Behind the algorithms: how AI chatbot recommendations work
AI chatbots like ChatGPT, Perplexity, and Copilot are rapidly emerging as new gateways to the internet, offering direct answers instead of traditional link lists. In this article, we help you figure out how their recommendations work – which is the first step to building an informed Generative Engine Optimization (GEO) strategy.
Key takeaways
- AI chatbots recommend from two sources: what the model learned in training, and pages it finds through a live web search for the question.
- For many questions they search, breaking one question into several sub-queries (query fan-out), then quoting and citing the passages that best answer each part.
- Brands that are mentioned often in trusted sources are recommended more. Brand mentions correlate with AI Overview visibility at 0.664 (Ahrefs).
- They don’t use “collaborative filtering” or learn from your clicks in real time. Personalisation comes from the conversation and, if enabled, the assistant’s memory.
- In shopping answers, structured product data matters: ChatGPT uses product metadata from merchants and data providers, and says product results are not ads.
AI chatbots have become a new way to find products, services and information. ChatGPT alone reached 900 million weekly users in February 2026. Instead of a list of links, users get a direct answer, often with a short list of recommended brands. If your brand isn’t on that list, many users won’t consider it.
To influence that list, you first need to understand how it’s put together. This article explains, step by step, how assistants such as ChatGPT, Perplexity, Gemini and Copilot decide what to recommend, and what that means for your website.
Do people really use AI chatbots to find products and services?
Yes, and the traffic is small but valuable. In an Ahrefs study of 3,000 websites, 63% received some traffic from AI chatbots, although it made up only 0.17% of the average site’s visits. Those visits convert well: Adobe found AI referrals to US retail sites converted 31% better than other traffic during the 2025 holiday season. Users arrive after the assistant has already compared options for them.
How does an AI chatbot build a recommendation?
Most AI assistants follow the same five steps. Details differ between platforms, but the logic is similar:
| Step | What happens | What influences it |
|---|---|---|
| 1. Understand the question | The model identifies the intent (learn, compare, buy), entities (product type, brand, features) and constraints (budget, size, location) | Wording of the question and earlier messages in the conversation |
| 2. Decide whether to search | Simple or general questions may be answered from training data; questions about products, prices or recent events usually trigger a web search | Type of question, need for current data, user settings |
| 3. Search and retrieve | The assistant runs several related searches (query fan-out) and fetches candidate pages | Search index (Bing, Google, own crawlers), rankings for sub-queries, crawlability |
| 4. Select and combine | It picks the passages that answer each part best and checks them against each other | Relevance, clarity, consistency across sources, freshness, brand reputation |
| 5. Write and cite | It writes the answer, names brands or products and links to some sources | Platform design; answers vary between runs |
1. Understanding the question
Language models interpret meaning, not exact keywords. “Elegant men’s winter shoes with a leather sole” is understood as a request for smart, warm shoes, with a constraint on the sole. Follow-up questions add context to the same conversation. More on this in intent-based search.
2. Training data or live search?
Every model has general knowledge from training data, collected up to a certain date. That’s often enough for definitions and well-known brands, but not for current prices, availability or new products. For those, assistants search the web: ChatGPT ran a search for 34.5% of queries in Semrush’s February 2026 data, while Perplexity, Google AI Overviews and AI Mode search for almost every question.
3. Searching with query fan-out
Assistants rarely run one search. They split the question into sub-queries, such as “best leather-soled winter shoes”, “are leather soles good in snow” and “Goodyear-welted winter boots price”. In a Peec AI analysis of 20 million ChatGPT fan-outs, these sub-queries roughly doubled in length, to about 12 words, between October 2025 and January 2026. Pages that rank for the sub-queries as well as the main query are more likely to be cited: 161% more likely in AI Overviews, according to Surfer.
4. Selecting sources
From the retrieved pages, the assistant picks passages that answer the question clearly and are confirmed by other sources. Signals that consistently show up in research:
- Brand mentions across the web. Branded mentions correlate with AI Overview visibility at 0.664; YouTube mentions at about 0.737 across ChatGPT, AI Mode and AI Overviews.
- Freshness. Pages cited by AI assistants are on average 25.7% fresher than organic search results.
- Clear, specific answers. Passages that answer one question directly are easier to quote than general marketing copy.
- Readable HTML. Major AI crawlers from OpenAI, Anthropic and Meta don’t run JavaScript, so content loaded by scripts may be invisible to them.
Cited pages often aren’t the top Google results: only about 12% of URLs cited by AI assistants also ranked in Google’s top 10 for the same prompt in an Ahrefs study.
5. Writing the answer
The assistant then writes the answer, names a few brands or products, and links to some of its sources. The result varies: in a SparkToro study, ChatGPT and Google’s AI returned the same list of brands in fewer than 1 in 100 repeated runs of the same prompt. Being mentioned consistently, across many runs and prompts, is what matters.
Let your customers find you, wherever they look.
How do personalisation and memory affect recommendations?
Within a conversation, assistants use everything the user has said: budget, style, brands to avoid. Some assistants also offer memory across conversations; in ChatGPT, users can turn it off or use a temporary chat. Location and language also shape results. What assistants don’t do is recommend based on what “similar users” bought, as shop recommendation engines do. Separately, OpenAI may use conversations to train future models unless users opt out, but this happens offline and doesn’t change recommendations in real time.
