Your competitor just got named in a ChatGPT shopping answer and you didn't. If "AI shopping doesn't show my products" is the thought you keep having, take a breath, because this is a data problem, not a talent problem. AI shopping citations get earned through product data, structured markup, and off-site signals you already control. Here are eight steps to fix them, in the order that pays off fastest.
Step 1: Map the shopping prompts your buyers actually type
Start with language, not tooling. Write down the exact questions a shopper asks an AI assistant before buying in your category.
You already have three sources for them. Your sales and support inboxes. Your on-site search logs. Your team's best guesses, typed straight into ChatGPT and Perplexity to see what comes back.
Then label each prompt by intent:
- Category: "best ceramic nonstick pan"
- Use case: "best running shoes for wet pavement marathon training"
- Comparison: "X vs Y"
- Problem: "what pan won't scratch a glass cooktop"
- Recommendation: "is this worth it"
Map every prompt to the one product or category page that should win it. Aim for 50 to 100 prompts. That sounds like a lot. It's really one focused afternoon.
Pro tip: long-tail use-case prompts are where smaller brands win. Big catalogs answer them generically. You can answer them precisely, with fit, material, and real conditions.
You'll know this step is done when every prompt carries an intent label and a destination page. Here's where people go wrong: they reuse their SEO keyword list. Keywords are two words. Prompts are full sentences, and sentences are what these engines answer.
Step 2: Audit where you stand on every engine
Now measure. You cannot fix what you have not counted, and guessing at ecommerce product recommendations AI engines make is how teams waste a quarter.
Run your prompt list through each engine and log three things per prompt:
- Mention rate. How often the answer names your brand at all.
- Citation rate. How often the answer links to one of your pages as a source.
- The top three sources cited. The actual URLs the engine leaned on.
That third column is the gold. It tells you exactly which Reddit thread, YouTube review, or roundup article is standing between you and the answer. You are not guessing at strategy anymore. You are reading a list of targets.
Add one more column while you're in there: your two or three closest competitors, and whether each one appeared. That gives you share of voice per prompt, and share of voice is the number that tends to get budget approved. It also shows you something reassuring. Most categories have a handful of brands winning nearly everything, which means the field behind them is wide open.
Doing this free? Query ten prompts a week across ChatGPT and Perplexity and keep a shared sheet. It works, and it will get old fast. Once the cadence hurts, move to tracking that runs on a schedule. DeepSmith's AEO module handles that side: you define the prompts, it checks them on a schedule and reports mention rate, citation rate, share of voice, and the sources AI cites most, with a competitor leaderboard next to your own numbers. Pro tracks ChatGPT, Grow adds Perplexity, Scale adds Gemini, and Enterprise covers Claude and Google AI Mode too.
You'll know this step is done when you have a sheet with 50-plus prompts, per-engine mention and citation rates, and three named sources per prompt. The common failure is tracking Google AI Overviews only. Most shopping conversations now happen inside ChatGPT and Perplexity, so a Google-only audit tells you almost nothing about AI shopping citations.
One thing worth knowing before you read your results: each engine builds ecommerce product recommendations AI shoppers see from a different mix. ChatGPT's shopping research reads product pages directly and leans on reliable third-party coverage plus accurate price and availability. Perplexity pulls from merchant feeds and surfaces a short list per category query, and it weighs open-web signals like forums and knowledge graph entries heavily. Google's AI Mode goes to Merchant Center feeds first and falls back to Shopping Graph data when no feed entry exists. Same catalog, three different doors.
Step 3: Fix your product feed before you touch anything else
This is the least glamorous step and the highest leverage one. Feed errors are a leading cause of invisibility in Google's AI Mode, and a broken feed quietly cancels out everything you do downstream.
Submit and maintain feeds in three places: Google Merchant Center, Microsoft Merchant Center (the feed source behind ChatGPT shopping surfaces and Copilot), and the Perplexity Merchant Program, which is free to join and takes a CSV or XML file. Keep GTIN, MPN, and brand identical across all three.
Then work this checklist per product:
- Title under 150 characters, no promotional language, no all-caps.
- Description of 500 to 1,000 words covering use cases, limitations, and fit, not feature spam.
- A GTIN assigned to every single variant.
- Multiple high-resolution images: product, in context, lifestyle. No decorative stock.
