AI Engines · 9 min read · July 15, 2026
Product data for AI: how your items get cited by ChatGPT and Perplexity
When a customer asks ChatGPT "Which running shoes for overpronation under 120 euros?", it's not your Google ranking that decides, but whether the AI understands your product data. Generative Engine Optimization makes items readable for language models: structured data, clear attributes, real facts. Whoever ignores this disappears from the answer before the click can even arise.
Why product search is tipping over right now
Your customers search differently than just two years ago. Instead of typing 'best coffee machine with grinder' into Google and clicking through ten test reports, they ask ChatGPT or Perplexity directly: 'Which fully automatic coffee machine under 500 euros is quiet enough for the office?' The AI answers with three concrete models, and only whoever appears in this answer is even noticed anymore. The classic ten-blue-links moment falls away, and with it the chance of getting into the game via position four or five after all.
For online shops this is a quiet threat. In Google Analytics you may still see stable numbers, but the question is for how long. Perplexity cites sources visibly, ChatGPT increasingly delivers product mentions via the search mode. If your competitor is permanently named there and you're not, you lose revenue at a point that appears in no classic SEO dashboard. GEO is therefore not a trend to wait out, but the channel where market shares will be redistributed over the coming months.
The good news: the lever lies exactly where you as a shop have homework to do anyway, namely with the product data. Language models reward clarity, completeness and structure. Whoever describes their items cleanly wins twice, with the AI and with the human who buys in the end anyway.
How ChatGPT and Perplexity even read your product
A language model doesn't see your shop like a customer with images, hover effects and discount badges. It reads text and structured data. What sits in a JavaScript carousel, in a graphic or in a PDF data sheet often simply doesn't exist for the AI. If your most important selling point, say 'waterproof to 10 bar', sticks only as an icon without alt text on the product image, it won't appear in any answer. The rule of thumb: what isn't present as clean text won't be cited.
What's decisive is structured markup, specifically Schema.org Product with fields like name, brand, description, offers, price, availability and aggregateRating. This data is the format in which machines have exchanged products for years, and AI crawlers access exactly that. A shop that delivers valid Product markup for every item makes the AI's work easy: price, availability and rating stand there unambiguously, instead of having to be guessed from cloudy running text.
Don't judge this from the gut. Take a real product page, copy the pure text version, for example via your browser's reader view, and ask yourself: are all the facts here that a customer needs for the purchase decision? If you yourself have to guess whether the shoe runs in normal or narrow width, the AI can't know it either.
Your customers' questions are your content plan
People ask AI in whole sentences and with context. Not 'women's winter jacket', but 'Warm winter jacket for cycling at minus 10 degrees that doesn't look like the Michelin man'. Exactly this long query is gold for you. It reveals the occasion, the constraint and the emotional criterion. A shop whose product texts take up such situations ('breathable enough for the bike commute') matches the question far more precisely than a description that only lists material and size.
Collect these questions systematically. Your customer-service tickets, the search-field logs in the shop, the reviews and the return reasons are a treasure trove. If thirty customers ask whether the espresso machine works with ground coffee instead of only pods, then the answer belongs verbatim in the product description and in an FAQ on the page. It is exactly these phrasings the AI picks up, because they're linguistically closest to the customer question.
Think in jobs, not in categories. A customer doesn't buy a 'Bluetooth speaker', he wants 'music in the bathroom without fear of splash water'. Whoever writes product texts around these jobs gets recommended for exactly the situational questions that make up the bulk of purchase advice in the AI era.
Attributes that really count: concrete instead of marketing prose
Language models love hard, comparable facts and struggle with superlatives. 'Best sound in its class' is worthless to the AI, because it's unverifiable. 'Battery life 18 hours, weight 240 grams, IPX7 waterproof', on the other hand, is exactly what lands in a comparison answer. If a customer asks 'Which headphones last a whole workday without charging?', the AI can only name your item if the 18 hours are machine-readable somewhere.
So build out your attributes completely and uniformly. For fashion that means fit, material, care instructions, sustainability certificate and real size charts instead of only S to XL. For electronics the technical specs, compatibility and scope of delivery. Gaps are expensive: if one of twenty similar products is missing the compatibility detail, exactly that one drops out of the 'Fits iPhone 15?' answers, even though it technically fits.
An often underrated field is the comparison to the competition within your own assortment. A short, honest paragraph like 'Model A is lighter, Model B has the stronger battery' helps the AI place your product in the right recommendation situation, and positions you as a source that thinks along instead of only selling.
