How Copilot, Perplexity and ChatGPT decide which sources to cite

When someone asks Microsoft Copilot or Perplexity to recommend a product, the model does not answer from memory. It runs one or more searches, pulls back a handful of pages, and writes an answer grounded in what those pages say. That intermediate step — retrieval — is where visibility is won or lost. If your page is not among the retrieved documents, no amount of brand strength puts you in the answer.

The three stages of an AI answer

Almost every consumer answer engine follows the same shape, whatever the model underneath.

• Query fan-out: your question is rewritten into several search queries, often more specific than what you typed.
• Retrieval: each query hits a search index and returns candidate pages. Only a small number survive to the next stage.
• Synthesis: the model reads the surviving pages and writes an answer, attaching citations to the passages it used.

Two consequences follow. First, a page must exist in the index the assistant queries. Second, the passage the model needs must be readable as plain text on that page — not locked in an image, a video, or a JavaScript widget that never renders for a crawler.

Which index each assistant reads

Microsoft Copilot is built on the Bing index, which is also the index behind Bing.com, Yahoo Search and a long tail of smaller search products. Perplexity operates its own crawler (PerplexityBot) and supplements it with third-party search results. ChatGPT's browsing mode uses its own crawler, OAI-SearchBot, alongside partner search data. Google's AI Overviews sit on Google's index.

This is why a single strong page can pay off across several assistants at once: they overlap heavily on the same well-crawled, high-authority domains. It also means blocking a crawler in robots.txt quietly removes you from that assistant's answers, however well you rank in classic search.

What makes a page quotable

Retrieval gets you into the candidate set. Being quoted is a separate hurdle, and it is largely a writing problem rather than a technical one.

• Answer the question in one self-contained sentence, near the top, before any preamble.
• Use the words people actually type. Models match on phrasing more literally than experienced marketers expect.
• State concrete facts — prices, dates, capacities, locations, who it is for — instead of adjectives.
• Attribute numbers to a named source and a date; unattributed claims are the first thing a cautious model drops.
• Keep one idea per paragraph so a passage can be lifted without losing its meaning.

A useful test: read a single paragraph from your page out of context. If it still answers a real question and names your brand, it is quotable. If it only makes sense after three paragraphs of build-up, it is not.

Why third-party pages carry more weight

Assistants are tuned to prefer sources that look independent, because a self-description is weak evidence. A claim on your own website supports one type of answer — what your company says about itself. The same claim on an established editorial domain supports a stronger one — what the wider web says about you. This is the mechanical reason placements on large publishers change AI visibility faster than another page on your own site.

No one can guarantee what an AI assistant will say about a brand. Retrieval and ranking change continuously, and models deliberately vary their sources. What you can control is whether an accurate, well-sourced, crawlable page about you exists on a domain these systems already read.

A short checklist

• Confirm your key pages are indexed and return plain HTML text to a crawler.
• Allow the assistant crawlers you want to be visible in, and check robots.txt for accidental blocks.
• Publish a plain-language facts page: what you sell, to whom, at what price, where.
• Add structured data that matches the visible page, not a marketing version of it.
• Earn at least one credible third-party page that states the same facts.