How AI Decides Who to Recommend, Explained Gently

A friend asked me the other week why, when she asked an AI assistant for “the best travel insurance,” it named three companies and not the one she’d actually bought. Good question. It’s the question, really. The whole industry I work in is currently rearranging itself around it. Let me try to answer it the way I’d answer it over coffee, without the jargon.

First: it isn’t “looking things up” the way you’d think

When you Google something the old way, a list of pages comes back and you do the choosing. When an AI answers, it does the choosing for you and hands you a conclusion. That’s the whole shift, and everything strange about the new world falls out of it.

So the real question isn’t “how does it find pages.” It’s “how does it decide which handful to trust enough to repeat.” And the honest answer is: partly things we understand, partly things nobody outside the labs fully does.

What it seems to lean on

From watching a lot of these answers, day after day, on sites I look after, here’s what appears to matter, stated plainly, and with the caveat that I’m reading tea leaves, not reading source code:

• Being said clearly, in one place. If a page answers the actual question in a plain sentence, “X costs €12 a month and covers Y”, it gets quoted far more than a page that buries the same fact under 800 words of throat-clearing. The machine likes a straight answer for the same reason a busy person does.
• Being said in more than one place. A claim that shows up, consistently, across several independent sources reads as “settled.” One lonely page saying something no one else says gets treated with suspicion, reasonably.
• Being attached to someone. Pages with a real author, a real date, and a track record seem to travel further than anonymous ones. Not always. But the trend is real, and it’s getting stronger.

The uncomfortable honest part

Here’s what I won’t pretend: a lot of it is opaque, and some of it is luck. Two nearly identical pages can get wildly different treatment. The systems change without warning. Anyone who tells you they’ve “cracked” how AI picks sources is selling something.

What I actually believe, after a couple of years of watching: you can’t game your way in, but you can be the obvious thing to quote. Answer the real question, answer it clearly, be a source a reasonable person would trust, and do it consistently. That’s not a trick. It’s just… being good, legibly. Which, annoyingly, is the same advice as before; it’s only the stakes that changed.

So when the assistant didn’t name my friend’s insurer, it probably wasn’t a snub. Her insurer just hadn’t made itself the easy, clear, corroborated thing to say. That’s fixable. Most of this is, once you stop looking for the trick and start looking for the plain sentence.