GEO for E-commerce: The Four Layers a Product Passes Through to Be Recommended
GEO for e-commerce, explained from measured data. Major AI crawlers do not execute JavaScript, so prices and stock may never reach them even on pages that rank well in Google. Drawing on our study where one retailer was named first in 15 of 15 refrigerator consultations, and an audit of nine major brands, this guide covers the four layers — reach, comprehension, selection and measurement — and where to start.
Key takeaways
- GEO for e-commerce begins not with adding product copy but with checking whether the prices and stock levels already published on your pages can be read by an AI at all
- Major AI crawlers including GPTBot and PerplexityBot do not execute JavaScript. A page ranking well in Google can still arrive at ChatGPT as a product with no price and no availability
- In our audit of nine major Japanese electronics retail brands, structured data was present on the product pages of just one. Adoption is limited even among the largest players, which leaves the gap open
What GEO for e-commerce means
GEO for e-commerce is the work of preparing your product data and information pathways so that when a generative AI such as ChatGPT or Gemini recommends a product, yours appears among the candidates. The discipline is called GEO (Generative Engine Optimization) in English-speaking markets and LLMO in Japan, where that term took hold around mid-2025. The full model is set out in our overview of the discipline.
The difference from traditional SEO is that the target of optimization moves from the screen a person sees to the data an AI reads. A search results page lists ten links, so ranking fifth still earns clicks. By contrast, an AI answer usually names about three candidates. Miss that shortlist and you are functionally invisible.
Most merchants begin this work with content: richer descriptions, more reviews, additional articles. The blockage is usually earlier than that. Polishing a description while the price remains unreadable leaves the model treating your product as one with no price.
What AI crawlers read, and what they miss
GPTBot, OAI-SearchBot, ClaudeBot and PerplexityBot do not execute JavaScript. They never wait for rendering and never return, so dynamically displayed prices and stock stay invisible to them (Gemini is the exception, running on Googlebot's infrastructure)
The behavior of an AI crawler is remarkably simple. It issues an HTTP request, parses the HTML that comes back, and ends the process there. It does not wait for JavaScript to execute, and it does not return later for a second attempt.
This applies to GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, Claude-SearchBot and PerplexityBot alike. In a joint analysis by Vercel and MERJ covering more than 500 million GPTBot fetches, no evidence of JavaScript execution was found. GPTBot downloaded JavaScript files roughly 11.5% of the time without ever running them. Google's Gemini is the exception, since it uses Googlebot's rendering infrastructure and can therefore read the painted page.
What disappears on an e-commerce site follows a predictable pattern: price, stock status, specifications that vary by option, spec tables held inside tabs, lazily loaded review text, loyalty point rates and promotional pricing. All of these are commonly fetched and rendered at request time.
The conclusion is straightforward. A Google ranking does not guarantee that an AI can read you. Googlebot executes JavaScript before indexing, while AI crawlers do not. A page that performs well in search can look thin and priceless to ChatGPT, and this combination is common rather than exotic.
Verifying it requires no special tooling. Disable JavaScript in your browser and open a product page, or view the page source and search for the price as a string. Fetching the page with curl works equally well, and each method takes only a few minutes.
What measurement reveals about fixed recommendations
Make the page readable and you will eventually join the shortlist. That is the assumption, and AI answers turn out to be more stable than it allows for.
We run an ongoing study that puts the same purchase question to five generative AI services repeatedly. Our study of refrigerator purchase consultations collected 75 responses comparing electronics retailers. ChatGPT and Gemini each named K's Denki first in 15 out of 15 runs. Claude placed Yamada Denki first in 11 of 15, while Copilot and Perplexity favored Yodobashi Camera in 12 and 9 runs respectively.
Two findings follow. Answers diverge between AI services, and within a single service the answer barely moves at all. Repeating the same question does not change which companies appear at the top.
The washing machine study reproduced the pattern, with ChatGPT and Gemini on K's Denki, Claude on Yamada Denki, and Perplexity and Copilot on Yodobashi Camera. Changing the product category left the per-service tendency intact.
A merchant currently outside the shortlist therefore does not join it by waiting. Getting there requires changing the material the AI uses to make its judgment.
Machine readability across nine major brands
How far along are real e-commerce sites? In our audit of nine major Japanese electronics retail brands, Product and Offer microdata was present within the reachable product pages of one brand out of nine.
The provisional distribution is equally instructive. Two brands stopped before being read, six before being understood, and one before being chosen, meaning the majority are blocked at the comprehension stage. Their official pages could be retrieved, but the structured data needed to present attributes, price, stock, delivery and returns in machine-readable form was limited.
