Key takeaways
- A mention, a citation and a qualified recommendation answer three different questions, and each needs its own count and its own denominator.
- Every combination of the three tells you something specific, and each points at different work.
- A single blended visibility score hides exactly the information you need to decide what to do next.
Most conversations about AI visibility start with one question: do we show up? It is a fair place to begin and a poor place to stop. A brand can show up in an answer in ways that range from decisive to worthless, and the word "visibility" covers all of them.
We use three words, and keep them apart on purpose: mention, citation and qualified recommendation. How an AI answer is assembled introduces them. This guide goes further. It sets out what each one actually asks, how each should be counted, what each combination of the three tells you and why folding them into a single score loses the thing you most need to know.
Three words, three questions
Each outcome is the answer to a different question about one AI answer.
A mention asks: does the answer connect you with this question? Your name appears in the text. It might be in a list, a comparison or a passing remark. A mention tells you the assistant associates you with the topic. It says nothing about whether you were put forward or why.
A citation asks: is the answer built on your evidence? A page on your own domain is used as a source. That means something you published was found, read and used to write the answer. It says nothing about whether the answer then recommended you. Your documentation can be used to explain the category while a rival is put forward.
A qualified recommendation asks: are you put forward for the buyer's stated need? The answer names you as the choice, or as a recommended alternative, for the specific situation the buyer described. This is the outcome that shapes a shortlist. It is also the hardest to earn and the one most easily confused with the other two.
The word "qualified" matters. An answer that says "popular tools include Pennant Metrics, Larchmont Data and your brand" has mentioned you. It has not recommended you for anything. An answer that says "for regulated reporting at a UK fintech, your brand is the strongest fit because it documents UK data residency" has.
What counts as recommended
Between "named" and "put forward" sit several distinct roles, and reading an answer properly means telling them apart. These are the roles we record for each brand in each answer.
| Role | What it looks like | Counts as a qualified recommendation |
|---|---|---|
| Primary recommendation | Put forward as the fit for the stated need | Yes |
| Recommended alternative | Put forward as a good option beside the main choice | Yes |
| Neutral shortlist | Listed among options with no preference expressed | No |
| Compared | Set against other vendors on features or price | No |
| Contextual mention | Named in passing, to describe the category | No |
| Discouraged or negative | Named, then advised against or criticised | No |
| Not present | Not named at all | No |
Two things follow. First, a neutral shortlist is not a recommendation, however good it feels to see your name in a list of three. Second, a discouraged mention is still a mention. A brand counting mentions alone can watch its number rise while the answers are steering buyers away.
Each one needs its own denominator
A rate is a count over a base, and the three outcomes have different bases. Getting the base right is most of what makes a figure honest.
Mention rate is the share of answers to your agreed questions that name you. The base is every answer that was read.
Citation rate is the share of answers that cite any source and cite your own domain among them. The base is answers with citations, not all answers. An answer that cites nothing cannot cite you, and counting it would make your citation rate look worse for reasons that have nothing to do with your evidence.
Qualified recommendation rate is the share of answers that put you forward as the choice or a recommended alternative. Because some questions matter more to the business than others, we weight each answer by the commercial weight agreed for its question. Being recommended for the question your sales team lives on should count for more than being recommended for a question nobody buys from.
Put the three side by side and they tell a story no single number can. A brand mentioned in most answers, cited in few and recommended in fewer has a name problem solved and an evidence problem open. A brand cited often but rarely named has the opposite problem.
Treating any one of the three as proof of the others is the most common mistake in this work. Each is reported with its own counts and its own base, and they are never added together.
What each combination tells you
Look at one priority question across a set of answers and your brand will usually sit in one of a few recognisable patterns. Each points at different work.
| Pattern | What it usually means | Where to look first |
|---|---|---|
| Not mentioned, not cited | No evidence connects you to this question | A page that answers the question in the buyer's words |
| Cited, not mentioned | Your page is read, but your name is not attached to the claim | Make the brand name explicit on the page and in its title |
| Mentioned, not cited, not recommended | Known in the category, no evidence used | Specific, findable pages for the constraints the question names |
| Mentioned and cited, not recommended | Your evidence is used, but a rival's fits the need better | Specific fit claims and independent support for them |
| Recommended, not cited | Put forward on someone else's evidence, or on general knowledge | Your own page for the claim, so the recommendation has a firmer base |
| Recommended and cited | Put forward, on your own evidence | Independent corroboration, and checking the role holds across repeated runs |
The pattern easiest to overlook is the second: cited but not mentioned. It happens when an answer uses your page as a source for a general point and never names you. It is one of the more fixable positions to be in, because the hard part, being found and read, has already happened. The work is to make sure the page says clearly who is making the claim.
The pattern easiest to misread is the fifth: recommended but not cited. It looks like success, and it is a good outcome. But a recommendation resting on someone else's description of you is fragile. If that source changes, or a rival's own documentation becomes clearer, the recommendation can go with it.
Why a blended score hides the work
It is tempting to combine the three into one visibility score. It fits on a slide and moves in one direction. It also destroys the information you need.
Consider a worked example with two fictional vendors, each observed on the same twenty answers to the same agreed questions.
| Pennant Metrics | Larchmont Data | |
|---|---|---|
| Mentioned | 16 of 20 answers | 10 of 20 answers |
| Cited (of 15 answers with citations) | 3 | 11 |
| Qualified recommendation | 3 of 20 answers | 8 of 20 answers |
Add these up with any reasonable weighting and the two land close together. Yet they are in completely different positions. Pennant Metrics is widely known and rarely put forward: its name travels further than its evidence. Larchmont Data is named less often, but when it is named it is usually on the strength of its own documentation, and it is put forward far more often. Pennant's next move is evidence. Larchmont's is reach.
A single score would tell both teams the same thing: you are somewhere in the middle. It would also let a rise in mentions mask a fall in recommendations, which is exactly the movement that matters most. That is why we report each outcome separately, with its counts beside it, and never combine them.
Putting the three to work
You can start keeping the three apart this week, with nothing more than a spreadsheet and some patience.
- Pick five questions that matter commercially, written the way a buyer would ask them.
- Read several answers to each. For every answer, record three things for your brand and your two closest rivals: named or not, your domain cited or not, and the role each brand played.
- Count each outcome over its own base. Mentions over all answers. Citations over answers that cite anything. Qualified recommendations over all answers, weighted if some questions matter more.
- Find your pattern for each question in the table above, and write down the one piece of work it points at.
- Report the three side by side. Resist the urge to add them up. If a board needs one line, make it the qualified recommendation count, with the other two beneath it.
Do this on the same questions each month and you will see something a blended score never shows: which of the three is moving, and whether the movement is the kind that wins shortlists.
The AI visibility checklist
A three page working checklist: what you can change in how AI answers describe you, the method to run in order, and a sheet to record what you find.
Free with your work email
Request the download