# How AI reads your store

> Where each AI assistant searches and which part of a product page reaches it: we measured ChatGPT, Claude, Gemini and Perplexity, full table included.

URL: https://clione.ai/studies/how-ai-reads-your-store/
Lang: en

When someone asks an AI assistant to look at a product, the assistant opens the [product page](/glossary/product-page/) right then and answers with what it receives. **ChatGPT, Claude and [Perplexity](/glossary/perplexity/) receive the page the shopper sees, not the [structured data](/glossary/structured-data/)**; [Gemini](/glossary/gemini/) receives both, and markup added by a script reaches none of them.

That's the short answer to the question every store working its catalog for [AI search](/glossary/ai-search/) asks: of everything I put on a product page, which part actually reaches each assistant? We didn't take it from each company's documentation or from an opinion. We measured it in October 2026, with a code hidden in every spot on the page, and the whole table is here, what gets through and what doesn't.

## Before it reads: where each assistant searches

For almost any product question, the assistant doesn't start with your store. It first searches a search engine's index, picks a few results and only then opens those pages. It's the rule that always held on Google: **a store that isn't in the index doesn't exist for the assistant**, however well its product page is written. The path has four steps:

1. **Memory.** From its training, the model already knows the brands that are cited everywhere. A small store can't change that.
2. **Search.** It queries a search engine's index. If the product page isn't indexed, it doesn't come up.
3. **Selection.** It picks a few results by [indexing](/glossary/indexing/), authority and relevance.
4. **Reading.** It opens those pages and reads a reduced version of each one. That's the part this study measures.

That's why [ecommerce SEO](/glossary/ecommerce-seo/) is still the base of [GEO](/glossary/geo/) and [AEO](/glossary/aeo/). And since each assistant searches a different index, ranking well on Google says nothing about [Copilot](/glossary/copilot/) or Claude.


| Assistant                                                                                | Index it searches                                                                  | How we know                                                                                                    |
| ---------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------- |
| Gemini, [AI Overviews](/glossary/ai-overviews/) and [AI Mode](/glossary/google-ai-mode/) | Google Search's index                                                              | Google documents it[^google-ai-guide]                                                                          |
| Copilot                                                                                  | Bing's index                                                                       | Microsoft documents it[^microsoft-ai-search]                                                                   |
| Claude                                                                                   | Brave Search                                                                       | An analysis by Profound: 13 of the 15 results Claude cited matched Brave's top results[^profound-claude-brave] |
| ChatGPT                                                                                  | Its search, where [OAI-SearchBot](/glossary/oai-searchbot/) surfaces websites      | OpenAI documents it; a site that blocks OAI-SearchBot doesn't show in its search answers[^openai-bots]         |
| Perplexity                                                                               | Its own results, where [PerplexityBot](/glossary/perplexitybot/) surfaces websites | Perplexity documents it[^perplexity-bots]                                                                      |


Claude's row is the weakest in the table: it comes from a small sample, from March 2025, and Anthropic hasn't confirmed it. Still, it points at a gap few stores check. **A store that's on Google and Bing may not be on Brave.** Google and Bing have dashboards to submit the sitemap and report changes, [Google Search Console](/glossary/google-search-console/) and [Bing Webmaster Tools](/glossary/bing-webmaster-tools/); Bing also accepts [IndexNow](/glossary/indexnow/), up to 10,000 URLs per notice.[^indexnow-docs] Product pages can't be submitted to Google one by one: its Indexing API only takes job postings and livestreams.[^google-indexing-api] Brave has no dashboard, no sitemap submission and no IndexNow. Its index is fed by its own [crawler](/glossary/web-crawler/) and by the Web Discovery Project, through which Brave browser users who opt in share the pages they visit anonymously, and its submission form is for asking it to read a page again.[^merj-brave] A product page nobody visits with Brave can stay out of its index.

## How we measured it


We built test product pages and hid a unique code in each of the 16 places a fact can sit on a page: the visible text, the title, the image's [alt text](/glossary/alt-text/), the [JSON-LD](/glossary/json-ld/) in the HTML, JSON-LD added by JavaScript or Google Tag Manager, [microdata](/glossary/microdata/), the [meta description](/glossary/meta-description/), the noscript block, content drawn by JavaScript, [llms.txt](/glossary/llms-txt/) and a few more. The codes were made up for the test and exist nowhere else on the internet. **If one shows up in an assistant's answer, the only place it could have come from is that part of that page.**


Then we asked each assistant about the product: 210 conversations per assistant and model, in three modes. In one, the assistant decides whether to open the URL; in another, we ask it to open it; in the third, it gets no URL at all. We measured through each company's API, not in their apps: ChatGPT with gpt-6-luna; Claude with Sonnet 5.5 through the API and with Claude Code on a subscription; Gemini with three models, 3.1 Pro, 4 Argon and 3.8 Flash; and Perplexity with its Agent API.


