Why ChatGPT recommends some websites

How ChatGPT and Perplexity retrieve sources, choose what to cite, and recommend brands—explained through retrieval, evidence, and trust.

Quick answers

How do AI assistants decide which websites to recommend?

For current questions, assistants like Perplexity and search-connected ChatGPT run live web searches, fetch a handful of top results, and compose an answer from what those pages state. A site is recommended when it is retrieved for the underlying searches, states liftable facts, and looks safe to repeat — not because of a private quality ranking.

Why does ChatGPT recommend some brands and not others?

Usually because the recommended brands have retrievable pages that match the question and state concrete, corroborated facts, while others are missing from the search results the assistant used, hide key facts behind JavaScript or vague copy, or lack the third-party corroboration that makes a claim safe to repeat.

How do I get my product recommended by AI assistants?

Publish pages that match the questions buyers actually ask, state what the product is, who it is for, and what it costs in plain server-rendered text, keep facts consistent across pages, and earn third-party corroboration. Then monitor the questions over time to verify whether retrieval and citations change.

Recommendation is retrieval, not memory

When someone asks an AI assistant "what's a good invoicing tool for freelancers?", the model doesn't consult a private ranking of every product in the category. For anything current, assistants like Perplexity and search-connected ChatGPT run a live retrieval step: they issue searches, fetch a handful of pages, and compose an answer from what those pages actually say.

That has a blunt implication: to be recommended, you first have to be retrieved. If your site doesn't surface for the underlying searches — because the pages are thin, client-rendered, or blocked — the model never sees you, and no amount of product quality changes the answer.

What survives the retrieval step

Pages that state plainly what the product is, who it is for, and what it costs — the exact facts an answer needs to include.
Direct-answer copy: a sentence the model can quote is worth more than a paragraph it has to interpret.
Specific intent pages — comparisons, use cases, pricing — that match the question as asked, not just your brand name.
Server-rendered HTML. Retrieval fetches are fast and rarely execute your JavaScript.
Corroboration: third-party mentions, reviews, and consistent details across pages make a claim safe to repeat.

Why models hedge, and how trust breaks the tie

Assistants are penalised for confidently recommending something that turns out to be wrong, so they hedge when the evidence is thin. A site with no about page, no visible team, no pricing, and template copy reads as unverifiable — and unverifiable products get dropped from answers or buried in a "there are many options" list.

Trust signals are how you become the safe answer: an identifiable entity (Organization schema, a real about page), claims that match across your site, visible pricing, and pages that answer the follow-up questions a cautious buyer would ask next.

The pipeline, step by step

How an AI assistant turns a buyer question into a recommendation
Step	What happens	Where sites drop out
Query rewriting	The assistant reformulates the question into one or more web searches	Sites with no page matching the rewritten searches are never requested
Retrieval	A search index returns top results; the assistant fetches a handful	Pages outside the top results, or blocked and slow pages, are not fetched
Reading	Fetched HTML is parsed for facts that answer the question	Client-rendered, vague, or fact-free pages contribute nothing
Synthesis	The model composes an answer from claims it can support	Unverifiable or contradictory claims are hedged or dropped
Citation	Sources supporting specific claims are linked	Pages that informed the answer without a liftable claim may go uncredited

How an answer engine gets from a question to a shortlist›

A buyer question goes to retrieval, which returns candidate pages. Only some of those pages are used: the engine synthesises an answer from a small number of them. Each brand in the result holds a role — cited when the answer used its page as a source, recommended when it was put forward as an option, mentioned when merely named, and absent when it appears at all. The step that decides most outcomes is retrieval: a page that is never fetched cannot be cited however good it is.

See how you do on the questions that matter

You can test this directly: ask an assistant the questions your buyers ask and see who gets named. A free Robot Visible scan does the systematic version — it grades whether your site gives retrieval something to find and answers something worth citing, then ranks the gaps by impact.^1,2

Sources and further reading

ChatGPT Search — OpenAI. Describes query rewriting and the use of search providers when ChatGPT retrieves current web information.
AI features and your website — Google Search Central. Explains retrieval eligibility, query fan-out, and textual-content requirements for Google's AI features.

Continue learning

How to improve AI visibility: the AEO loop AI visibility guide: get found and cited How long do ChatGPT and LLM citations take?AI visibility for B2B SaaS

See where your website stands

Run a free scan and get your AI readiness score across all six categories, with the gaps to fix first.

Run a free scan More guides