# How AI assistants decide which products and stores to recommend

> What's known and unknown about how AI assistants pick products and stores, and the three gates (read, verify, agree) your store must pass first.

By Ghost Agent Lab · Start here · https://ghostagentlab.com/articles/how-ai-assistants-choose-products/

## Key takeaways

- No AI company publishes how it ranks products in an answer, but they do document how their crawlers reach sites and what product data they accept.
- Before any ranking applies, an assistant has to read your pages, verify price, stock and policies, and find that your sources agree.
- Blocking an assistant's search crawler or fetcher can keep you out of its answers, whatever your products are like.
- When your page, structured data and feeds disagree on price or stock, the assistant has to guess, hedge or leave you out.
- This week, check that your ten best sellers show price, currency and availability in the HTML and in matching structured data.

When a shopper asks an AI assistant for "the best waterproof trail shoe under $150", a short list of products and stores comes back. This guide explains what is publicly known about how that list gets made, what isn't, and the groundwork that decides whether your store can be on it at all.

## What we know and what we don't

No AI company publishes how it ranks products in an answer. Anyone who claims to know the exact formula is guessing. What the companies do publish is the plumbing: which crawlers and fetchers they use, how to let them in, and what data their shopping features accept. That is enough to see the shape of the process, even if the final weighting stays private.

Here is what the vendors' own documentation says today:

- **Google** says its AI Overviews and AI Mode have no additional requirements beyond normal Search. A page must be indexed and eligible to show with a snippet. Its [guidance on AI features](https://developers.google.com/search/docs/appearance/ai-features) lists the basics that still matter: allow crawling in robots.txt and in your CDN, put important content in text, make structured data match the visible text, and keep Merchant Center and Business Profile information up to date. It also says both features may run many related searches behind one question (it calls this "query fan-out").
- **OpenAI** says sites that opt out of its search crawler, OAI-SearchBot, won't be shown in ChatGPT search answers, though they can still appear as navigational links ([OpenAI crawlers](https://developers.openai.com/api/docs/bots)). Separately, approved partners can send ChatGPT a product feed with titles, descriptions, images, price and availability ([OpenAI commerce docs](https://developers.openai.com/commerce/guides/get-started)).
- **Anthropic** says blocking Claude-User or Claude-SearchBot may reduce your site's visibility in answers to user searches ([Anthropic's crawler help page](https://support.claude.com/en/articles/8896518-does-anthropic-crawl-data-from-the-web-and-how-can-site-owners-block-the-crawler)).
- **Perplexity** describes PerplexityBot as the crawler that surfaces and links websites in its search results, and Perplexity-User as the fetcher that visits a page when someone asks a question ([Perplexity crawlers](https://docs.perplexity.ai/docs/resources/perplexity-crawlers)).

What none of them say: how much reviews count against price, whether a cheaper store beats a better-known one, or how often a feed is preferred over a page. Treat any checklist that promises a ranking as a hypothesis, not a fact.

- ✓ Which crawlers and fetchers each assistant uses
- ✓ That blocking them can keep you out of answers
- ✓ That Google wants structured data to match the page
- ✓ What fields a product feed accepts
- ✗ How products are ranked inside an answer
- ✗ How much reviews, price or brand weigh

**Known and unknown.** The vendors document access and data formats, not the ranking itself.

## Three gates before a recommendation

The ranking is private, but the steps before it are not. Before an assistant can recommend your product, it has to get through three gates. Fail any one and the ranking never matters, because you aren't in the pool.

1. **Read:** Can it reach your pages at all?
2. **Verify:** Can it find price, stock and policies it trusts?
3. **Agree:** Do your page, data and listings say the same thing?
4. ✓ **Compare:** Only now does the private ranking apply

**Three gates, then the ranking.** You can't influence the last step directly, but you control the first three.

### Gate 1: can it read you?

Assistants reach your store in several ways: a search crawler that indexed your pages earlier, a fetcher that opens a page while the shopper waits, or a product feed. See [the AI agents that visit your website](https://ghostagentlab.com/articles/kinds-of-ai-agents/) for the full cast. Each one can be stopped by something you control:

- A robots.txt rule that blocks an AI search crawler or assistant fetcher. See [robots.txt for AI agents](https://ghostagentlab.com/articles/robots-txt-ai-agents/).
- Bot protection that challenges or blocks automated requests, which is what an assistant's fetcher sends. See [bot protection and CAPTCHAs](https://ghostagentlab.com/articles/bot-protection-ai-agents/).
- Product details that only appear after JavaScript runs. Many fetchers read the HTML your server sends and nothing more. See [JavaScript-only content](https://ghostagentlab.com/articles/javascript-content-ai-agents/).

### Gate 2: can it verify the facts?

A shopper's question usually has hard constraints: under $150, in stock, ships to Canada, free returns. An assistant can only include you if it can check each one. The more of them it can read without guessing, the safer you are to recommend.

| The shopper asks about | What the assistant needs to read | Where to put it |
| --- | --- | --- |
| Price | A number and a currency | In the page HTML and in Product structured data |
| Stock | In stock, out of stock or preorder | Visible text and the `availability` field |
| Fit and features | Specs in words, with units | Product description and attributes |
| Quality | Ratings and review text | Reviews in the HTML, marked up honestly |
| Delivery and returns | Costs, times, windows, conditions | Policy pages written as plain sentences |

In [our scan of 255 online stores](https://ghostagentlab.com/blog/agentscore-store-study-2026/), product details were where readability slipped most. Our guides to [product data](https://ghostagentlab.com/articles/structured-data-ai-shopping-agents/) and [shipping and returns pages](https://ghostagentlab.com/articles/policy-pages-ai-assistants/) cover the rows in detail.

