Answer engine optimization (AEO) is how your store gets cited when a shopper asks ChatGPT, Perplexity, or Gemini “what is the best insulated water bottle under $50” instead of typing it into Google. Those questions are being asked right now, in volume, and the answers name specific stores and products. Either the assistant can find, parse, and trust your pages, or it recommends someone else with a straight face.

The good news, and the mild anticlimax: AEO is not a bag of new tricks. It is SEO with the tolerance for ambiguity removed. A human squints past your inconsistencies. An answer engine does not squint; it either extracts a clean fact from your page or it moves on to a page where the fact is clean. Three things determine which side of that line you land on.

1. Category pages that state their case

When an assistant assembles a recommendation, it leans on pages that define a category and say who each option is for. A collection page that says “Insulated bottles: 12 to 64 oz, keeps drinks cold 24 hours, from $28” is quotable. A bare grid with the title “Bottles” is invisible. This is exactly the split-description pattern from tip 60: the explanatory text lives below the product grid, written in plain declarative sentences, answering the questions a buyer would ask. If your category page could not be read aloud as an answer, it will not be borrowed as one.

2. Comparison content nobody else will write

The queries assistants get are comparative by nature: best X for Y, X vs Z, is X worth it. Pages that honestly compare, including your own products against each other (“which of our three jackets for which winter”), are disproportionately cited because they match the question’s shape. Here is the part that takes nerve: honest comparison content admits what each option is not good for. That admission is precisely what makes an engine, and a human, trust the rest of the page. A comparison where your product wins every row is an ad, and everyone’s parser knows it.

3. Product data that agrees with itself

Consistency is the unglamorous core of AEO. The price in your structured data, your page copy, and your feed must match. Sizes, materials, and availability must be filled in, not “see description.” Shopify emits schema.org Product markup from most themes automatically, which means the floor is already under you; your job is making sure the data flowing into it, titles, variants, metafields, prices, is complete and coherent. An assistant that finds one contradiction on your page discounts everything else it found there. Machines hold grudges at scale.

Then check your work: ask three assistants your own money questions (“best [your category] for [your ideal customer]”) and see who gets named. If it is not you, the transcript usually shows you why, and it is nearly always one of the three gaps above. AEO is the new SEO in the only sense that matters: the stores that treated it as real two years early collected the compounding returns last time, too.

Structured data, the schema.org product markup embedded in your pages, is how LLMs, shopping agents, and search engines actually see your Shopify store. Your theme is for humans. The colors, the typography, the carefully art-directed hero: machines read none of it. They read the machine-readable layer underneath, and for a growing share of your future customers, that layer is your storefront. When someone asks ChatGPT or Perplexity for “the best insulated water bottle under $40”, the systems answering that question shop your data, not your design.

Back in tip 26 we made the argument that agentic shopping is coming. This tip is the practical half: audit what the machines currently see, because most merchants have never once looked, and what is there was decided by theme developers and app defaults, not by anyone who cares about your revenue.

The 20-minute audit

• Run your bestseller’s URL through Google’s Rich Results Test. It shows exactly what structured data your page emits, plus errors and warnings. Most Shopify themes emit basic Product markup out of the box; the question is whether it is complete and correct.
• Check the fields that sell: price, currency, availability, review rating and count, brand, identifiers like GTIN or MPN. “Missing rating” means every review you collected is invisible to this entire audience. “Availability: out of stock” on an in-stock bestseller is a five-alarm bug you would otherwise never notice.
• Test one of each page type, not just one product: a collection page, a variant-heavy product, a sale product (is the sale price the one in the markup?).
• Check your feeds too. Your Google Merchant Center feed is structured data with a different hat on, and agents lean on the same product attributes. Clean titles, real categories, filled attribute fields: the boring plumbing is the marketing now.

Fixing what you find

Gaps usually trace to one of three places: a theme that emits skeleton markup (fix in the theme, or choose a review or SEO app that injects the missing pieces), product data that was never filled in (no markup can emit a GTIN you never entered), or an app conflict double-emitting contradictory markup, which reads as spam to machines. None of these are glamorous fixes. All of them compound: the same clean data feeds Google’s shopping surfaces today and every agent that shops on your customers’ behalf next year. The stores that treat product data as a first-class asset are the ones the machines will be able to recommend.

