closer look, because it's the easiest one to trip over by accident. If your structured data shows an old price, a sale price without the regular price, or the wrong currency, your listing can look like an outlier and get cut before anyone sees it. Beyond the rules in that file, AI shopping agents generally say they look at the same familiar signals: the shopper's request and budget, price, availability, reviews, delivery and returns, and shopping history. None of that is exotic. It's the same basic upkeep that decides how ChatGPT's shopping filters choose products. Fixing it once usually helps you with every assistant at the same time, which is why I'd start there.

How do you make product pages an AI shopping assistant can read?

You make product pages readable for an AI shopping assistant by stating every important fact plainly, both on the visible page and in the structured data behind it. That means Product markup with an accurate price, currency, availability, brand, GTIN or SKU, plus review ratings if you really have them. It also means a product feed that says the same thing. Consistency is the whole game. If your feed says $89, your schema says $99 and your page shows a crossed-out price, an agent can't tell which one is true and has every reason to move on. If you sell on Shopify, your products have reportedly been shared with Meta's catalog by default since September 2026. Whatever data you have is probably already going there, checked or not.

The browser path needs its own attention, because early testing suggests that's where most final picks come from. Muse browses retailer sites much like a person would, and its traffic can be hard to tell apart from ordinary visits. Your key details should load quickly on the page itself, not sit behind a pop-up, an age gate or a script that only runs after the visitor scrolls. Check your firewall and bot-management settings too. Blocking every unfamiliar crawler can quietly shut you out of AI discovery. On the other hand, some requests that claim to be Meta's indexer come from IP addresses Meta doesn't own. The sensible middle ground is to verify bots by their IP addresses instead of banning or trusting them all at once.

Then there's the copy itself. Shoppers don't just type 'puffer jacket'. They ask for a warm, packable jacket for New York winters under a set budget. A product page with only a material code and three adjectives gives the agent nothing to match that request against, and nothing to put in its short 'why we picked it' bullets. Write descriptions that cover who the product is for, when to use it, sizing and fit, materials, care, and what it isn't good for. Add several clear images taken from different angles, since visual detail makes listings richer everywhere. The same feed work that gets you two product images in Bing's mobile results carries over to AI assistants too.

Why do buying guides help get products recommended by AI?

Buying guides help get products recommended by AI because they answer the questions the assistant is trying to settle for the shopper. They also give it reasons to pick one item over another. When Muse's browser visits three stores, it isn't only collecting prices. It's looking for evidence that a product fits the request: which jacket is warmer, which serum suits sensitive skin, which running shoe works for wide feet. A store that has already written that comparison, in plain language, on a page crawlers can reach, makes the agent's job easy. That's the core of ecommerce GEO: content structured so that AI systems can pull out a clear, accurate answer and attach your product to it.

In practice, the most useful content tends to look like this. Category guides that explain how to choose, with honest trade-offs. Head-to-head pages that compare your own models, or set your product against the usual alternative. Sizing and fit guides. Use-case pages, such as the best bag for carry-on-only travel or a coffee grinder that fits a small kitchen. Short FAQ blocks on product pages that answer real questions from your support inbox. Each piece should open with a direct answer and then back it up with specifics. Before you write anything new, run a quick GEO score check on one product or category page to see how readable it already is for AI engines.

Reviews finish the job, both on your own site and elsewhere. On-site reviews with proper markup feed the signals agents say they use, and detailed reviews often answer fit and quality questions better than your own copy can. Mentions on other sites matter more than many store owners expect. The assistant has to judge 'well-known' somehow, and the most likely evidence is what the rest of the web says about you. Here's my contrarian take: a product feed makes you a candidate, but content makes you a pick. Plenty of stores will connect their catalog and stop there, adding themselves to the pile of sixty candidates. Stores that also publish helpful content other sites cite have a much better chance of landing among the final six.

What's still unknown, and should you pay for AI shopping placement?

Plenty is still unknown, and anyone who claims to know Meta's full ranking formula is guessing. Meta hasn't published its methodology. It hasn't said whether it earns referral commissions on recommendations, or whether brands that advertise on Facebook and Instagram get any priority. The instruction file that surfaced in September doesn't define 'well-known', and Meta hasn't confirmed it's still current. The independent test covered one category and fifty prompts. Even usage is at an early stage: one survey in September found that only 2% of US digital buyers started their product searches with an AI assistant. So the honest position is this: we know where things are heading, plus a few rules, but not how much each factor counts.

That uncertainty is exactly why paid placement looks tempting, so think it through before you commit budget. Meta runs a huge ad business, and in October 2025 it said it would start personalizing ads based on people's conversations with Meta AI, shopping chats included. Paid units inside AI answers are spreading across the industry, as early tests of single-brand shopping ads in ChatGPT show. But paid visibility in an assistant works like any other PPC: it stops the day the budget stops. Product data, descriptive pages and buying guides keep earning shortlist spots across Meta, ChatGPT, Gemini and Google long after they go live. If you're weighing the two, this cost-per-lead comparison of SEO and Google Ads is a useful way to frame it.

So what should you do this quarter? Check your feed and schema for consistent price, availability and brand details. Make sure product pages load their key facts quickly, and that your bot rules verify crawlers instead of blocking them blindly. Rewrite thin descriptions around real buyer questions, and publish a few buying guides for your best-selling categories. Then measure the results. Watch your referral traffic and learn how to track clicks from AI shopping assistants so you can see what's working. To estimate whether the effort pays off, try the AI content ROI calculator. If you'd rather have that content produced for you, book a short demo with Artiql and we'll scope it for your catalog. Pricing is tailored to your business and set on that call.

Frequently asked questions