A test like this only counts if it can fail, so it has controls. The visible text is the positive control: if an assistant doesn't bring it back, that conversation didn't read the page. A decoy code we never published didn't appear once, so the assistants aren't making codes up. The runs with no URL scored 0%: none of this was known to them beforehand. And the server logs name the robot behind each visit: [ChatGPT-User](/glossary/chatgpt-user/), Claude-User, GoogleAgent-URLContext and Perplexity's browser. Every percentage carries its margin of error (a 95% Wilson interval), and each arm of the test (an assistant, a model and its tool for reading pages) is counted on its own, never pooled. We measured on October 22, 23 and 29, 2026; we voided Gemini's run of the 23rd because the pages carried the nosnippet directive, and ran it again without it.

## The results


Each cell under the four assistants is the share of conversations in which the code from that spot made it into the assistant's answer. The table covers nine of the sixteen places we measured. The first column isn't our measurement: it's what Google and Bing document they do with each part when they index.


| Where the fact is                    | Google and Bing, when indexing                       | ChatGPT    | Claude                  | Gemini  | Perplexity |
| ------------------------------------ | ---------------------------------------------------- | ---------- | ----------------------- | ------- | ---------- |
| Visible text                         | Used                                                 | 100%       | 88–97%                  | 98–100% | 100%       |
| Title                                | Used                                                 | 84%        | 88–97%                  | 98–100% | 100%       |
| Image alt text                       | Google Images                                        | 100%       | 88–97%                  | 98–100% | 0%         |
| JSON-LD in the HTML the server sends | [Rich results](/glossary/rich-results/) and Shopping | 0%         | 0%                      | 97–100% | 0%         |
| JSON-LD added by JavaScript or GTM   | Google processes it, less reliably for Shopping      | 0%         | 0%                      | 0%      | 0%         |
| Markup in microdata                  | Used                                                 | 0%         | 0%                      | 0%      | 0%         |
| Meta description                     | Used in the search result                            | 0%         | 88% API, 0% Claude Code | 0%      | 0%         |
| Content drawn by JavaScript          | Google renders it                                    | 0%         | 0%                      | 0%      | 100%       |
| llms.txt                             | Google doesn't need it                               | 0 requests | 0                       | 0       | 0          |


**Three places reached no assistant at all: JSON-LD added by a script, microdata and llms.txt.** In Claude's column, the range runs from the API (88%) to Claude Code (97%), and the meta description reached the API (88%) and not Claude Code (0%). Gemini's three models behaved the same.

## What it means for each assistant

### ChatGPT


ChatGPT gets the text of the page as the server sends it: the visible text and the image alt text arrived in 100% of conversations, and the title in 84%. In the test it didn't run JavaScript, so content drawn by a script didn't reach it, and none of the markup did either: not the JSON-LD, not the microdata, not the meta description. The visit comes from ChatGPT-User, the [user agent](/glossary/user-agent/) OpenAI uses when a person asks it to open a page.

### Claude


Claude reads like ChatGPT where it counts: the visible text, the title and the alt text reach it, from 88% through the API to 97% in Claude Code, and the JSON-LD reaches it through neither. The difference is the meta description: **it reached Claude through the API in 88% of conversations and Claude Code in none**. It's the only one of the four the meta description reached, and only through one of its two routes, so it's not a good place to tell an assistant anything. The visit comes from Claude-User.

### Gemini


**Gemini is the only one of the four that receives the JSON-LD**: in 97% to 100% of conversations, as long as the block is in the HTML the server sends. JSON-LD added by JavaScript or Google Tag Manager after the page loads didn't reach it, and neither did microdata. The visible text, the title and the alt text reached it in 98% to 100%. All three models behaved the same, and the visit comes from GoogleAgent-URLContext.

### Perplexity


Perplexity reads with a real browser that runs JavaScript: **it's the only one that received content drawn by a script**, in 100% of conversations. Even so, JSON-LD added by a script didn't reach it, nor did the JSON-LD in the HTML, nor the image alt text. The visible text and the title reached it every time. One caveat: Perplexity's Agent API runs on an OpenAI model under the hood, so this column describes how Perplexity reads the page, not a Perplexity model of its own.

## What it means for your product pages

Assistants read your page like a shopper in a hurry: **if a fact isn't in view, it doesn't exist**. The table leads to six decisions for a product page.

**Write what matters on the page.** The price, the dimensions, the materials, what it's compatible with, what's in the box: if it's in the visible text, it reaches all four assistants. A visible FAQ counts as visible text. Anything that lives only in an image's alt text doesn't reach Perplexity.


**Put the JSON-LD in the HTML the server sends.** It's what Google and Bing read for search: Google documents which properties product markup needs for its [merchant listings](/glossary/merchant-listings/),[^google-merchant-listing] and Bing confirmed that it reads [Schema.org](/glossary/schema-org/) markup in JSON-LD.[^sel-bing-jsonld] Microsoft has also said that Bing uses [schema markup](/glossary/schema-markup/) and, according to Search Engine Roundtable's account of a talk by Fabrice Canel of Microsoft Bing, that the markup helps its language models understand content.[^ser-microsoft-schema] [Copilot](/glossary/copilot/) wasn't part of this test. That's the documented part; what we measured is that Gemini receives it. Google also asks that structured data match the visible text on the page.[^google-ai-features]

**Schema injected by a script reaches no assistant.** If your theme, an app or Google Tag Manager adds the JSON-LD after the page loads, in the test it reached none of the four. Google can read it after [JavaScript rendering](/glossary/javascript-rendering/), but it warns that dynamically generated markup can make Shopping crawls less frequent and less reliable, right where [price and availability](/glossary/price-and-availability/) change fastest.[^google-merchant-listing] To find out where yours lives, open the product page, choose "View page source" in your browser (not the inspector, which shows the page after rendering) and search for application/ld+json. If it isn't there, a script adds it.