- **63%** of stores give price, currency and availability as product structured data
- ! **19%** have no usable product structured data at all
- ✗ **15%** don't have prices in the product page's HTML

**Where verification slips.** From our scan of 255 online stores in October 2026.

### Gate 3: do the facts agree?

An assistant may see your price in three places: your product page, your structured data and a feed or marketplace listing. It may also read a review site, a comparison article or an old forum thread. When those disagree, the assistant has to choose, hedge or leave you out. None of those help you.

Google is explicit on one part of this: structured data should match the visible text. The same logic applies everywhere. If the page says $129, the markup says $139 and a feed says "out of stock", nothing tells the assistant which is true.

**Conflicting:**

- ✗ Page shows $129, markup says $139
- ✗ Feed says out of stock, page says in stock
- ✗ Returns page says 30 days, FAQ says 14

**Consistent:**

- ✓ One price, in the page and the markup
- ✓ Stock updated in page, data and feed together
- ✓ One returns policy, linked from the FAQ

**Give it one answer.** When every source agrees, the assistant has nothing to hedge about.

## Where the answers come from

Assistants don't rely only on your site. Depending on the product and the question, an answer may draw on any of these:

- **Your own pages:** Fetched live or from a search index
- **Product feeds:** Merchant Center, or a feed sent to an assistant where offered
- **Business listings:** Such as a Google Business Profile, for local stores
- **Other sites:** Reviews, comparisons, articles and forums

**Four kinds of source.** You control the first three directly; the fourth you can only influence.

Two practical points follow. First, your own pages are the one source every assistant can use and that you fully control, so they should be the most complete and the most current. Second, when another site describes you wrongly, the fix usually starts on your own site: a clear, quotable page that states the fact correctly gives assistants something better to cite. Feeds are covered in [product feeds and merchant listings for AI shopping](https://ghostagentlab.com/articles/product-feeds-ai-shopping/).

> **No special file required:** Google says you don't need new machine-readable files, AI text files or special schema.org markup to appear in its AI features. An llms.txt file can still help other assistants find your key pages, but it is no substitute for pages they can read.

## What you can and can't influence

It helps to be honest with your team about where effort pays off. You can't buy your way into an organic answer or tune it the way you tune a bid. You can remove every reason for an assistant to skip you.

- **In your control:** whether assistants can reach your pages, whether facts are in the HTML and in structured data, whether those facts agree, how clearly your policies are written, and whether your product and merchant listings are complete.
- **Partly in your control:** reviews (you can ask for them and show them honestly, but not write them), and how other sites describe you (you can correct the record with them, but they decide).
- **Not in your control:** the ranking itself, which assistants a shopper uses, and how often each one refreshes what it knows.

This is the same split SEO teams already know, with one difference: the assistant often picks for the shopper rather than showing ten links.

## What AgentScore checks

AgentScore doesn't measure whether assistants recommend you. It tests the groundwork for the first two gates:

- **Reading:** whether robots.txt lets AI assistants and AI search agents reach your home page, and whether your bot protection gives requests as ChatGPT-User, Claude-User, Perplexity-User and OAI-SearchBot the same home page as a browser. It also checks that a product, pricing or cart page opens for an AI assistant as it does for a browser, and isn't closed to assistants in robots.txt.
- **Verifying:** whether a product page gives price, currency and availability in Product structured data in its HTML, and whether prices appear in the page HTML at all, without JavaScript.

It doesn't compare your page with your feeds, or check what other sites say. To see what an assistant actually says about you, track a fixed set of questions over time. See [tracking how AI assistants describe your brand](https://ghostagentlab.com/articles/brand-answers-ai-assistants/).

## Where to start

> **Try it:** Ask two or three assistants the question your best customer would ask, without naming your brand. Note who they name and which pages they cite. Then open your own product page for that question and check that price, stock and returns can be read in plain text.

1. Run an [AgentScore](https://ghostagentlab.com/agentscore/) scan and fix anything under Access first. If assistants can't read you, nothing else matters.
2. Pick your ten best-selling products. Check that each page has price, currency and availability in the HTML and in Product structured data, and that they match what a shopper sees.
3. Read your shipping and returns pages as a stranger would. Make sure each key fact is one plain sentence, and that no other page contradicts it.
4. Compare your feed or Merchant Center listing with the live page for the same ten products.
5. Start a monthly record of how assistants answer your key questions, so you can see whether the fixes change anything.

## Sources and further reading

- [AI features and your website](https://developers.google.com/search/docs/appearance/ai-features) (Google Search Central)
- [Overview of OpenAI crawlers](https://developers.openai.com/api/docs/bots) (OpenAI)
- [Agentic Commerce: Get started (product feeds)](https://developers.openai.com/commerce/guides/get-started) (OpenAI)
- [Does Anthropic crawl data from the web, and how can site owners block the crawler?](https://support.claude.com/en/articles/8896518-does-anthropic-crawl-data-from-the-web-and-how-can-site-owners-block-the-crawler) (Anthropic)
- [Perplexity crawlers](https://docs.perplexity.ai/docs/resources/perplexity-crawlers) (Perplexity)