Alt text on your Shopify product images is the rare job that serves three masters at once: shoppers using screen readers, Google Images, and the AI shopping agents that increasingly read your store instead of looking at it. It is also, on most stores, either empty or garbage. Open any product and check: the alt attribute is blank, or it is the filename (“IMG_4382-final-v2.jpg”), or someone from 2019 stuffed it with “buy cheap best hoodie sale free shipping”.

This used to be a defensible gap, because writing good alt text for 800 images was a week of mind-numbing work. It is not defensible anymore. A vision-capable LLM (ChatGPT, Claude, and Gemini all qualify) writes excellent alt text from the image itself, and it does not get bored on image 400. This is one of the cleanest batch jobs in all of AI-assisted ecommerce: well-defined input, well-defined output, easy to review, hard to get sued over.

What good alt text actually is

Describe what is visible, concisely, as if telling a friend on the phone what the photo shows. The W3C’s image tutorial is the canonical reference, and its rules happen to be exactly what search engines and shopping agents want too: no “image of” preamble, no keyword stuffing, the product named naturally, the details a buyer cares about (color, material, angle, context) actually mentioned. “Sage green linen shirt, back view, showing the box pleat” beats both the empty string and the spam string everywhere that alt text matters.

The prompt

Work product by product: paste or upload a product’s images together with its title, and let the model handle the whole set in one pass so repeated angles get distinct descriptions. Here is the prompt we use. It is ready to paste.

Getting it back into Shopify

For a small catalog, paste results straight into the media alt fields in the admin. For hundreds of products, export your products to CSV, work through them in batches, fill the image alt column, and re-import. Either way, spot-check ten products before you trust the batch: you are reviewing for invented details, not for style. An afternoon of this clears a debt most stores have carried for years, and every reader of your images, human, crawler, or agent, gets the benefit at once.

Rewriting product descriptions with an LLM usually produces the same beige paragraph every other store is publishing: “elevate your everyday,” “crafted with care,” “the perfect blend of style and comfort.” The tool is not the problem. The prompt is. Ask ChatGPT or Claude to “improve this product description” and you have asked it to average you into the internet.

The fix is to make the model earn the rewrite in two steps: first it must prove it understands your voice by studying pages you are proud of, then it rewrites the weak page under constraints that forbid the beige. Voice is not magic; it is patterns: sentence length, how you open, whether you joke, which words you would never use. LLMs are exceptional at extracting patterns when you actually hand them examples instead of adjectives like “friendly but premium.”

The two-step rewrite prompt

Pick three product pages whose copy you would defend in a fight (they do not have to be your own store’s; a brand you admire works too, as long as the register fits yours). Then paste this, filled in, into your LLM of choice:

Rules of engagement

• You are the editor, not the audience. Read the rewrite out loud. If a sentence does not sound like you, kill it. The model drafts; you decide.
• Chase down every [NEEDS FACT]. That marker is the prompt protecting you from confident fabrication, which is the one unforgivable sin in product copy.
• Ship it as a test, not a truth. One product, two weeks, judged on add-to-cart rate (tip 33). If the rewrite wins, run the same prompt across the rest of the catalog with the voice profile it already wrote.

The lazy version of this workflow makes your store sound like everyone. The two-step version makes the machine sound like you, at a speed no human copywriter can match. The difference is fifteen minutes of pasting.

Agentic shopping, where an AI assistant researches and shortlists products on a customer’s behalf, is moving from demo to habit, and it changes a foundational assumption of your store: that the entity doing the shopping has eyes. People already ask ChatGPT, Claude, Perplexity, and Gemini “what should I buy” questions every day, and the assistants answer by reading the web, including your store. The next step, agents that compare, shortlist, and hand a human a filled cart, is being built by every major platform right now.

Here is the uncomfortable part. An agent does not see your beautiful theme. It does not feel the hero video, admire the color story, or scroll your lovingly art-directed lookbook. It reads your data: titles, descriptions, prices, variants, availability, reviews, policies, structured markup. Two decades of visual merchandising instinct, and the newest customer in the room is functionally blind. The stores that win agent-mediated sales will be the ones whose product data answers questions without a human squinting at photos.