Reviews and real usage data as a trust signal
AI systems weight signals that point to real experience. Authentic customer reviews are ideal for this, because they contain language no marketing team writes: 'runs half a size small', 'slightly stretched out after three washes', 'perfect for wide feet'. Exactly these phrasings answer later customer questions and make your product page a dependable source. A shop that integrates reviews in a structured way and marks them up gives the AI these proofs directly in hand.
What matters is honesty instead of gloss. Walls of five stars without a single critical note seem incredible to humans as well as models. An honest four-star balance with a comprehensible weakness ('build quality top, but manual thin') creates more trust and at the same time gives the AI context for which customer the product fits and for which not.
Complement this with concrete usage scenarios from real life. A short practical section 'Tested over two weeks as a commuter backpack' or 'often bought by customers for the home office' gives evidentiary weight that pure manufacturer specs never reach.
The most common mistake: thin, interchangeable product texts
Many shops adopt the manufacturer description one to one. The problem: the same sentences then appear at a hundred other retailers. For the AI there's no reason to cite you of all people if your text is identical to the competition's. Duplicate content was already harmful in classic SEO, in generative search it's even more dangerous, because the model simply prefers the source with the most own contribution and context.
The way out is not more text at any price, but your own added value. Complement the manufacturer data with your perspective: for whom the product is worthwhile, for whom not, what it can be combined with, which typical mistakes happen when buying. With a tent, for example, 'great for festivals, too little storm-stable for alpine tours'. Such classifications no manufacturer copy can deliver, and it is exactly they that make you citable.
Pay attention to substance instead of keyword stuffing. Sentences that cram 'cheap women's winter jacket buy' three times do harm. Language models recognize unnatural patterns and downgrade the source. Write for the human who asks the question, then it's also right for the machine.
Landing in the answer: feeds, FAQ and currency
Beyond your product pages, it's worth sharpening up in three places. First, a clean, current product data feed, because many AI systems access merchant feeds and marketplace data. Outdated prices or wrong availability there lead the AI to falsely dismiss you as 'not available'. Second, structured FAQ blocks per product that answer real customer questions verbatim and are marked up with FAQPage markup.
Third, currency counts more than before. A guide 'The best gas grills 2024' won't be recommended in 2026 if the models are off the market. Maintain the year figures, swap out discontinued items and keep test-winner references fresh. AI systems prefer visibly maintained sources, because outdated recommendations are their own trust problem. A 'last updated' date is therefore more than cosmetics.
Measure success concretely. Regularly ask ChatGPT, Perplexity and Gemini the ten most important purchase questions of your industry and note whether and how you're named. This manual spot check replaces no tool, but shows you faster than any dashboard whether your product data reaches the AI.
Concrete roadmap for the coming weeks
Start small and measurable. Take your twenty highest-revenue products and work them through completely: valid Product markup, complete attributes, a dedicated added-value paragraph, three real FAQs and integrated reviews. Experience shows these twenty pages make up the bulk of your revenue and are the fastest way to appear in AI answers, without touching the whole catalog at once.
After that you systematize. Define an attribute template per category so no field gets forgotten anymore, and build the maintenance into your product-creation process. New items shouldn't go online without complete data in the first place. In parallel you set up a fixed monthly check routine with the AI systems to see progress and setbacks.
Stay honest with yourself: GEO doesn't replace a good assortment and a fair price. It ensures a good product gets found at all. The effort pays off twice, because the same clear, honest product data also convinces your human customers and lowers your return rate.
Common questions
Is it enough to just add Schema.org markup, or do I also have to change the texts?
Markup alone isn't enough. It helps the AI read price, availability and rating reliably, but the actual recommendation arises from the content. If your description is thin or identical to the competition's, the cleanest markup does little. You need both: technically clean structured data and content-wise independent, concrete product texts with real attributes and usage scenarios.
My product texts come from the manufacturer. Is that a problem for AI visibility?
Yes, if you adopt them unchanged. The same sentences then appear at many other retailers, and the AI has no reason to cite you of all people. You don't have to rewrite the text completely, but add your own perspective: for whom the product is suitable, what it's combinable with, typical purchase mistakes. This own contribution makes you the preferred source.
How do I even tell whether ChatGPT or Perplexity recommends my shop?
Fastest via a manual spot check. Phrase the ten most important purchase questions of your industry the way real customers would ask them, and put them regularly to ChatGPT, Perplexity and Gemini. Note whether you're named or linked and which competitors appear. Perplexity shows sources directly, which makes the evaluation easier. This replaces no tracking tool, but gives you an honest picture of your current AI visibility.
Read on