Being mentioned, being cited, and being recommended are three different outcomes with three different fixes. Most tools report the first and let buyers assume the third.
They also come from different machinery. A mention can come from training-data memory with no retrieval at all, while a citation requires a live fetch of a specific page, so the same brand can score well on one and zero on the other.
The three get collapsed into “visibility” because they look alike in a dashboard. They do not behave alike, and a fix aimed at the wrong one produces months of work and no movement.
What are the three outcomes, exactly?
Each one is a different event with a different mechanism behind it and a different remedy in front of it.
| Outcome | What the engine did | What it proves | The fix when it is missing |
|---|---|---|---|
| Mentioned | Named your brand in prose, no link attached. | Your brand exists as a resolved entity in the model for this category. | Entity work. Get named on the sources the model learned the category from, and make the name unambiguous. |
| Cited | Attached a specific page of yours as a source. | Your content was retrieved, read, and judged usable for this question. | Content and access work. A page that answers the question, reachable by the retrieval crawler. |
| Recommended | Named you as the answer rather than as context. | The model ranked you above the alternatives it also knows. | Positioning work. Be the documented best fit for a stated constraint, not a generic option. |
Read the fix column from the bottom up and the sequencing becomes obvious. You cannot be recommended by an engine that does not know you exist, and you cannot be cited for a page that does not answer anything.
Why does a mention come from memory and a citation from retrieval?
Because the two are served by different systems, and the vendors document them separately. OpenAI’s crawler documentation separates them at the infrastructure level: “OAI-SearchBot is used to surface websites in search results in ChatGPT’s search features,” while GPTBot “is used to crawl content that may be used in training our generative AI foundation models.” Two crawlers, two purposes, and blocking one does not affect the other.
A mention can therefore arrive with no live fetch at all. The model is recalling that your brand belongs in this category, which is a fact it absorbed during training. A citation cannot arrive that way, because a citation names a URL and a URL has to be fetched.
Anthropic draws the line on the answer side. Its citations documentation describes the feature as returning the exact passages that support each claim, so that answers can be verified against sources. That is a stricter event than a name appearing in a sentence.
How do you measure each one?
Mention rate and recommendation rate need the answer text. Citation rate is the only one of the three that a platform will hand you directly, and only on one surface.
| Rate | Definition | Where the data comes from |
|---|---|---|
| Mention rate | Answers naming you, over answers in the set. | Sampling. Read every answer, count the names. |
| Citation rate | Answers attaching a page of yours, over answers in the set. | Sampling, plus first-party data on Microsoft surfaces. |
| Recommendation rate | Answers naming you as the choice, over answers in the set. | Sampling with a coding rule, because it needs a judgement about the sentence. |
The first-party exception is worth using. Microsoft’s AI Performance report in Bing Webmaster Tools shows “the total number of citations that are displayed as sources in AI-generated answers,” which is a census rather than a sample. It covers Microsoft surfaces only, and it counts citations specifically, so a Copilot answer that names you without linking you never appears in it. The mechanics of that surface are covered in how Copilot inherits Bing’s index.
Recommendation rate is the one that needs a written coding rule, because “named as the choice” is a judgement. Decide in advance whether a brand listed first in a neutral list counts, and apply that rule identically every period, or your trend line is measuring your own drift.
What actually fixes each one?
- Symptom
- Asked for options in your category, no engine names you.
- Work
- Get onto the sources that define the category. Third-party listings, comparison pages, and a consistent, resolvable brand name.
- Symptom
- Your name appears; the links go to competitors and review sites.
- Work
- Publish the page that answers the question, and confirm the retrieval crawler can reach it. This is access and content, not brand.
- Symptom
- You are the source for the explanation and somebody else is the answer.
- Work
- Positioning. Be documented as the best fit for a specific constraint rather than a capable general option.
Case two is where most B2B teams actually sit, and the access half of it is easy to get wrong invisibly. Google’s documentation on AI features sets the bar plainly for its own surfaces: a page “must be indexed and eligible to be shown in Google Search with a snippet,” and there are “no additional requirements to appear in AI Overviews or AI Mode.” No special markup, no separate file, just retrievability. The two Google surfaces still differ from each other in what they show, which is covered in AI Overviews vs AI Mode.
Can the three rates move in opposite directions?
Routinely, and the combination is more diagnostic than any single number. A rising mention rate with a falling citation rate usually means the category is discussing you more while your own pages are being retrieved less, which points at access or freshness rather than awareness.
The reverse pattern, citations up and mentions flat, is what a successful content programme looks like in its first quarter. You are being read before you are being remembered, because retrieval responds in days and training memory responds in model generations.
The pattern worth acting on fastest is recommendation rate falling while the other two hold. Nothing about your visibility broke. Somebody else became a better documented answer to the same question, and the answer text will usually name them.
Why do most tools report only the first?
Because mentions are the cheapest to detect. Finding a brand name in text is string matching. Deciding whether that sentence recommended you requires reading it, and reading every answer is the expensive part of this category.
A tool that reports one blended visibility score has decided, on your behalf, that the difference between being compared against and being chosen does not matter.
The practical consequence is that two brands with identical dashboard scores can need opposite work, and the dashboard cannot tell you which. Separating the two is the same diagnostic split as entity problem vs content problem AI visibility, and the reading discipline that keeps the three rates honest is in how to interpret an AI visibility report.
Report all three, per engine, with the run count attached. If you only have budget to track one properly, track recommendation rate on unbranded prompts, because it is the only one of the three with revenue on the other side of it.