How Meta Business Agent catalog training and website training actually work under the hood, the exact data sources it draws from, and the setup mistakes that leave it guessing.
Most setup guides for Meta Business Agent tell you to “connect your website” and “link your catalog” and move on, as if those two steps are self-explanatory. They’re not quite the same kind of connection. One reads pages and indexes what’s on them. The other reads a structured feed of rows and columns. Understanding the difference matters, because it’s usually the reason an agent answers a catalog question well and a policy question badly, or the other way around.
Here’s what actually happens when you point Meta Business Agent at your website and your product catalog, and where each one tends to fall apart in practice.
Meta Business Agent draws on five separate inputs: past conversations, a website URL, a product catalog, business details and FAQs you type in directly, and files you upload. We’ve covered the FAQ and brand-voice side of this in our full setup walkthrough and our breakdown of training Meta Business Agent to match your brand voice. This piece is about the two sources that don’t work the way a person typing in an FAQ answer does: the website and the catalog.
You give the agent a URL, and Meta crawls and indexes the pages at that address to pull out product and service details it can answer from later. That’s a meaningful distinction from “the agent reads your website live.” It’s not opening the page fresh every time a customer asks a question. It’s working from an index built by crawling the pages you pointed it at, the same general mechanism behind a search engine, not a live lookup against your CMS.
Two things follow from that.
First, whatever is publicly reachable on the pages you link is fair game to show up in a customer’s answer. If your site has a staging page, an old pricing sheet still live at some forgotten URL, or a page meant for internal use that never got a noindex tag, pointing the agent at your domain can surface it. Link the pages you actually want a customer reading from, not just your homepage and hope the crawler stays on the paths you meant.
Second, because it’s an index and not a live connection, a change you make to your site today doesn’t guarantee the agent’s answers reflect it today. If you update a return policy or discontinue a service, the safest move is to check the agent’s test answers against that specific page again rather than assume the update propagated automatically. This is the same crawl-based model behind most AI training pipelines, and it’s worth treating your website link the way you’d treat submitting a sitemap: useful, but not instantaneous.
The catalog connection is a different mechanism entirely. There are two paths into it. A small business can build a catalog by hand inside the WhatsApp Business app, adding items one at a time with a name, price, and photo. A business already running ads or managing a larger inventory typically connects a structured feed through Commerce Manager instead, in one of the formats Meta’s commerce platform accepts: CSV, TSV, RSS XML, ATOM XML, or a Google Sheet kept in the right shape.
A feed-based catalog needs specific fields to work: a product ID, the product name, an image link, a description, price and currency, a product URL, and an availability status. That last field only accepts a small set of exact values, in stock, out of stock, preorder, or available for order, not a free-text description of stock levels.
Once the catalog is connected, this is where “catalog training” earns its name over “catalog reading.” When a customer asks something like “do you have this in blue,” the agent isn’t describing your inventory in a sentence. It’s pulling an actual entry from the feed, image, price, and a link to buy, straight into the chat as a product card. That’s the mechanic behind Meta’s product-recommendation feature, a WhatsApp Business catalog AI capability that most plain FAQ bots don’t have, and it’s the reason a clean feed matters more here than almost anywhere else in the setup: a card is either accurate or it’s visibly wrong in the chat, there’s no vague middle ground the way there is with a text answer.
That structured feed is also what separates an AI agent product catalog from a plain product list: the feed isn’t just something the agent reads once, it’s the source it keeps pulling from every time a related question comes up.
The most common catalog mistake is a stale availability field. If your feed still says “in stock” on something you sold out of last week, the agent will recommend it with full confidence, because as far as it can tell, that’s still true. This is a sync problem, not an AI problem, and it’s worth checking how often your feed actually updates rather than assuming it’s near real time by default.
The most common website mistake is the mirror image of a good FAQ answer: pages that are all marketing copy and no specifics. A website page that says a product is “premium quality” and never mentions the actual price, sizing, or return window gives the agent nothing concrete to answer from, so it either goes vague or leans harder on whatever FAQs you did write out by hand. If a page you’re linking doesn’t have the answer to a question a real customer would ask, the agent can’t manufacture one from it.