Public information about AI-referred traffic or product feed integration was also absent at all nine brands. The area is difficult to observe from outside, so this does not prove that nothing exists, but no trace of measurement is visible in public. That the largest players sit here means the opportunity remains open for everyone else.
The four layers a product passes through to be recommended
The conditions depend on one another. Investing in an upper layer while a lower one remains unmet does not produce results.
Price, stock and specifications appear as text in the HTML
Attributes are machine-readable and do not contradict your official information
You appear in the results of each query the question decomposed into
You record your share of recommendations over time
Layer 1 — Reach: price, stock and specifications appear in the HTML
Server-side rendering or static generation places the key information in the initial HTML. Prices that change dynamically remain readable as long as the server fills in the value at request time.
Not every element needs to be there. Set a priority: price, stock, primary specifications and a review summary belong in the initial HTML, while the full review list and related-product carousels can continue to load afterward.
The unit of work is the template. It looks like a large engineering project, but fixing a single product page template applies the change to every product at once, so the actual effort is often smaller than expected.
Layer 2 — Comprehension (1): make product attributes machine-readable
Product, Offer and AggregateRating markup makes the name, price, stock and rating explicit. Identifiers such as GTIN are the key that allows platforms to judge two listings as the same item. We cover the implementation approach in structuring product data for agents.
Mistaking the purpose here leads to poor investment decisions. There is currently no evidence that adding structured data increases AI citations. In a controlled experiment published in June 2026, Ahrefs added JSON-LD to 1,885 pages and compared them against a 4,000-page control group, finding no significant increase. Google likewise states in its guidance for AI features that no special schema.org markup is required.
Structured data continues to drive rich results in search, and it provides the material for verifying your product feed against the page. Treat it as a way of handing over a confirmed price rather than leaving the model to infer one.
Layer 2 — Comprehension (2): make official information and AI answers agree
Once a page is readable, the next question is whether it is conveyed correctly. This holds up considerably worse than most teams expect.
We asked five major AI services about the returns and warranty policies of five electronics retailers. Across 100 question patterns, only 60 were correct when checked against the official information, with 25 incorrect and 15 producing no usable answer. Fifteen of the incorrect answers described return windows or extended warranty terms as more generous than the official policy, which carries real-world risk.
The four items that break most often are shipping fees, return windows, warranty length and delivery times. Check whether the numbers agree across the product page, the FAQ and the terms of service. An inconsistency that a human resolves from context is one an AI does not.
Layer 3 — Selection: appear in the results of each decomposed query
Being readable and understood does not always move the recommendation order. At this point the question is whether you appear in search results at all.
AI does not search the question it was given. It rewrites it into several search queries and runs them in parallel. The AI decides the search terms, and they do not match the words the user typed. "A quiet vacuum for a small apartment" breaks into angles like noise level, weight, night-time use in shared buildings, and price band.
The results for each of those queries hold your pages and your competitors' alike, and AI reads that set as context, summarizes it, and judges from there. Winning a single keyword does not work. You need to appear in the results for each decomposed angle.
In a study of 863,000 keywords and 4 million AI Overview URLs published in March 2026, Ahrefs found only 38% of cited pages ranked in the top 10 for the original query, down from 76% in July 2025. The rest were picked up from the results of a different, decomposed query.
On the product page itself, what decides it is whether you survive being filtered by conditions. If size, compatibility and installation requirements do not exist as attributes, a search carrying those conditions drops you.
Layer 4 — Measurement: record your share of recommendations
Recommendations do not come back identical every time. Unlike a search position, this is not one fixed number, so you read it as a share of recommendations across repeated runs. Prepare the purchase-consultation prompts your customers would plausibly use, run them several times on each AI, and record mention, recommendation order and cited URLs. Tracking your search position for each decomposed angle monthly shows whether the layer-3 work is landing.
Three routes by which information reaches AI
Improving your own site does not complete the picture. Information travels by three routes, and your site is only one of them.
| Route | How it reaches AI | What to prepare | Time to effect |
|---|---|---|---|
| Your own site | Search crawlers fetch it and the search index is referenced | Key information in the initial HTML, structured data, internal consistency | Days to weeks |
| Product feeds | Data is handed directly to Merchant Center or the AI provider | Required attributes, identifiers such as GTIN, update frequency | Weeks, after feed review |
| Third-party surfaces | Comparison articles, review sites and press coverage get cited | Earning placement, review volume and freshness, consistent naming | Months |
The standing of product feeds changed substantially over the past year. Google's guidance for AI features states that no special schema is required, while recommending Merchant Center feeds for product visibility in AI answers. Waving off one and encouraging the other reads as a preference for confirmed product data over inference.
