I took the summer off from blogging so that I could focus on AI and building an AI shopping concierge for Shopify. Wow, what a ride! When I came up for air I realized that there are some AI basics that even big stores need and most don’t have. This post is the first in a series dedicated to helping you make your Shopify store, not just visible to AI, but a great resource for AI recommendations. Let’s get started.
Somebody asked ChatGPT for a gift for her sister last week, and something decided whether your store was in the answer. It wasn’t luck, and it probably wasn’t your rankings either. It came down to a handful of small things about how your store presents itself to a machine, most of which you can change yourself this afternoon, from your Shopify admin and one browser tab.
There is no shortage of checklists telling you how to do that. I’ve read a lot of them. The trouble is that a good part of what they say is now wrong, because Shopify quietly rebuilt a chunk of this plumbing for AI agents and the checklists haven’t caught up. One item I see repeated everywhere has been measured more than once and found to be essentially inert. Ahrefs found that 97 percent of llms.txt files got zero requests across 137,000 sites, and Google has said Search ignores it. Another tells you to fix a problem that Shopify already fixed for you, which means you go looking for something that is not there and conclude the whole exercise is nonsense.
So this is the version I would give you if you were sitting across from me. What actually matters, what changed, what you can skip, and where the honest limits of the evidence are.
Two questions, not one
Almost every mistake in this area comes from treating AI search as one problem. It’s two, and they need completely different work.
The first is a matching question. Someone asks for something specific. A bulky Icelandic wool. A weighted eyeglasses holder. A replacement bobbin case for a 1954 Singer. The engine has to work out that what you sell is what she meant, and everything it uses to do that lives on your own website: your titles, your descriptions, your image descriptions, your product data.
The second is a recommendation question. Someone asks who to buy from. Best place to get X. Who ships fastest. Which shop knows what they are talking about. That answer is assembled almost entirely from what other people say about products and stores, in places you do not own and cannot edit.
You need both, and they fail in different ways. Win the first without the second and you are findable but never suggested. Win the second without the first and you get suggested, and then an AI cannot answer a simple question about your shipping policies so you lose the sale.
Most of what follows is the first question, because that’s the part you control. The second gets its own section, and it is slower and more valuable than it looks.
The file most merchants have never opened
Type your store’s address with /robots.txt on the end. That page decides what crawlers are allowed to read, and it’s where the biggest recent change happened.
Shopify rewrote its default. If yours opens with a comment saying that public product, collection, page, blog and policy HTML is crawlable, and it points at an agents.md file and something called a UCP endpoint, you are on the current version. Shopify has published, for every store on that default, a document telling AI shopping agents how to browse your catalog and how to transact without doing anything you would not want them doing. That is a substantial piece of work you did not have to do.
It also removed a block that used to matter enormously. The old default hid your /policies/ pages from crawlers, which meant your shipping policy, your return policy and your terms were invisible to every engine, while “what is their return policy” and “how long does shipping take” remained two of the first things anybody asks an AI about a shop they have never bought from. Most articles still tell you to fix that. On the current default it is already fixed.
Which brings us to the next question: Your policy pages can now be read. Are they worth reading?
I have been looking at small Shopify stores while writing this, and the pattern is consistent enough to be worth saying plainly: the refund policy is usually written with real care, and the shipping policy is frequently an empty page. Not thin. Empty. The page loads, the heading says Shipping policy, and there is nothing underneath it. So the crawler arrives, finds a blank, and a customer who asked a perfectly reasonable question gets no answer from anywhere.
If your robots.txt still carries a line reading Disallow: /policies/, that usually means somebody customized this file at some point and it froze at whatever the default was then. If you are not the person maintaining your website it’s worth asking whoever handles your theme whether anything in that custom file is still needed. Usually nothing is, and deleting it restores the current version, which is better than any patch.
How to do this today (October 2026). Open yourstore.com/robots.txt and read the first ten lines. Then go to Settings in your Shopify admin, then Policies, and read what is actually in each one.
Write your product pages the way people ask
A product title is the first thing an engine reads and the thing it leans on hardest. It earns its place by carrying three things: the brand or line, what the object actually is, and the attributes that separate it from the one beside it.
“Fernwood Chunky Merino Wool Yarn, Moss Green.”
“Sugarhill Farm Maple Syrup, Grade A Dark Robust, 12 oz.”
“Hearthstone Stoneware Mixing Bowl, 3 Quart, Speckled Cream.”