How do AI shopping recommendations work?
When a question has shopping intent, ChatGPT can show product cards with images, prices and links. According to OpenAI, it considers structured metadata from first-party and third-party providers (such as price and description) and other third-party content, and product results are selected independently and are not ads. Shopify merchants are included through Shopify Catalog, and other merchants can apply to submit a direct product feed.
For online stores, this means product data quality directly affects AI visibility: accurate names, specifications, prices, availability and reviews, consistent across your product pages, feeds and marketplaces. See product descriptions ready for AI search.
A real-life example: “elegant men’s winter shoes with a leather sole”
- Understanding: product type (winter shoes), style (elegant), gender (men’s), constraint (leather sole). The assistant may note that leather soles are slippery on ice.
- Context: if the user mentioned a budget or preferred brands earlier, it uses that.
- Search: sub-queries such as best leather-soled winter dress shoes, winter-proof leather soles and specific brands’ winter collections, plus reviews and forum discussions.
- Selection: shoes that meet the constraint, from brands mentioned in several sources, with current price and availability from store pages or product data.
- Answer: a short list of options with reasons, links, perhaps a note that a rubber sole is safer in snow, and a question about colour or budget.
- Refinement: “Anything in cognac?” or “Something cheaper?” starts new searches within the same conversation.
Brands that appear here are those with clear product pages, consistent data and frequent mentions in comparisons, reviews and discussions about the category.
What does this mean for website owners?
| What drives recommendations | What to do | Read more |
|---|---|---|
| Rankings for the main question and its sub-queries | Keep strong SEO; cover follow-up questions in focused pages and sections | Rankings and AI |
| Brand mentions in trusted sources | Earn mentions in media, comparisons, reviews, communities and video | Backlinks and mentions |
| Clear, quotable answers | Answer first, use question-style headings, tables and specific facts | GEO guide |
| Trust and expertise | Named authors, sources, first-hand experience | E-E-A-T |
| Accessible, current content | Server-rendered HTML; update key pages regularly | Structured data and crawling |
| Accurate product data | Complete, consistent product information across site, feeds and marketplaces | Product descriptions |
| What AI says about you | Monitor answers and correct outdated information at its sources | Brand safety, Metrics |
Boost sales with SEO + GEO.
Summary
AI chatbots recommend brands by combining what their models already know with what they find in a live search, broken into many sub-queries. They pick passages that are relevant, clear, current and confirmed by other sources, and favour brands that are widely mentioned. For your brand to be recommended, it needs to be findable in search, described clearly and consistently on your own site, and talked about elsewhere. Generative engine optimization builds on SEO to make that happen. See our GEO guide.
FAQ: how AI chatbot recommendations work
How does ChatGPT decide which brands to recommend?
It combines knowledge from training data with a live web search for many questions. It favours brands that appear in relevant, trusted and current sources, and that are mentioned consistently across the web.
Does ChatGPT learn from my clicks in real time?
No. It uses the context of your conversation and, if enabled, its memory. OpenAI may use conversations to train future models unless you opt out, but this doesn’t change answers in real time.
Why does ChatGPT recommend my competitor and not me?
Usually because your competitor is mentioned more often in the sources the assistant finds, ranks for more related sub-queries, or describes its offer more clearly. Answers also vary between runs, so test many times.
How does ChatGPT Shopping choose products?
OpenAI says it considers structured product metadata from merchants and data providers, other third-party content, its models and its product policies. Product results are not ads.
Do Perplexity and Copilot work the same way?
Broadly, yes: they search, select sources and write an answer with citations. Perplexity searches for almost every question and shows sources prominently; Copilot relies on Bing’s index.
Can I pay to be recommended by AI chatbots?
Not in organic answers or ChatGPT’s product results, which OpenAI says are not ads. Some platforms offer separate, labelled advertising.
Sources
- OpenAI Help: Shopping with ChatGPT
- OpenAI Help: How your data is used to improve model performance
- TechCrunch: ChatGPT reaches 900M weekly active users (Feb 2026)
- Ahrefs: AI traffic study (2025)
- Adobe: AI-driven traffic surges across industries
- MediaPost: AI chat relies on search to fill gaps (Semrush) (2026)
- Peec AI: ChatGPT fan-outs have doubled in length (Feb 2026)
- Surfer: Query fan-out impact on AI Overview citations (2025)
- Ahrefs: AI Overview brand visibility factors (May 2025)
- Ahrefs: Brand visibility factors in ChatGPT, AI Mode and AI Overviews (Dec 2025)
- Ahrefs: Do AI assistants prefer to cite fresh content? (2025)
- Ahrefs: AI search overlap with Google’s top 10 (Aug 2025)
- Vercel: The rise of the AI crawler (Dec 2024)
- Search Engine Journal: AI recommendations change with nearly every query (SparkToro) (Jan 2026)
About the author: This article was prepared by the Insightland SEO and GEO team. Insightland is a Wrocław-based agency, part of the Hexe Capital Tech Hub, that connects business, tech and data: SEO and GEO, performance marketing, UX and analytics, and SEO and AI training. Contact us.