How AI Assistants Choose Which Products to Recommend
Ask an AI assistant for “a good espresso machine under $400 for a small kitchen” and you’ll get two or three specific recommendations, often with links. Those picks aren’t random, and they aren’t paid placements. They’re the output of a pipeline you can understand — and optimize for.
Two sources: model memory and live retrieval
Every assistant blends two kinds of knowledge:
- Model memory — what the model absorbed during training. This is broad but slow-moving: it reflects how the web described your brand months or years ago, and you can’t change it quickly.
- Live retrieval — what the assistant fetches from the web while answering. For shopping questions, this is now the dominant mode: the assistant runs searches, reads a handful of pages, and builds its answer from them, usually with citations.
The practical takeaway: model memory rewards long-term brand consistency, but live retrieval is where you can win this quarter — it re-reads the web every time someone asks.
The four steps between a question and a recommendation
Most assistants follow roughly the same funnel:
1. Interpret the question
The assistant decomposes the query into constraints: category (espresso machine), budget (under $400), context (small kitchen → compact footprint). Your content can only match constraints it actually states. If your product page never mentions dimensions, you can’t win the “small kitchen” half of the question.
2. Retrieve candidates
The assistant issues search queries against a web index and its own crawl data. What surfaces here looks a lot like classic search results: product pages, category pages, buying guides, comparison articles, forum threads. If your pages aren’t crawlable and indexable, you’re eliminated in this step — before any judgment of quality happens.
3. Read and synthesize
Now the model actually reads the retrieved pages, more like a careful human than a keyword matcher. Pages that state facts plainly win here:
- Specs in visible text (not just images or JS-rendered widgets)
- Clear pricing and availability
- Honest, specific descriptions (“38 cm wide, fits under standard cabinets”) rather than adjectives (“sleek, game-changing design”)
- Reviews and ratings visible in the HTML
4. Select and cite
Finally the assistant picks what to name. Corroboration matters heavily: a product that appears on the merchant’s own site and in an independent buying guide and in a Reddit thread is a much safer recommendation than one the model has seen exactly once. Ambiguity is penalized — if your product name is inconsistent across your site, marketplaces and press mentions, the model may not connect them into one entity.
Signals that repeatedly matter
Across assistants, the same factors keep showing up in what gets recommended:
- Crawl access for AI bots (
GPTBot,ClaudeBot,PerplexityBot, etc.) - Server-rendered, readable HTML — content that exists before JavaScript runs
- Complete structured data — Product, Offer, AggregateRating, Organization
- Consistent naming of brand and products across every channel
- Independent mentions — editorial lists, niche reviewers, communities
- Freshness — updated prices, availability and content
- Comparison-friendly content — pages that answer “X vs Y” and “best for Z” in your own words
What silently filters stores out
The inverse list is just as important, because these failures are invisible until you go looking:
- Blocking AI crawlers in robots.txt or via aggressive bot protection. The assistant doesn’t complain — it just recommends someone else.
- JavaScript-only content. Many AI fetchers read raw HTML and execute little or no JS. If your specs live in a JS-rendered tab, they may not exist as far as the model is concerned.
- Thin manufacturer descriptions duplicated across dozens of stores. The model has read that exact paragraph elsewhere; it gives you no reason to be the cited source.
- Inconsistent entity data — different product names, prices or claims on your site vs. marketplaces vs. social profiles.
How to act on this
Start from the buyer’s side, not your catalog: collect the questions your customers actually ask an assistant, then check — for each one — whether you survive step 2 (retrieval), step 3 (readability) and step 4 (corroboration). That audit is exactly what AppearAI runs for Shopify stores: it scores your visibility across 8 dimensions, shows the real buyer questions for your category, and lists which competitors AI names when you don’t come up.
For the fixes themselves, work through our 12-point GEO checklist for Shopify stores. And if you’re new to the topic, start with What is AI search visibility?