Third-party surfaces are the area you cannot control directly, covering comparison articles, review sites and press coverage. According to Ahrefs Brand Radar, the most frequently cited domains in Japanese include YouTube, note, Wikipedia and PR TIMES. For the decomposed angles where you do not rank, appearing on the third-party pages that do is the only route in. That mechanism is why improving your own site alone leaves ground you cannot reach.
Beyond the recommendation — the path to purchase
Separate from the four layers, retail and e-commerce have a continuation: making the shortlist produces no revenue if the journey breaks afterward.
In our survey of 400 people who purchased an appliance within the previous six months, 71.2% reported some form of anxiety before or during purchase. For installed goods such as air conditioners, washing machines and refrigerators, support-related concerns alone accounted for 50.4%. Whether stock and lead times are legible, whether installation and disposal terms are stated, and whether in-store pickup is available all shape the buying experience while simultaneously supplying material for the AI's answer.
What the Instant Checkout retreat revealed
The largest shift in this space arrived in 2026. OpenAI launched in-chat checkout on 29 September 2025 and discontinued it in March 2026. ChatGPT now surfaces product recommendations and routes shoppers to merchant storefronts and apps.
Two implications follow. The first is that discovery happens in the AI while purchase happens on your own site, at least for the time being. Appearing in the answer and converting on the landing page are both required; neither alone produces revenue.
The second is that feed and data work survives the change. Product data continues to be used as material for recommendations, and the wider agentic commerce architecture keeps expanding.
Priorities for getting started
With limited resources, impact and dependency largely settle the order of work.
Stage one: remove the JavaScript dependency. Output price, stock and primary specifications in the initial HTML. Nothing else takes effect until this is done, which makes it the first priority, and the work centers on the product page template
Stage two: reconcile official information. Align the numbers for returns, warranty, shipping and lead times across the site. This is checking and correcting rather than engineering, so it proceeds without development capacity while also eliminating the risk of misleading guidance
Stage three: feeds and third-party surfaces. Maintain the Merchant Center feed, operate reviews, and pursue placement in comparison media. This is ongoing operation rather than a one-off project
How to redesign a whole store around this is covered in e-commerce SEO in the AI era. With a large catalog, treating every product identically is not realistic. Starting from the revenue leaders, the categories where buyers compare on conditions, and the items whose price and stock move most makes the return on investment easier to read. Criteria for deciding whether to outsource are set out in how to choose a GEO agency.
Frequently asked questions
Marketplace pages follow the marketplace's own structure, so what you can change is limited. AI does read those pages, however, so product titles, descriptions and accumulated reviews still have an effect. If you also operate your own store, making it machine-readable carries separate value.
Whether price and stock are emitted server-side or drawn by JavaScript depends on the theme. Disable JavaScript, open a product page, and observe what disappears. Whatever vanishes can usually be addressed through theme modifications or app settings.
Both are necessary. Structure your own reviews so they are machine-readable, and pursue external review sites separately as third-party evidence. Because the roles differ, this is not a choice between one and the other.
Clearing technical faults can change answers within weeks once pages are re-fetched. Reputation on third-party surfaces accumulates over months. The timeline depends on which stage you are working on.
Eligibility and cost follow the provider's current policy, so check the official documentation at the time you evaluate it. The mechanics have also changed quickly, with the purchase path substantially revised in March 2026.
Summary
GEO for e-commerce is not about adding information. It is about putting the information you already publish into a form that arrives. The first step is to disable JavaScript and open your own product page. If the price disappears, that is your starting line.
The conditions come in order: readable, structured, consistent, purchasable. Skipping a lower condition and investing above it does not produce results. Alongside that, check which of the three routes, your own site, product feeds and third-party surfaces, is currently empty.
Specifications in this field will keep moving. Instant checkout lasted five months, and feed requirements may change again. What does not change is the requirement that price and stock be machine-readable, whichever route ends up carrying the most weight. Starting with the part that holds still is the least wasteful order.
How AI currently handles your products is something you have to check to find out. We offer a free AI visibility assessment for retail and e-commerce businesses. Which AI, which questions, and how your products sit alongside competitors. Once that is clear, the condition where you are blocked becomes visible as well.
