How AI engines pick what to recommend

The engine doesn't judge which product is better. It reads a handful of pages and repeats what they say. So the names in an answer are decided by whoever bothered to write those pages.

The three engines work differently

ChatGPT searches Bing, reads three or four results, and builds the answer from those. It doesn't always search — if the question looks like something it already knows, it answers from training data instead, which means anything newer than the cut-off is invisible.

Perplexity searches on every question, runs its own index, and reads five to ten sources. Because it reads more, smaller names appear there that don't appear anywhere else.

Google AI Overviews doesn't run a fresh search. It builds the answer from pages that already rank. Whatever wins in normal Google search tends to win here too.

Same question, three different source sets. Overlap between them is small — one study found only about 2.4% of cited URLs appeared in all three engines for the same question.

What the sources actually are

Across categories, the same three types keep appearing:

Listicles. "Best X software 2026." Usually written by someone who competes in the category, or by a site that monetises the traffic.

Comparison pages. "A vs B." Almost always written by A or B.

Reddit and forums. Where people answer each other, which the engines treat as unbiased.

Review sites like G2 and Capterra show up too, but less often than most people assume. In some categories they don't appear at all.

Your own site is a minority of what gets read. Measured across SaaS, ecommerce and finance, roughly 70–85% of cited sources are third-party.

Why the same few names keep winning

The pool of sources is small. If four pages carry a category's answers and three of them name the same company, that company is in almost every answer.

This compounds. Being named in answers gets you mentioned in more articles, which become more sources, which get you named more.

In one category I measured, one tool appeared in fifteen of the sixteen answers that named anything. Everything else split what was left.

Where the smaller names appear

Narrow questions.

Nobody writes a listicle for "software with client photo history" or "tool that works offline on iPhone." So there are two or three pages answering it, and a small company can be one of them.

The broad question in a settled category is usually lost. The specific one isn't.

What this means

You can't make an engine like your product more. It has no opinion about your product.

What you can change is which pages exist about you, and whether the pages the engine already reads say your name.