Answering Pre-Sale Questions Automatically, Without Making Things Up
How to answer pre-sale questions automatically: the taxonomy of questions that stall checkout, and how to automate answers honestly without guessing wrong.
A pre-sale question is an abandoned cart that hasn't happened yet. The shopper found the product, they're interested enough to type a question, and one missing fact is standing between them and checkout. Answer it in the next thirty seconds and you probably get the order. Answer it tomorrow morning and you probably don't. That's the whole case for automating this, and it only works if the automated answer is actually correct.
This is a practical rundown of what these questions look like, why the usual fixes (FAQ pages, size charts, canned replies) only get you partway, and how to automate the rest without shipping a bot that makes things up.
The taxonomy
Pull your store's chat and email history and filter to messages sent before an order existed. They sort into a short, repeating list.
Shipping cost and timing. "How much to ship to Alberta?" "Will this get here before the 22nd?" These are answerable facts, but they depend on a destination and a date, so a static "flat rate shipping" banner only covers the average case.
Sizing and fit. "I'm usually a 10 in Everlane, what's that in your sizing?" "Does this run small?" This is the single biggest category for apparel and footwear, and it's the one static content handles worst, because the right answer depends on the shopper's own measurements against a specific product's cut, not a generic chart.
Stock and variants. "When is the black one restocked?" "Do you have this in a 42?" Answerable only if whatever answers it can see real-time inventory, not a snapshot from last week.
Compatibility. "Will this fit my 2021 model?" "Does this pod work with the older base station?" These need to be checked against a spec sheet, not answered from a general impression of the product category.
Which one to buy. "What's actually different between the Pro and the Standard?" A comparison question, answerable from the product data, but it requires connecting two products' specs rather than reciting either one.
Returns and risk. "What if it doesn't fit?" This one is asked before purchase specifically because the shopper is trying to lower the risk of buying, not because they already have an order to return.
Every one of these is a shopper who is, functionally, already sold on the product and stuck on one fact. That's a different problem from someone who's just browsing, and it's why the fix for it looks different from general on-site content or marketing.
Why the standard fixes only get you partway
The default answer to "shoppers have questions" is to write more static content: a bigger FAQ page, a detailed size chart, a shipping calculator. Do this. It's cheap and it removes the easy 60% of the volume. But it structurally can't close the gap, for two reasons.
First, static content answers the average case, and pre-sale questions are disproportionately the exception case. A size chart answers "what are your measurements for a large" but not "I'm 5'4", between sizes, and this fabric has no stretch, should I size up." Baymard Institute's usability testing on apparel sites found that 90% fail to let shoppers properly assess appearance, size, or fit, and separately that 83% of desktop and 87% of mobile apparel sites don't provide sufficient sizing information to make a confident decision. That gap is exactly the space where a shopper types a question instead of clicking "add to cart."
Second, static content requires the shopper to find it. Baymard's research also notes that size guide links get ignored when they aren't placed right next to the size selector, and a chart the shopper has to scroll through to find their specific product often gets abandoned before it's read. A page can have the right answer on it and still lose the sale, because the shopper never got to that answer before they closed the tab.
That leaves a real gap between "we have a policy page" and "the shopper standing on the product page right now got their specific question answered." Closing that gap is what automation is actually for.
Speed is part of the answer, not a separate problem
Even a correct answer loses if it arrives too late. Live chat benchmarking in 2026 puts the industry average first response time at under a minute, and separately finds that 71% of consumers expect a chat reply in under 60 seconds, with satisfaction starting to drop once the wait passes three minutes. A pre-sale question sent by email and answered the next business morning is, for a meaningful share of shoppers, answered after they've already bought elsewhere or lost the thread entirely. This is the part manual support can't fully solve regardless of team size: a Shopify store takes orders at 2 AM, and unless someone is awake to answer at 2 AM, the fastest possible manual response still has a floor measured in hours.
Three ways to automate this, ranked by how honest they are
Canned responses and macros. A human agent picks from a library of pre-written answers. Fast to build, and fine for genuinely uniform questions like "what's your return window." It falls apart on anything that needs a specific input, like a real address for a delivery estimate or a shopper's actual measurements, because a canned reply can't adapt to the specifics without a human editing it live.
Decision-tree chatbots. "Click here if your question is about shipping, then click here if it's international." These scale further than macros and can pull in real data at each branch, like a live inventory count. They're honest in a narrow sense: they only ever say what they were explicitly configured to say. Their failure mode is coverage. The moment a shopper's question doesn't match a branch, in a different order or a slightly different phrasing, the tree has nothing, and "please rephrase" is not an answer.
Language-model agents grounded in your actual store. These can handle "I'm 5'10 and 170 lbs, will the medium fit" in the shopper's own words, by reading it against your size chart, and can compare two products or check a compatibility question the same way. This is also where the accuracy risk concentrates. A model that answers from general knowledge instead of your specific content will produce something fluent and specific-sounding that's wrong, because that is what language models do when they aren't grounded in real data. Gorgias has written about this directly: their examples include a chatbot claiming a jacket is waterproof when it's only water-resistant and inventing a discount code that doesn't exist. A wrong "yes, that's in stock" or "yes, that fits" costs more trust than the hundred right answers that came before it earned.
The honest version of automating this means picking the most capable tier and constraining it hard enough that it can't answer from anywhere except your actual data, rather than just picking the most capable tier and trusting it by default.
What "automating honestly" actually requires
If you're evaluating or building anything in the third category, hold it to these before trusting it with a real shopper:
- It answers from your content, not the model's general knowledge. Ask it something your store's content doesn't cover and see what happens. The honest failure mode is "I don't know, let me get someone," not a confident guess.
- It can see current state, not a stale export. Stock counts and prices change; an agent working from a snapshot will eventually tell a shopper something sold out is available.
- It names the specific product or fact, not a category. "We have a few options that might work" is what a chatbot says when it can't actually check. "The medium runs true to size on this style based on our size chart" is what one says when it can.
- It escalates instead of guessing on anything it can't verify. A wholesale inquiry, a custom request, a question with no source in your content, all belong with a human, with the conversation attached so nobody has to re-ask.
Ernest is built around that constraint for Shopify and WooCommerce stores: it ingests your actual site content and product catalog on connect, searches the live catalog rather than a stale export, and answers sizing, shipping, stock, and compatibility questions grounded in what your store actually says. When a question falls outside what it can verify, it hands off to a human by email with the full conversation attached rather than guessing. It's one honest example of the third category done carefully, not the only way to do this well.
Testing before you trust it
Before any automated answer goes in front of real shoppers, write down your ten hardest actual pre-sale questions, the between-sizes case, the compatibility edge case, the product comparison your team gets asked weekly, and run them against whatever you're evaluating. A tool that gets these wrong is disqualifying no matter how fluent its other answers sound. Then watch the transcripts for the first week live: the questions it couldn't answer are a direct map of what your product content is missing, which is useful information whether or not you keep the tool.
For a broader look at what a dedicated sales-side agent does across a full evaluation, including pricing and the current market, see our breakdown of what an AI sales agent does for a Shopify store and our piece on the specific mechanisms behind AI-driven sales lift. If most of your incoming volume is post-purchase, order status, cancellations, returns, rather than pre-sale, that's a different operational problem; our guide to ecommerce customer service covers what to fix there first.
Pre-sale questions aren't a nuisance to route around. They're the clearest signal you get of exactly where a sale almost fell apart. Automate the answer, but only once it earns the shopper's trust rather than spending it.