**Nobody requested llms.txt.** No assistant asked for the file during the test, and Google says you don't need to create AI text files or special markup to appear in [AI Overviews](/glossary/ai-overviews/) or [AI Mode](/glossary/google-ai-mode/).[^google-ai-features] Its optimization guide names it outright: llms.txt isn't needed for Search's generative AI features.[^sej-llms-txt] Having one does no harm, but it doesn't replace anything above.

**The meta description is still SEO.** It's work for the search result, and that hasn't changed. Among the assistants, it reached only Claude through the API. Whatever you want an assistant to know about a product goes in the visible text, not in the meta description.


**SEO, [GEO](/glossary/geo/) (Generative Engine Optimization) and [AEO](/glossary/aeo/) (Answer Engine Optimization) are the same work.** The same product data, for the search engine and for the assistant, delivered two ways: the markup for Google, Bing and Gemini, and the page text for all four. It's not two projects: a product page that says the same thing in the markup and in the text serves both readers at once.

## How to check your store


The basics can be checked today, with no special tools and with or without Clione. The first two checks are about being found; the last three, about being read.


1. **The store is in all three indexes.** Search `site:yourstore.com` on Google, on Bing and on search.brave.com, and see whether your product pages come up. On Google, Search Console's page indexing report says which ones are indexed and why the rest are missing.
2. **The sitemap and the robots' permissions.** The [XML sitemap](/glossary/xml-sitemap/) has to be up to date and list every product page. The [robots.txt](/glossary/robots-txt/) mustn't block [Googlebot](/glossary/googlebot/), [Bingbot](/glossary/bingbot/) or the assistants' robots, such as OAI-SearchBot, ChatGPT-User or Claude-User.
3. **Key facts in view.** Price, availability, dimensions, materials, compatibility and the answers to the usual questions, written as text on the product page, not only in images, in the code or in tabs that load later.
4. **JSON-LD in the page source.** In "View page source" (Ctrl+U on Windows, Cmd+Option+U on Mac) you should find application/ld+json. If you only see it in the inspector, a script or Google Tag Manager adds it, and it reaches no assistant.
5. **Alt text on every image** that describes what the photo shows, not the file name. It reaches ChatGPT, Claude and Gemini, and Google Images uses it.

## How Clione uses this


Clione writes the same product data both ways: a JSON-LD graph with the product and its FAQs, for Google, Bing and Gemini, and text on the page, the title and a visible FAQ, for ChatGPT, Claude and Perplexity. On Shopify, both halves are in the page the server sends, no scripts, through [Shopify app embeds](/glossary/shopify-app-embed/). Clione also serves an llms.txt on your store's own domain.

Clione doesn't replace your SEO agency: it doesn't do ranking strategy, link building or editorial content. It makes sure the product data is complete, well formed and where each reader looks for it, so that work reaches the assistants too.

## What this study doesn't say


- **That reaching an assistant means being recommended.** We measured whether a fact reaches the assistant, not whether it uses it to answer or whether it changes what it recommends. The next layers of the study measure that.
- **How much search and selection weigh.** The index table gathers what each company documents and one outside analysis; we didn't measure it. How much of a recommendation depends on the index and on selection is for a later stage.
- **Anything about the consumer apps.** We measured through the APIs. The ChatGPT app, Claude.ai, the Gemini app, Copilot and Google's AI Mode may read differently, and we didn't measure them.
- **Anything about visibility, traffic or sales.** No figure in this table says a product page will show up more, get more visits or sell more.
- **That schema is useless.** JSON-LD is how Google and Bing read a product page, and Gemini receives it. What reached no assistant is the JSON-LD a script adds.
- **Anything about crawlers.** We measured assistants reading a page because someone asked, not any [web crawler](/glossary/web-crawler/): neither the training crawlers ([GPTBot](/glossary/gptbot/), [ClaudeBot](/glossary/claudebot/), CCBot) nor the search crawlers ([OAI-SearchBot](/glossary/oai-searchbot/), [PerplexityBot](/glossary/perplexitybot/), Claude-SearchBot), which are different robots. What each of them does with JavaScript is covered in the guide [What an AI crawler sees in your store](/guides/what-an-ai-crawler-sees/).
- **That it's final.** The results are provisional: we close the criteria once every layer is in, and if a later layer changes a figure, we correct it with a dated note in the change history.

The rest of the concepts behind how assistants read a store are laid out in the [AI search reading path](/learn/ai-search/).