Lead with the part that matters most, because the first seventy characters or so are as far as anyone reads, human or otherwise. That is worth checking on your own store, because a lot of shops want to personalize the title with a fun name, some even add quotation marks, which is charming and spends the most valuable real estate on a phrase no customer will ever type. ‘”Coastal Calm” Linen Throw Pillow’ would be an example of that.
Below the title there are two more places a fact can live, and they are not interchangeable. There may be a list of specifics, and there is usually the description. On most stores both sit in the same field, one after the other, and they look like the same information said twice.
The list is for values. Dimensions, weight, material, wattage, wash temperature, what comes in the box. Anything that is a number or a short specific. A customer scans it. A machine reads it as data.
The description is for benefits and worries. These are the questions where the value alone doesn’t settle it, because what the customer is actually asking the AI is whether she is about to make an expensive mistake. Write those as a plain FAQ, in her words rather than yours.
“Can I put this in the dryer?” Yes. Low heat, and it comes out softer than it went in.
“Will it crack if it freezes?” No. It is built to live outdoors all year.
Answer in the first sentence and explain afterwards. AI engines lift short passages that answer something directly. Copy that circles the point for a paragraph before landing does not get lifted, because there is nothing clean to quote.
Your most-asked questions belong in both places, and that is not repeating yourself. The list serves the customer comparing two products. The FAQ serves the one who has already decided and is looking for a reason not to buy.
Your photographs say nothing unless you describe them
Alt text is the written description sitting behind an image. Google says outright that it uses alt text, along with what it can make out from the picture itself and the rest of the page, to work out what an image shows.
Software can recognize the obvious. Roughly what kind of object it is, roughly what color. It cannot tell what the thing is actually made of, how it’s sized, what separates it from the one next to it, or what you call it. Those live in your words or they do not exist anywhere.
Here is the failure that turns up again and again, and it costs stores with variants the most. Alt text gets written per product rather than per image. So a product with eight photographs carries the same eight words eight times, or the main image gets a good description and the rest are left blank.
Think about what that does to a shop selling anything in twelve finishes. Each finish has its own photograph. Every one of those photographs says the same thing, which is the name of the product line. The finish, which is the entire basis on which the customer is choosing, appears nowhere a machine can read. I have seen stores where the color name was sitting in the image filename, which AI doesn’t read, and absent from the alt text, which AI does read.
Give every image its own description. Same stem, different ending. And add the plain word when your name for something does not give it away, because your customer types “dark red” and not “Cranberry.”
The shots that are not variants still have to say what they are, whether that’s the detail close-up, the one held for scale or the room shot. Keep the identification and spend the rest of the sentence on what that particular picture shows that the others do not. One exception worth knowing: if a photograph is reused across every variant, don’t name a variant in it. Alt text describes the picture, not the product.
How to do this today (October 2026). In your Shopify admin go to Products, then your product, then click an image, then Add alt text. Start with your twenty best sellers rather than the whole catalog. That should take you about an hour.
The data behind the page
Structured data is a block of machine-readable facts your theme publishes alongside the human-readable page. Shopify generates it for you, and it is decent and incomplete in ways worth knowing.
What you get automatically covers name, description, image, brand, category, price, currency, availability and the page address. Brand comes from the Vendor field, which a surprising number of stores leave blank, so that one is worth two minutes of your time.
What you do not get is the barcode, the SKU, review ratings, and item condition. Those are the fields that let an engine cross-check your listing against the manufacturer’s own record and treat it as verified rather than merely asserted. If your products carry barcodes, putting them in is the single highest-value data entry on this list. If they do not, because you make them yourself, that is completely normal and Google’s own instruction is to leave the field blank rather than invent one.
Review ratings deserve an honest note, because the marketing around them oversells. Publishing star ratings as structured data is a Google search win. It is not what gets you recommended by an AI and, even though review apps recommend it, I have not found evidence that it is worth doing.
Shopify also publishes a free app called Knowledge Base that generates a set of facts about your store from settings you have already filled in, shows you what it came up with so you can correct it, and reports which questions AI agents are actually asking about you. That last part is worth the install by itself. Shopify is straight about the limits: it makes AI answers about your store more accurate. It does not make you come up more often.
Where AI answers actually come from
This is the section that changed my mind while writing this, so I will give you the data before the argument.
An analysis of 3,312 AI prompts by cloro, an AI-visibility firm, ran them across six AI engines and tracked the sources behind the answers. YouTube accounted for 19 percent of citations. Reddit, 19 percent. One independent testing site, RTINGS, 16 percent. Brand-owned websites barely registered.
