Cited Is Not Chosen
Most brands measuring AI visibility are measuring citations. Citations prove a model could read you. However, they do not prove it recommends you, and the gap between the two is where deals are lost.
Here is a scenario that should worry anyone reporting on AI visibility.
Your citation count is climbing. Your pages are being pulled into answers across ChatGPT, Gemini and AI Overviews. Every number on the dashboard is moving the right way.
Then a customer asks which one they should actually buy, and the model names someone else while using your content to explain why.
That is not a measurement error. It is a measurement gap. Being cited and being chosen are two different outcomes, and most AI visibility programmes only track the first one.
Three outcomes, routinely reported as one
| Outcome | What it means | What it proves |
|---|---|---|
| Cited | Your URL appears as a source beneath or beside the answer | A model could find, read and use your page |
| Mentioned | Your brand name appears in the answer text | The model considers you part of the category |
| Recommended | You are named as the thing to do or buy | The model trusts you enough to put its answer behind you |
These are not three steps on a ladder. A model can cite you without naming you, name you without backing you, and back you without linking to you at all.
Getting cited means a model could reach your page, read it, and find something worth using. Getting mentioned means it thinks your brand is worth naming. Getting recommended means it will stand behind you, and nothing you publish about yourself can settle that, because you are not a neutral source.
Most GEO work chases the first and assumes the other two follow. They do not.
The gap is not small
SEMrush, working with Kevin Indig, analysed thousands of domain appearances across multiple AI engines and found that 62% of AI citations are what Indig calls ghost citations: your site is used as a source, but the answer never says your name. The user reads the response, absorbs the information, and walks away with no idea it came from you.
For a publisher, that is an attribution problem. For a consumer or B2B brand, it is worse than invisibility. You supplied the research that helped someone choose a competitor.
The same study found the gap runs in opposite directions depending on the engine. Of the brands that appear at all, Gemini names 83.7% in the answer text but cites only 21.4% as a source. ChatGPT does close to the reverse: an 87% citation rate against a 20.7% mention rate.
If your reporting is citation-based, you will look strong in ChatGPT and weak in Gemini. If it is mention-based, you will look strong in Gemini and weak in ChatGPT. Neither picture is wrong. Both are partial. And a team that reports only one number is unknowingly choosing which engine it appears to be winning.
The self-citation trap
The worst version of this gap is the one brands build for themselves.
A recent analysis of B2B software queries found that Google's AI Overviews frequently cited brands' own "best of" listicles, the ones ranking their own product first, and then recommended a competitor anyway in 69% of those cases.
The mechanism is not mysterious. The model treated the page as a legitimate source of category information while discounting its ranking as self-interested. You wrote the comparison, the model read the comparison, and the model disagreed with your conclusion using your own data.
Publishing more of that content raises citations and does nothing for recommendations. It may be the single most common way brands manufacture the exact gap they are trying to close.
Most of your citations were never yours anyway
A recent BuzzStream study found that about 80% of citations sat on third-party sites rather than the brand's own domain.
That is the uncomfortable structural point. The surfaces that decide whether you are recommended are mostly not surfaces you publish on: review sites, forums, trade press, comparison pages, community threads. You can influence them, but you cannot write them.
Which means closing the citation-to-recommendation gap is not a content calendar problem. It is a PR, community, review and partnership problem, funded from a content budget, owned by nobody in particular. That is why so few teams are making progress on it.
And it varies by market more than anyone expects
The SEMrush study also broke its numbers down by country. Brands are named in roughly 50% of AI answers in India and Sweden, but only 18% to 22% in Italy, Brazil and the Netherlands.
Same category. Same brands. Wildly different odds of being named rather than quietly used.
If you run a global brand and report a single blended AI visibility figure, you are averaging across markets where the underlying mechanics differ by a factor of two or more. The average is describing a situation that exists in none of your markets.
What to actually do about it
Separate the three metrics before your next report. Cited, mentioned, recommended. If your dashboard collapses them into one visibility score, it cannot tell you which problem you have, and the three problems have completely different fixes.
Diagnose which gap you're in. Cited but not mentioned is a ghost citation problem: you are feeding the category and getting no credit. Mentioned but not recommended is a trust problem: the model knows you and will not back you. Neither is a discoverability problem, and neither is solved by publishing more.
Stop counting your own comparison content as progress. Track what proportion of your citations come from your own domain. If it is high, your citation growth is likely to be inflating a number that recommendations do not follow.
Move budget toward earned surfaces. Third-party validation is what converts a citation into a recommendation. That is review platforms, independent comparisons, trade coverage and community presence. These are the places where the case for you is made by someone who does not work for you.
Break the numbers out by engine and by market. A single global figure averages Gemini against ChatGPT, which run in opposite directions, and India against Italy, where mention rates differ by more than double. One engine or one market can collapse without the average moving.
The bottom line
Citations tell you a model can read you. They do not tell you it whether it backs you.
A brand can hold high citation share across every engine, appear in thousands of answers, and still be left off the shortlist every time. The citations came from its own pages. The recommendations went to brands other people vouched for.
That is the whole gap. You can publish your way to being cited. You cannot publish your way to being chosen.