- Price and currency matching the landing page exactly.
- Availability synced to live inventory.
Google now flags eight attributes for AI Mode inclusion: material, fit, durability, sustainability, gender, age group, size, and pattern. For apparel merchants, gender, age group, size, and pattern are explicitly required. If you sell anything physical, material, fit, durability, and sustainability apply wherever you can back the claim honestly.
Every one of those fields has to match your on-page product copy. Google cross-checks the feed against the page, and a mismatch reads as unreliable.
Common mistake: deleting out-of-stock products from the feed. Flag them unavailable instead. Deletions teach the engines your catalog is unstable, and that reputation is slow to rebuild.
Treat the feed as a live connection, not a monthly CSV upload. Product feed AI visibility falls apart the moment your feed and your inventory disagree, and if you sell into several markets, each one needs its own localized title, description, and attributes in the local language. A feed management platform wired to your inventory API handles this. So does a good developer and a scheduled sync.
Your product feed AI visibility work is done when Merchant Center shows zero errors and zero warnings on the eight flagged attributes, across every shipping market you serve. The usual failure is stopping at title and price, because those are the fields that move paid clicks. The flagged attributes don't move clicks. They move inclusion.
Step 4: Put live structured data on every product page
Structured data is how a crawler reads your page without guessing. Deploy it in JSON-LD, which Google and most AI parsers prefer.
The set that matters for shopping:
- Product with name, image, description, brand, sku, mpn, and gtin.
- Offer nested inside Product, with price, priceCurrency, availability, and priceValidUntil.
- AggregateRating on Product, with ratingValue and reviewCount, backed by real reviews.
- Review markup on individual reviews shown on the page.
- BreadcrumbList across your category hierarchy.
- Organization and Brand at the site root, with logo, contactPoint, and sameAs pointing to your verified profiles.
- FAQPage on buying guides and category pages.
The word "live" is doing heavy lifting here. Schema tied to a static export goes stale within a week, and a crawler that sees last month's price and a sold-out item marked in stock learns not to trust you. Wire your markup to the same inventory source your feed uses.
One honest caveat: schema's weight is contested. Some tests show it doesn't lift AI visibility on its own, and other 2026 analyses treat it as a baseline requirement. Treat it as required hygiene rather than a magic lever. You need it, and it won't carry the whole load by itself.
Validate before you ship. Run key product and category pages through Google's Rich Results Test and the Schema Markup Validator, and check that Product rich results come back eligible. Test one page per template, not one page per product, so a storefront with 4,000 SKUs is still a short morning of work.
You'll know it's done when every template passes, not just your one favorite hero product. The classic failure is set-and-forget: markup goes live, someone celebrates, and nobody notices when the price field drifts two weeks later. Put a monthly validation check on somebody's calendar and that failure never happens to you.
Step 5: Get your reviews out of JavaScript and off your own site
Two problems live in this step, and both are quiet.
The first is technical. Reviews that load inside a JavaScript tab or a widget that needs execution are partially invisible to AI crawlers. Your five hundred glowing reviews may as well not exist. Render them in server-side HTML, on the page, in the source. If your review platform can't do that, ask it to, or replace it.
The second is placement. AI engines lean on third-party review platforms more heavily than on your own testimonials, which makes sense: your site is a party with an interest. Trustpilot, Google Reviews, Amazon, and the niche Q&A sites your category actually uses carry more weight for shopping answers.
So activate your happy buyers there, deliberately. Ask at the moment of delight, not thirty days later.
Pro tip: the ask shapes the answer. "Tell us what you think" produces five stars and no signal. "What size did you order and how does it fit?" or "What did you use this for?" produces the specifics that AI answers quote back to shoppers.
One rule you cannot bend: AggregateRating has to reflect real, verifiable reviews. Synthetic ratings are against Google's review snippet policy, and getting caught costs you the rich result and the trust behind it.
Specificity is what gets quoted. A review that says "runs half a size small, wore them for a wet trail half marathon and had no slipping" answers a shopper's question directly, so an engine can lift it into an answer. "Love these!" cannot be used for anything. Ask better questions and your customers will hand you the evidence.
This step is done when your review content appears in raw page source, and when you're adding detailed third-party reviews every month instead of every launch.