That reorders everything. When somebody asks an AI who to buy from, it reads what independent reviewers and ordinary people say about the product, and then it names a shop that stocks it. What you say about yourself, on your own site, in your own words, is close to irrelevant to that particular question.
Which is uncomfortable, because your own site is the only part you control. It is also the strongest argument I know for doing the slow unglamorous thing.
The first place to start promoting what you sell is where your people gather, and I can’t guess at that, because it is different for a tea importer, a bike shop and somebody selling replacement parts for vintage sewing machines. Here’s how to use AI to show you the way.
Ask an AI the question your customer would ask. “Best place to buy X.” “What should a beginner look for in Y.” Then read the citations. Whatever it names is where the answers about your category are being assembled, and that is your target list, discovered from your own market rather than borrowed from someone else’s.
Then do three things. Get reviews somewhere you don’t own, because a shop with ratings in three places reads as more real than one with ratings only at home. Be genuinely useful where your category gets discussed, which is something a hundred purchased links can’t imitate. And keep your titles, prices and descriptions consistent everywhere they appear, because contradictions make an engine less certain it’s looking at one trustworthy business, and uncertainty is how you get left out of the answer.
It’s slow. It compounds. It’s the highest-value work on this entire page.
How to tell if it is working, and the part it cannot measure
Most store owners have no idea whether AI engines send them anything at all.
The easiest place to look is your own Shopify admin. Open Sessions by referrer under Analytics and filter on referrer name for ChatGPT, then Perplexity, then Gemini, then Copilot. Run the same filters on Sales by referrer to see what those visits were worth. In Google Analytics, there is now an AI Assistant channel under Traffic acquisition, though I’d check the individual sources as well, because Google hasn’t published which referrers it folds into that group.
Two caveats, and they matter more than the numbers. When somebody copies a link out of a chat window, no referrer travels with them, so that visit lands in Direct and looks identical to someone typing your address. Your real AI traffic is higher than what you’re seeing. And clicks from Google’s AI Overviews do not appear as AI traffic at all, because the click still comes off a Google results page. For those you need Search Console’s generative AI report, which shows impressions and no click data whatsoever.
When Ahrefs looked at 300,000 keywords in February 2026, they found that pages ranking number one receive 58 percent fewer clicks when an AI Overview sits above them. Pew Research, measuring actual browsing behavior, found people click a result on 8 percent of searches that show an AI summary against 15 percent when there is not one.
That’s a business problem wearing an SEO costume. Fewer people are arriving. The ones who do arrive are further along, because an AI already answered the easy questions, and they came to decide rather than to browse.
Which is why engagement stopped being a vanity metric somewhere in the last two years. How long people stay, how much they look at, and whether they come back are now among the strongest signals a store gives the algorithms, and they are the only part of this that gets better rather than worse as traffic thins out.
That is the problem I built Stylaquin to solve. It turns a Shopify product grid into something closer to a magazine: a shopper drags a product onto the Stylaquin Bar and it opens in a Look Book, an editorial layout of that product’s photographs sitting right beside the collection page she was browsing, so she never navigates away. Anything she likes goes onto an Idea Board she can rearrange, reconfigure and buy from.
On HorseWorldEU.com, across six months, shoppers using it viewed twice as many products per session, 10.0 against 4.9, and stayed 78 seconds longer. Returning visitors converted at 8.13 percent with it against 3.76 percent without. Fourteen percent of visitors engaged with it, against the one to three percent that wishlist tools typically see, and that same group produced 21 percent of session revenue from 13 percent of the traffic.
Over the same period that store’s organic traffic rose 700 percent, during a stretch when most stores were losing ground, and the owner confirmed nothing else had changed. I want to be careful here, because this is exactly the kind of number that gets over-claimed. Stylaquin did not produce that traffic. Google’s algorithm did. What Stylaquin produced was the engagement that the algorithm had just started rewarding.
Where to start
If you do one thing from this page, open your robots.txt and then read your own shipping policy. It takes four minutes and there is a decent chance you find a blank page where an answer should be.
If you do two, write proper alt text on your twenty best sellers.
Everything else here will get its own article, and I’ll link them as each one publishes.
And if you want to take a quick pre-launch peek at the amazing AI concierge I built this summer, head over to our demo site at https://quinn-demo.netlify.app/ I think you’ll have fun, and be engaged, which is what Google is rewarding these days.
