How AI Assistants Pick Products to Recommend

When a shopper asks an assistant "what's the best automatic litter box under $400," the answer is assembled from two very different sources. The first is training data — the text the model learned from, which is months to years old and encodes brand-level knowledge: which names are established and roughly what they are known for. The second is live retrieval — a real-time search step that fetches current pages and quotes or summarizes them.

The split matters because the two are influenced on completely different timescales. Training data is effectively frozen; a new pet supplies brand cannot appear in it until models are retrained on a web that already talks about the brand. Retrieval, by contrast, updates as fast as the underlying index — a page published this month can be fetched and cited this month.

Every practical lever in this article targets retrieval, because that is the layer a merchant can move this quarter. Training-data presence follows later, as a byproduct of the same third-party coverage that helps retrieval today.

"AI search" is not one system. Each assistant has its own retrieval supply chain, and knowing them tells you where indexing effort pays off.

• ChatGPT search retrieves from established web indexes and fetches pages with OAI-SearchBot — a separate user agent from GPTBot, which collects training data.
• Microsoft Copilot leans on Bing's web index, so a store invisible to Bing is invisible to Copilot.
• Perplexity operates its own crawler, PerplexityBot, and is unusually generous about citing sources with links.
• Google Gemini and AI Overviews draw on Google's index — ordinary Google indexing is the prerequisite for both.
• Claude fetches pages with ClaudeBot when browsing is invoked.

The practical consequence: your existing SEO indexing work is not separate from AI visibility — it is the substrate. A page that Google and Bing cannot see will not be recommended by Gemini or Copilot no matter how well it is written, so fix plain indexing problems before polishing anything else.

Across assistants, the retrieved pages that make it into answers share a shape. They are answer-shaped: the direct answer appears early and unhedged, followed by the specifics that justify it. An assistant summarizing five pages under time pressure quotes the one that already reads like an answer.

Concretely, for a coffee gear store: a page titled "What grind size for a moka pot" that opens with "slightly finer than drip, coarser than espresso" and then explains why will beat a 1,200-word brand story that finally mentions grind size in paragraph nine.

Structure is the second half. schema.org markup — Product with a nested Offer carrying price and availability, FAQPage for question blocks — lets an assistant extract facts without parsing prose. Most Shopify themes emit Product JSON-LD already; verify rather than assume, using a structured-data validator on your top ten pages.

Third: assistants trust corroboration. A product that appears in multiple independent comparison articles gets recommended with confidence; a product described only by its own store page is a single unverified claim, and models hedge on those.

None of this matters if the assistants cannot fetch your pages, and plenty of stores block AI crawlers without knowing it — often via a robots.txt edit made during a content-scraping panic, or a blanket rule copied from a blog post.

1. Open yourstore.com/robots.txt in a browser.
2. Search the file for GPTBot, OAI-SearchBot, PerplexityBot, and ClaudeBot. A Disallow: / under any of those user agents removes you from that assistant's retrieval.
3. On Shopify, edit through the robots.txt.liquid theme template (Online Store → Themes → Edit code). Shopify's default file does not block these bots, so if they are blocked, someone added the rule — find out why before deleting it.
4. Know the distinction the names hide: blocking GPTBot opts you out of OpenAI's training data but leaves ChatGPT search working; blocking OAI-SearchBot does the reverse. Google-Extended only controls Gemini training — it does not affect your normal Google rankings or AI Overviews.

Decision rule for a store: allow the search-facing bots. Being retrievable is the whole game here.

Assistants hedge on stores they cannot confidently identify. If your brand appears as "Whisker Works," "WhiskerWorks Pet Co.," and "whiskerworks-store" across your site, social profiles, and marketplace listings — each with a different one-line description — a model has weaker evidence that these are the same entity. Weak entity confidence loses recommendations to competitors with cleaner footprints.

The fix is unglamorous: choose one canonical brand name and one two-sentence description, and apply them everywhere — storefront footer, about page, every social bio, every marketplace seller profile. Same spelling, same spacing, same claims. Machines reward the repetition that copywriters find boring, and the cross-referencing an assistant does before naming you is exactly where that repetition pays.

A curated /llms.txt file plays a small supporting role here: it states in one place who you are and which pages represent you best. The format — and the Shopify-specific problem of serving a root file at all — is covered in our llms.txt guide; file it under cheap-but-unguaranteed rather than essential.

Ask assistants a "best X" question and inspect the citations: they are dominated not by store pages but by comparison and best-of articles on third-party sites. Assistants treat an independent roundup as pre-digested judgment — someone already compared the options — and lift recommendations from it. Your product appearing in those articles matters more than almost anything on your own domain.

Earning that presence is classic PR at niche scale:

• Identify the roundups that already rank for your money questions ("best dog puzzle feeders," "best hand grinders for travel") and pitch the authors with a sample and a spec sheet.
• Prioritize genuinely independent reviewers over pay-to-play listicles; assistants cite the former's domains far more often.
• Make the reviewer's job easy: exact product names, current prices, measurements, and photos they can reuse.

Expect the assistant to cite the roundup rather than your store — that is fine. The recommendation still names your product, and the shopper's next search is for you.

If you act on one section this week, make it the crawler check — it takes ten minutes and gates everything else. After that, work in this order: verify structured data on your top ten pages, rewrite your three most-asked-about category pages into answer shape, standardize your entity footprint, then start the slower outreach for third-party reviews, which pays off over months rather than weeks.

Measure with monthly spot-checks — ask the assistants your customers' actual questions and log whether you appear — because no analytics dashboard reports citations directly. A full month-by-month sequencing of this work, including that measurement loop, is in our generative engine optimization guide.

The through-line: assistants recommend what they can fetch, parse, and corroborate. If keeping a store's worth of answer-shaped content published and current is more production than you can staff, Hushwork generates and hosts that content on your own domain and checks weekly whether assistants mention your brand.