What does it mean to be recommended by AI?

Being recommended by AI means three things happen in the same answer. The assistant names your business. It offers your business to the person asking. And it does not warn them away. Miss one and you were not recommended. At Vivander Advisors I call that moment a recommendation event, and it is the only outcome I count as a win, because a business can be named ten times in AI answers and recommended zero times.

I learned that the hard way, reading answers.

A highlighted recommendation path leads a traveler to one hotel entrance.

The answer that looked like a win

Someone asks ChatGPT for a quiet place for an anniversary dinner near the Seaport.

ChatGPT answers with three restaurants. Your name is in the answer. The owner's dashboard lights up. Visible.

Read the sentence again. Your restaurant is in there as the landmark. "Two blocks past X, you will find Y." The diner is being steered to Y. You were furniture.

I have watched this happen to a Boston restaurant that appeared in most of the neighborhood prompts I ran. A vendor tool had told the owner it "appeared in 80 percent of AI answers." True. Every one of those appearances was directions to the place the assistant recommended.

There is no page two in an AI answer. Nor is there partial credit.

Here is the mechanic. An assistant can name you for reasons that have nothing to do with the question. A reference point for directions. Or a comparison foil: "if you want something quieter than X, try Y." Maybe a historical aside. A warning. In every one of those cases your name is in the answer and the customer is sent elsewhere.

So I classify every response on six values. Invisible. Named without being offered. Negative mention. Mentioned. Cited. Recommended. Only the last three count as visible. The two in the middle keep the fact that the assistant knows you exist while holding it out of the number. That is where the landmark and the foil belong.

What the research says about the gap

Recognition and discovery are two different abilities in these systems, and the distance between them is enormous.

A 2026 study by Sharma tested 112 startups. ChatGPT recognized 99.4 percent of them when named in the prompt. It surfaced 3.32 percent of them in unbranded discovery queries.

Ninety-nine percent known. Three percent offered.

That is why a branded prompt tells you nothing. Ask ChatGPT about your hotel by name and it will describe you fluently. Then ask it where to stay in your city and see whether you exist.

Rank does not rescue you either. Across 17 million AI citations Ahrefs analyzed, about 12 percent of the cited URLs sat in Google's top ten for the question. Being findable and being recommended have come apart.

Cited means the answer carries a link to your own domain. Stronger than a mention, because the customer reaches you in one click and because the link is evidence the assistant read your page rather than someone else's description of you. A citation without an offer still fails the test. I record both.

Recommended means the response puts you forward as an answer to the question asked. A single recommendation counts. So does a shortlist. Position nine on a list of twelve counts, and I record the position separately, because always last is a different problem from never present.

Endorsement strength is recorded and never enters the yes or no. "The obvious choice" and "worth considering" are both recommendations. A warning is not.

A traveler considers a destination after receiving a conversational recommendation.

Three things worth sitting with

For your business. Stop counting mentions. Ask how many times, out of the questions a real customer would type, you were offered as the answer.

For your role. Somebody has to read the whole answer, not the first line. That is not an SEO job or a PR job in most org charts. It sits between them, and right now nobody owns it.

For your customers. They asked one question and got three names. Nobody told them yours was in the fourth paragraph as a landmark. To them, you did not exist.

What I do not know

How often the "named, not offered" case occurs across a whole industry. I report the count per client, and one client is not a base rate. Anyone quoting an industry percentage for this is guessing, and I include myself in that.

The measurement I would start with

Run unbranded prompts. Read every answer to the end. Mark landmark, foil, offered, recommended. Then count the rows where you were named and a competitor was recommended in the same breath. I call those Ghost Rankings. They are the most actionable rows in any audit I have run, because you were in the room and lost.

The dashboard said you were visible. Meanwhile the diner drove past.

Investigate what buyers hear about your business

Compare Vivander’s AI visibility audit with ongoing visibility monitoring. Use the findings to prioritize work with your team and existing partners.