The customer asked the AI. It didn't say your name.

A growing share of buying decisions now start with a question typed into a model, not a search bar. When that model answers, most businesses aren't ranked low — they're absent. Here's why AI can't see them, what the newest Harvard research says to do about it, and the one move a rented platform structurally can't make.

Ten years ago a customer looking for a landscaper, a signage company, or an HVAC crew opened Google and scrolled a list of blue links. The whole SEO industry existed to move a business up that list. That customer still exists — but a fast-growing number of them now open ChatGPT, Gemini, Perplexity, or simply read the AI Overview that sits above Google's own results, and they ask a plain-language question: "Who's the best commercial sign company near me?" Then they act on the two or three names the model returns.

That is a different game with different rules, and most businesses are still playing the old one. The uncomfortable part isn't that they rank poorly in the new one. It's that they never enter it. The model gives an answer, confidently, and their name is nowhere in it.

Ranking was a list. Recommendation is a sentence.

Search ranking was forgiving. Even at position nine, you were on the page — a customer who scrolled far enough could find you. An AI answer has no page nine. It names two or three businesses and stops. You are either in the sentence or you do not exist to that customer, and there is no scrolling to change it.

This is why "we're on Google" is no longer the same as "we can be found." Being crawlable got you into the index. Being recommendable is a separate property, and it's the one that now decides whether a model puts you in the answer.

Why the model can't see you.

The instinct is to assume this is about being famous — that the models only name big brands. It isn't. It's mechanical, and two threads of recent Harvard Business Review research explain it cleanly.

The first, on how AI surfaces brands, found that models recommend what they can compare, verify, and connect to a specific customer problem. When your site describes what you do in vague, self-referential language — "quality craftsmanship, unmatched service" — there's nothing for the model to line up against a competitor and nothing to tie to the question the customer actually asked. You're not being rejected. You're being skipped, because you gave the machine nothing to reason with.

The second, on how large models read premium and specialized brands, found something subtler: models routinely misread businesses whose value lives in visual, spatial, or cultural cues rather than plain text. A signage firm's craft, a builder's community, a designer's taste — the things that make them worth choosing — are exactly the things that don't survive the trip into a language model unless someone deliberately translates them into structured, machine-parseable form. The value is real. It's just invisible at the layer where the recommendation gets made.

Put together: models recommend businesses that are legible to them. Most businesses are illegible by accident — not because they're small, but because no one ever wrote their presence for the reader that now matters.

Measuring the thing everyone is guessing at.

Almost every business is flying blind here. They don't know whether ChatGPT names them, whether Google's AI Overview cites them, or how they compare to the competitor down the road inside the answer itself. They feel the leads getting quieter and assume it's the market.

The first thing PRAGMA does is end the guessing. It probes the surfaces that actually make recommendations now — Google's AI Overview and Perplexity on the web-grounded side, and ChatGPT, Claude, and Gemini on the consumer side — with the real questions a customer would ask, and reports, per surface, a plain answer: did the AI name this business, and did it cite it as a source? Alongside it runs a small-model scan, because the on-device assistants shipping in phones and browsers are becoming their own quiet channel of recommendation. The output isn't a vanity score. It's a map of exactly where you're in the sentence and where you're absent.

That map is the free part. Anyone can be measured. The number on its own changes nothing.

The move a report can't make.

Here is where the old model of this work — the audit, the slide deck, the list of recommendations handed back for someone else to implement — quietly fails. Knowing you're invisible to AI is not the same as becoming visible to it. The gap between the finding and the fix is where every "AI visibility report" dies.

PRAGMA closes that gap because it doesn't stop at the diagnosis. When a scan finds that a page gives the model nothing to compare, PRAGMA drafts the rewrite — clear, benefit-anchored, tied to the questions customers actually ask. When it finds the value trapped in images and layout, it injects the structured data and schema that translate that value into something a model can read. When a competitor is being named and you aren't, it stages the specific content and local-signal changes that put you back in the running. Each fix is prepared, shown to you, approved with one tap — and shipped to your own site.

And it can do that for one structural reason the whole PRAGMA thesis rests on: the work runs inside a site you own. A rented marketing platform can post to its own template and answer a review, but it cannot restructure your pages, rewrite your content, and inject schema at the depth AI legibility requires — because it doesn't operate your site at that level, and it never will. The thing that makes a business recommendable is exactly the thing a rented platform can't reach.

This is not a trend to wait out.

The share of decisions that begin with a model instead of a search box is going up, not down, and the answer surfaces are consolidating — Google's own AI Overview, the assistants baked into phones and browsers, the chat tools people already trust. The businesses that get named in those answers over the next year will compound an advantage that's hard to reverse later, because models learn from what already looks authoritative. Being recommendable early is a moat. Being invisible while a competitor becomes the default answer is a slow, quiet loss you won't see in any report you're currently reading.

The customer is already asking the AI. The only open question is whose name it says back.

Frequently asked.