Before an AI model can name your business, it has to work out who you are. That comes first, and everything else you try depends on it.
To get AI to recommend your business, a model has to do two things. It has to work out who you are, meaning what you do, who it's for, and where you work. Then your business has to be the answer when people ask AI about the place you want to be. The first part comes first.
You asked ChatGPT about your own field last month and a competitor's name came back. Not a partial match. Not a maybe. Their name, in the answer, in the slot where yours belongs.
Most of them didn't do anything clever. A model recommends a business when it can confirm three things fast: what you sell, who you sell it to, and that somebody other than you has said so.
The question you typed is the one a buyer types when they're close to hiring. A missing name is a call that goes somewhere else.
Two things cause this, and they show up together. Your details read differently everywhere they appear, so the model can't settle on one version of you. And almost nobody else on the web writes about you, so it falls back on the version it already holds. That version has somebody else's name on it.
Retrieval is the step where the model goes looking. It finds pages on the web that mention you and reads them before it answers. That step comes before every tactic you'll hear about, because a tactic only changes what the model does with a page it has already found. If the pages it finds hold no clean version of your business, there's nothing to change.
The part I'd want a buyer to know is this. Olivier Martinez reviewed 45 studies on generative engine optimization in a preprint posted on 15 July 2026. One author, and it hasn't been peer reviewed. In it, content a model has already retrieved can change how that content gets cited or used. No trick he reviewed showed a steady effect on whether a page gets found in the first place.
The oldest studies in that pile go back to November 2023. Nobody has shown the steady version since. Meanwhile the work gets sold, week after week, as if finding pages were already settled.
This is the order I'd work in. A model carries one version of a business, and it leans on whichever version it reads most often. So the foundation is the part that has to be settled before anything else, because a model that can't work out who you are has nothing to recommend.
So write the sentence you want it to hold. What you do, who for, where you work. One sentence. Not a paragraph, and not a page.
Make that sentence read the same everywhere your business appears, starting with your own site. Then get other people to write about you, because models lean on pages they didn't get from you. A professional body. A directory. A podcast write-up. A local news mention.
Then look at what a machine reads that you don't. Behind every page sits text no visitor sees: the page title, the short description under it, the labels on images, the code that tells a machine what the page is about. A model reads all of it. If that background says one thing and your front page says another, you've told it two stories, and it can't tell which one is you.
Only then does the second part start, and it doesn't end. Find out what people ask AI about the subject you want your business to be the answer on, then answer those questions, in the buyer's words, on a page of your own. A model that can't find a straight answer on your site will lift one from a competitor's.
Then check it, and keep checking. I open a model in a fresh window, because my own chat history feeds me back the answer I want. Ask what it knows about you. Then ask who it recommends. Run two models, and run it twice. One answer on a Tuesday is a reading, not a verdict.
I sell AI visibility work, so weigh this part with that in mind. I can't prove a citation turns into a customer. Nobody can.
The people who say they can prove a citation turns into a customer can't show you one either. Pew Research Center followed 900 U.S. adults through 68,879 Google searches in March 2025. An AI summary appeared in 12,593 of those searches, and in about 1 percent of those visits somebody clicked a link inside the summary. A model can point at a page and the reader still won't follow it.
So the work I'd defend is narrow. Can a model work out who you are, and is that description strong enough to get picked ahead of the names it already reaches for.
Find out which of those two is broken before you pay to fix either one. That's what the AI Visibility Check does. It asks the models what they know before they ever see your site, scores what comes back, and names the competitor a model reaches for instead. If it can't describe you at all, the repair is mostly technical. If it describes you correctly and still hands the buyer to someone else, more site work is money spent on the part that already works. I wrote about that split in Cited is not named. Named is not recommended..