Getting structured early

• Get the facts out of the photos. If the material, dimensions, weight, or compatibility only exist in an image or a lifestyle caption, they do not exist to an agent. Put every buying-decision fact in the description or, better, in metafields.
• Check your structured data. Most Shopify themes emit schema.org Product markup with price and availability; verify yours does, and that it is not stuck on placeholder values from a theme customization long ago.
• Write titles that identify, not just seduce. “Midnight Wanderer” is a mood; “Midnight Wanderer Wool Overshirt, Navy” is an answer. Agents match on the second half.
• Keep policies findable in text. Shipping costs, return windows, and warranty terms are exactly what an agent checks before shortlisting you. A policy that lives in a checkout FAQ image might as well not exist.

The good news: none of this is wasted if agentic shopping arrives slower than the keynotes promise. Structured, complete, honest product data is also what Google Shopping ranks on, what your own site search matches on, and what your human customers skim for. Preparing for the customer without eyes makes the store better for the ones with them. Start with your top ten products this month.

Use an LLM as a CRO consultant: screenshot your product page, paste it into ChatGPT, Claude, or Gemini, and ask the model to shop it as a skeptical first-time buyer. Ten minutes, zero dollars, and you get the one thing nobody on your team can give you: a first impression. You have looked at your PDP a thousand times. You literally cannot see it anymore.

To be clear about what this is and is not: it is not an audience, and it will not replace watching real sessions or running real tests. What it is exceptional at is fresh-eyes work: spotting the missing shipping answer, the buried size chart, the benefit your headline forgot to claim, the seventeen competing links above the buy button. The stuff a good consultant flags in the first five minutes of a teardown, before they start charging you for the rest of the day.

The workflow

• Screenshot like a shopper, not a designer. Full-page capture of the live PDP on a phone-width window, because that is how most buyers meet it. Include what loads first as its own shot: the above-the-fold view is a separate exhibit.
• Pick your hardest product, not your best. The bestseller converts despite its page. Run the teardown on the product with traffic but weak add-to-cart, the one your tip 11 funnel flagged.
• Paste the screenshots with the prompt below. Then, and this matters, push back on two of its findings and ask it to defend them. The second round is reliably sharper than the first.

Two habits make this compound. First, run the identical prompt on a competitor’s PDP for the same kind of product; the side-by-side tells you which doubts your whole category leaves open, and answering those is free positioning. Second, treat every finding as a hypothesis, not a verdict. The model saying “the guarantee is invisible” is a reason to test moving the guarantee, and you already know from tip 7 how to run that test without fooling yourself.

Paste your last 200 product reviews into an LLM and ask for the top five objections and the exact words customers use. Ten minutes of work, and you walk away with the market research agencies charge five figures for: a ranked list of why people hesitate, written in the vocabulary that will dissolve the hesitation.

Here is the thing about product copy: you wrote yours from inside the building. You say “premium full-grain leather”; your customers say “doesn’t scratch when my dog jumps on it.” Those are different languages, and shoppers only buy in their own. Your reviews, support tickets, and Q&A are a corpus of thousands of sentences in customer-native language, and until about two years ago, reading all of it was an intern-week. Now it is one paste.

Export reviews from your review app (Loox, Judge.me, and the rest all have CSV export), grab the most recent 200 or so including the 2-star and 3-star ones, which are the most honest documents your business produces, and feed them to ChatGPT, Claude, or Gemini with the prompt below.

What to do with the output

• Answer the top objection on the PDP, above the fold. If fifteen reviews say “worried it would run small,” your size section just wrote itself: “Runs true to size, say 94% of reviewers.”
• Put the stolen language into bullets and headlines. Customer phrasing consistently outperforms brand phrasing because it sounds like the voice in the shopper’s own head.
• Route objections you cannot copy-write away (real product flaws, shipping speed) to the operations to-do list. The analysis does not care whose department the problem belongs to. That is a feature.

One caveat: the model summarizes what you give it. Feed it only five-star reviews and it will cheerfully report that everything is wonderful. The 2-star reviews are where the objections live, so make sure they are in the paste. Rerun the exercise quarterly; objections drift as your product, price, and audience do.