GEO for e-commerce: 5 actions that change AI citations

Sistrix measured in April 2026 that 58% of French Google queries now trigger an AI Overview, and the pattern is comparable on US and UK markets. For an e-commerce merchant, this means a growing share of product, comparison and advice searches end on a synthesized answer, often without a click back to the source page. On the ChatGPT side, Semrush counted 150,000 citations in a single month, and Yext tracked 6.8 million AI citations across platforms. The problem: stores optimized only for traditional Google SEO barely appear in these answers anymore. Models cite pages that directly answer, expose their data via schema.org, and allow AI bots to crawl. Here are five concrete actions to turn an e-commerce catalog into a citable source, testable on a sample of pages this week.

The starting point is simple. Ahrefs tested in March 2026 more than 1,885 pages for structured data coverage and the verdict is harsh: most e-commerce product pages expose an incomplete Product schema, with no GTIN, no MPN, no review, no dynamic availability. For an LLM, these missing fields mean one thing: the page is not a trustworthy source to generate a buying answer.

Before any costly rebuild, identify the 20 to 50 pages capturing 80% of traffic and revenue. That is the core on which to apply the five actions below. The rest will follow through your Shopify, BigCommerce, WooCommerce or Magento template.

Action 1: enrich Product schema beyond the bare minimum

The first action is to fill in every Product schema field that Google and AI bots can read. A Product schema reduced to name, image and price is useless to an LLM trying to answer a comparison query.

Critical fields to expose, in order: name, long description of 300 to 600 characters, brand, sku, gtin13 or mpn, multiple images, offers with price, priceCurrency, availability, priceValidUntil, shippingDetails, hasMerchantReturnPolicy, and aggregateRating plus individual review nodes. That last block is what flips the page from a transactional asset to a citable source.

Point 1: expose gtin13 if you sell globally identifiable products, since Perplexity and Google Shopping AI use this field to dedupe and pick their source. Point 2: shippingDetails became required in Google Search Central public docs to appear in merchant rich results in 2026. Point 3: aggregateRating and review must be real structured data, not a JavaScript widget loading after page render, otherwise GPTBot does not see them.

Quick test: take 10 strategic pages and run them through Google's Rich Results Test. Any error or warning is a citability gap. On Shopify, the default theme covers about 40% of the fields; the rest needs a dedicated schema app or custom Liquid. On WooCommerce, Yoast SEO handles the core but not shippingDetails or hasMerchantReturnPolicy: add a PHP snippet in functions.php or a specialized plugin.

Action 2: rewrite product pages answer-first

The second action is to turn each product page's description block into a self-contained citable answer. Today most pages open with a marketing tagline or a vague promise. An LLM looks for a paragraph that directly addresses the intent query.

The winning format is simple: a first paragraph of 50 to 130 words answering Who is this product for, for what use case, with what key difference, at what reference price. This paragraph must sit above any marketing copy or tab. This is the answer-first pattern Princeton, Allen Institute and Georgia Tech identified in their November 2023 GEO paper as the factor most correlated with LLM citation.

Concrete example for a running shoe page: The ASICS Gel-Nimbus 26 is a max-cushion road running shoe for heavier or neutral runners doing long-distance training. Drop is 8 mm, weight is 305 g in size 9 US, suggested retail is $200. Key difference vs. Gel-Nimbus 25: new FF Blast Plus Eco midsole with higher rebound. Available in seven colorways, free US shipping. This paragraph contains the entities, the numbers, the use case, and the version comparison. An LLM can cite it verbatim.

Action 3: structure a product FAQ that answers real buying queries

The third action is to add at the bottom of every strategic page a JSON-LD FAQ with 5 to 7 questions buyers actually type. This FAQ feeds AI answers directly when a user asks a related question.

Questions to include are not What are the benefits or Why choose this product. They are natural phrasings: What size should I pick if I'm between two sizes, Is the product compatible with, What's the difference vs. last year's model, What's the delivery time to Canada, How long is the warranty, How do I care for it. Source these questions from internal search, support chat, and customer reviews.

