I've been thinking a lot lately about a moment that happens millions of times a day and that almost nobody in marketing has fully reckoned with yet.
Someone opens ChatGPT, or Claude, or Perplexity, and types something like "what's the best provider for X?" The assistant thinks for a second and names three or four brands. Everyone else — every other company that does the exact same thing — simply doesn't come up. They're not ranked lower. They're just absent.
For the brands that get named, this is about the best thing that can happen to them: a trusted recommendation delivered at the precise moment someone is deciding. For everyone else, it's a quiet kind of invisibility, the kind you don't even notice is happening to you.
So the obvious question, the one I keep getting asked, is how the assistant actually makes that call. And the reassuring answer is that it isn't random, and it isn't (for the most part) something you can just buy your way into. Once you understand what's going on under the hood, you can actually do something about it.
Three things are happening at once
When an AI decides who to mention, there are really three forces at work, and it helps to separate them:
- What the model already knows. These systems were trained on an enormous slice of the web, and brands that showed up often, consistently, and credibly across all that text became part of the model's baseline sense of a category. If you were well represented in good sources when the model learned about the world, you start with an edge. That's the part you can't retroactively change — but it's also not the whole story, which is good news if you weren't a household name three years ago.
- What the model retrieves in the moment. Most modern assistants don't rely purely on what they memorized — they go and pull live information to ground their answers. When someone asks a question, the system fetches current sources and builds its answer from them. This is the part I find most encouraging, because fresh, well-structured, findable content can get you recommended even if you were nowhere in the training data. You're not locked out by history.
- How much the model trusts you. This, in my experience, is the one that actually decides things. The AI isn't just tallying how many times your name appears — it's making a judgment about authority and independent validation.
That last point deserves dwelling on, because it's where most brands get it wrong. When review platforms, industry publications, and ordinary people in communities all say roughly the same thing about you, that reads as genuine reputation. When the only place your brand is described in glowing terms is your own website, that reads as marketing — and the model is specifically built to tell the difference.
Independent signals carry far more weight than anything you say about yourself, and I don't think that's going to change, because it's the whole reason people trust these answers in the first place.
The simple versionThe AI recommends the brands it recognizes, can find, and trusts.
The uncomfortable part
What follows from that is a truth I've watched frustrate a lot of good companies. The brands that win here aren't necessarily the biggest or even the best at what they do — they're the ones most legible to a machine.
I've seen genuinely excellent businesses lose this game simply because their reputation lived in places an AI can't see: closed communities, sales calls, word of mouth between people who already know them. Meanwhile a more AI-legible competitor gets named in their place. It isn't fair, exactly. But it is learnable.
What you can actually do
"You can't program the AI" is true, but it's also a bit of a cop-out, because you can shape every input it draws on. If I were prioritizing, this is the order I'd go in:
- Build real independent validation. Get genuinely talked about on the third-party sources these systems lean on — review platforms, industry press, the corners of Reddit and LinkedIn where your customers actually are. This matters more than anything on your own site, because trust is what tips the decision.
- Publish things worth citing. Original data, honest comparisons, direct answers to the questions people really ask — with the point made in the opening lines, not buried six paragraphs down.
- Strengthen your entity signals. Consistent brand descriptions, structured data, the same story about who you are everywhere you appear, so the model can connect the dots instead of defaulting to a competitor.
- Fix the boring technical stuff. None of the above matters if the crawlers can't get in. Make sure your key pages aren't blocked or hidden behind scripts an AI can't parse.
The reframe I'd leave you with
Be worth recommending — and make sure a machine can tell.
Getting recommended by an AI isn't really about gaming an algorithm, and I'd be a little suspicious of anyone selling it that way. It's about building a brand that's genuinely recognized, findable, and trusted across the open web — and then doing the unglamorous work of making that reputation legible to machines and not just to humans.
That's a healthier game than the keyword-stuffing era it's replacing, because the thing it rewards is closer to actually deserving the recommendation. The gatekeeper changed. The way to earn its blessing is, more or less, to be worth recommending — and to make sure a machine can tell.