The third, less obvious one is conflicting information between the two sources. If your website lists a price you haven’t updated in months and your catalog feed has the current one, or the reverse, the agent has no way to know which is authoritative, since both are legitimate inputs it was told to trust. Keeping the two in sync isn’t a Meta Business Agent requirement specifically, it’s just good practice that starts mattering a lot more once an AI is reading both and answering from whichever one it happens to pull from.
Before switching the agent live, use the built-in test flow and ask it the exact kind of catalog and product questions your customers actually send, not the easy ones. Confirm the product cards that come back have working images and the right price attached, and check that at least one out-of-stock item returns an accurate answer rather than a confident recommendation. If you’ve just updated your website with new information, ask a question tied to that specific page and see whether the answer reflects the change before assuming it does. For what to do once the agent hits something it can’t answer from either source, see our guide to how Meta Business Agent handles human handoff and escalation.
Meta has also opened a separate Business Agent Platform for connecting third-party systems, like Shopify or a support tool, more directly than a catalog feed or a crawled website page allows. That’s a distinct layer from the setup covered here, closer to what we described as the enterprise connector tier in our Meta Business Agent roadmap piece, and worth checking if your catalog already lives somewhere with more structure than a WhatsApp-managed feed.
Everything above is specific to how Meta trains its own agent on a catalog and a website, crawled pages, a structured feed, an index that updates on its own schedule rather than yours. That’s a reasonable trade for an agent that answers on its own, but it also means the accuracy of an automated reply is only as good as how recently you refreshed those two sources.
The Chat Quotient runs differently because it isn’t reading from an index at all. Its AI Reply Suggestions generate a draft from your own knowledge base at the moment you ask for one, and its AI Agent works from the business information and goals you give it directly inside WhatsApp Web, without a separate catalog feed or a crawl step to keep synced. For a business that’s just connected its catalog to Meta Business Agent and wants a second layer that doesn’t depend on the same crawl-and-index timing, that’s the practical difference: one less thing to keep updated in two places. Its built-in CRM and visual pipeline then pick up wherever a conversation goes after the product question gets answered, tracking the lead through your pipeline instead of ending at the reply.
Does Meta Business Agent read my website in real time? No. It crawls and indexes the pages at the URL you provide, and answers from that index rather than fetching the live page for each customer question. Recheck your test answers after a significant website update rather than assuming the change is reflected immediately.
What file formats does the Meta Business Agent catalog accept? A structured feed through Commerce Manager accepts CSV, TSV, RSS XML, ATOM XML, or a properly formatted Google Sheet. Smaller businesses can also build a catalog by hand, one item at a time, directly inside the WhatsApp Business app.
Why did the agent recommend a product that’s actually out of stock? The agent trusts your catalog feed’s availability field. If that field still says “in stock” after an item sells out, the agent has no separate way to know otherwise until the feed is updated.
Can I stop Meta Business Agent from using certain pages on my website? The safest approach is not linking pages you don’t want it drawing from in the first place, since anything publicly reachable at the URL you connect is a candidate for the crawl to index.
Does the catalog and website training replace writing FAQs? No. FAQs and business details you type in directly are a separate data source, and they’re what cover the specific policy and process questions a website page or a product feed usually doesn’t spell out. Our setup walkthrough covers how the two work together.
Meta Business Agent catalog training and website training are two different mechanisms wearing the same “connect your data” label. The website connection is a crawl and an index, useful for descriptive and policy content but not a live mirror of your site. The catalog connection is a structured feed, and it’s only as trustworthy as the availability field inside it. Neither one fixes itself automatically when your business changes, which is the part worth building into a routine, not just a one-time setup step.
Install The Chat Quotient and turn that chat into a tracked lead, a faster reply, and — eventually — a closed deal. Free to start, right inside the WhatsApp Web you already use.