The FAQ should be a FAQPage JSON-LD with acceptedAnswer in plain text, not nested HTML. Each answer is 40 to 120 words and gives a number, a condition, or a verifiable reference. This is the format GPTBot favors based on the patterns observed across the 500 million fetches measured by Vercel and MERJ in 2025.

If you want to go deeper on structured data pitfalls, our piece on the most frequent GEO errors covers the traps to avoid. And to understand the framing difference between classic ranking and AI citation, the GEO vs SEO article sets the baseline.

Action 4: allow GPTBot and publish a commerce-oriented llms.txt

The fourth action is the cheapest and most overlooked: verify your AI bots can actually crawl the site and publish an llms.txt to guide them. If robots.txt blocks GPTBot, OAI-SearchBot, ClaudeBot or PerplexityBot, none of the optimizations above matter.

Step 1: open /robots.txt and confirm no User-agent: GPTBot Disallow: / line is hiding. Per OpenAI public docs, the user-agents to explicitly allow are GPTBot for training crawl, OAI-SearchBot for ChatGPT Search results, and ChatGPT-User for real-time calls. On the Anthropic side, ClaudeBot is publicly documented. On Google's side, Google-Extended controls usage by Gemini and AI Overviews.

Step 2: publish an llms.txt file at the root of the domain. For an e-commerce site, the useful content points to: product sitemap, cornerstone category pages, buying guides, return policy page, shipping page. Avoid pointing at cart, account or checkout pages. The llms.txt format is markdown, short, hierarchical. About thirty lines is enough for a 5,000 SKU catalog.

To go further on this specific point, the ScoreGeo methodology details the 13 weighted criteria and explains why bot accessibility weighs 12% of the total score. If you suspect a configuration issue, a manual GEO audit on the top 20 strategic pages often surfaces a cache or CDN bug silently blocking AI bots.

Action 5: build off-site authority with comparisons and structured reviews

The fifth action steps out of the site perimeter and tackles off-page authority. LLMs do not only cite the product page; they also cite comparisons, buying guides and review sites that mention the brand. Ahrefs showed on a 75,000-brand sample that the correlation between AI citations and external brand mentions is stronger than the correlation with classic backlinks.

Three concrete levers. Lever one: create 5 to 10 comparison pages by use case, hosted on your domain or an editorial subdomain. Format: Best running shoes for marathon 2026, Best urban e-bike under $2,000. Each page compares your product with 3 to 5 alternatives, with a structured table and numeric criteria. This is the format LLMs prefer on comparison queries.

Lever two: structure customer reviews in Review schema with author, datePublished, reviewRating and full reviewBody. Avoid third-party widgets that inject reviews via JavaScript after load, since GPTBot does not execute them. Trustpilot and Yotpo expose server-side JSON-LD widgets: activate that option.

Lever three: target 3 to 5 press or media mentions per quarter in specialized industry outlets. A mention in a Wirecutter, The Verge or Engadget buying guide for tech, or in a specialized magazine comparison for sports, weighs as much as a dofollow backlink in the AI citation algorithm. The ideal form is a textual brand-name citation tied to a quality criterion.

If you want to prioritize these five actions on your catalog without guessing where to start, our GEO accompaniment for e-commerce starts with a Product schema audit on 50 pages then delivers a sequenced 6-week battle plan. For brands based in France, a Paris GEO audit or a Lyon GEO audit can be run on-site with internal presentation.

Measuring real impact: KPIs to track from week 2

The actions above produce measurable effects over 4 to 8 weeks depending on catalog size and recrawl frequency. Three indicators to set up from day one.

Indicator 1: number of brand-name mentions in ChatGPT and Perplexity across 20 typical queries from your vertical. Test manually each week in incognito. Indicator 2: referrer traffic from chat.openai.com, perplexity.ai and google.com via AI Overview, visible in GA4 reports by filtering source/medium. Indicator 3: share of pages with a valid error-free Product schema, measured via Rich Results Test or a monthly Screaming Frog crawl.

A typical pattern on this kind of optimization: pages with enriched Product schema and structured FAQ start appearing in Perplexity answers in 2 to 3 weeks, and in Google AI Overview in 4 to 8 weeks. ChatGPT Search is slower because its crawl cache refreshes less often, expect 6 to 12 weeks.

Frequently asked questions