Step 6: Earn the off-site mentions engines actually pull from
Here's the part most ecommerce teams skip, and it's the part that moves the needle most. For product recommendations, AI engines cite third-party sources more than they cite brand websites. Editorial authority carries more predictive weight than the volume of content on your own domain. If you want to get products in AI best lists, some of that work has to happen somewhere other than your own site.
You already know where to start, because Step 2 handed you the list. Take the ten sources cited most often across your tracked prompts and work those, specifically.
What that looks like in practice:
- Communities. Reddit threads, Quora answers, and niche forums where your category gets debated. Show up as a real participant with real detail. One thread with authentic use-case specifics can outweigh fifty claims on your homepage.
- Creators. YouTube and short-form reviewers matter enormously in electronics, beauty, fitness, and home, where buyers want to see the thing work.
- Editorial. Wirecutter-style roundups and the vertical publications your category respects. These are slow to earn and durable once earned.
- Roundups and gift guides. Submit products. Somebody is building next season's list right now.
- Entity presence. Wikipedia and Wikidata entries, plus consistent sameAs links across your profiles. Knowledge graph presence lifts citation rates noticeably in Perplexity, which weighs entity data heavily.
Common mistake: treating this as promotional posting. Engines and communities both punish that. Answer questions you can genuinely answer, including the ones where your product is the wrong fit.
You'll know it's working when you can point to ten or more new third-party mentions per quarter, each tied to a prompt you're tracking. If you can't tie a mention to a prompt, it was PR, not AEO.
Step 7: Rewrite product content to answer real buyer questions
Your product pages were probably written for category keywords. Chat answers need something else: natural language responses to specific questions.
Every product page should answer five things in crawlable HTML:
- What is it, in plain words?
- Who is it for, and who is it not for?
- What problem does it solve?
- How does it compare to named alternatives?
- What should someone check before buying?
Then build the layer above it. Buying guides that cover category, use case, persona, and budget. Comparison pages that name real alternatives instead of dancing around them. FAQ blocks that use the literal phrasing buyers type, marked up with FAQPage schema.
Depth pays here. Longer expert content with a named human author gets cited more often than short anonymous copy, and a visible author bio block is the cheapest credibility upgrade available to an ecommerce site. Add one.
This is also where knowing what to write next stops being guesswork. DeepSmith's Content Intelligence shows which competitor pages are winning the prompts you're losing and drops those gaps into an idea bank, so one losing prompt becomes one scheduled article instead of one more line on a wish list. Its writing pipeline builds heading structure, schema, internal links, and metadata during creation rather than after.
You'll know this step is done when each priority product page answers five real buyer questions in the page source and each tracked prompt has a page pointed at it. Where people go wrong: they write for the category term and stay invisible to the sentence.
Step 8: Measure, iterate, and scale what works
You've built the stack. Now keep it honest with a rhythm you can actually sustain.
- Weekly: re-run tracked prompts, log mention and citation rates, note share-of-voice moves.
- Monthly: review which sources engines are citing, audit feed and schema health, spot new winners.
- Quarterly: expand the prompt set, refresh what's underperforming, and copy winning patterns into adjacent product lines.
Set expectations kindly, including with your boss. Most brands see measurable movement in four to twelve weeks once feeds, schema, and off-site signals line up. Competitive categories take longer. What it takes to get products in AI best lists is consistency across all four layers, not a heroic sprint on one of them.
Keep a simple win/loss log: prompts gained, prompts lost, and the source responsible for each shift. That single document turns AI visibility from a mystery into a workflow.
The common failure is judging the whole program on the first 90 days and quietly abandoning it in month four, right before the compounding starts.
What to do next
You don't have to do all eight steps this week. Pick the one that's most broken.
If you have never measured, start at Step 2. Ten prompts, two engines, one spreadsheet.
If your feed has warnings sitting in Merchant Center, start at Step 3. Those warnings are costing you inclusion right now.
If your data is clean and you're still invisible, it's Step 6. Off-site authority is almost always the missing layer.
Then keep going. Momentum matters more than perfection here, and every layer you add makes the next one work harder.
If you'd rather see your real numbers before you plan anything, start a free DeepSmith trial and watch which prompts you're losing, on which engines, to which competitor pages. It's seven days, no contract, and you'll know where you stand.
