A mention is your name appearing in an AI answer. A citation is your name appearing as the source the answer was built from, usually with a link. A recommendation is the answer telling the buyer to choose you. They are three different outcomes with three different mechanisms behind them, and treating them as one thing called visibility is how companies end up optimising for the one that matters least to them.
The reason to separate them is that they do not come in order and they are not degrees of the same achievement. You can be cited constantly and never recommended. You can be recommended without being cited at all, because the model learned about you somewhere you cannot see. And the cheapest of the three to earn turns out to have a measurable commercial effect, which is not what most people assume.
The three, separated properly
| Outcome | What the answer does | What earns it | What it is worth |
|---|---|---|---|
| Mention | Names you, no link, no endorsement | Presence in the corpus and topical relevance | More than people think, and measurable |
| Citation | Names you as the source of a claim, usually linked | A passage that answers something standalone | Traffic, plus authority by association |
| Recommendation | Tells the buyer you are the right choice | Third-party corroboration and comparative standing | Most, and it is the hardest to influence |
What each one looks like in a real answer
The difference is easier to recognise than to define, so here is the same fictional company appearing three ways in response to a buyer asking about contract management software.
Options in this space include Ironclad, Lexion, Concord and Clausewise, with pricing models varying between per-seat and per-contract.
A mention. Named, unlinked, unranked, in a list
You are in the consideration set and nothing more. No judgement has been made about you and no traffic will arrive. This is still the outcome most companies are failing to achieve, and per the study above it moves buyer behaviour.
Mid-market legal teams typically hit a ceiling on manual review at around 500 contracts a month, after which error rates rise faster than headcount can absorb (Clausewise).
A citation. You are the evidence the claim rests on
Notice what earned this. Not a claim about how good the product is, but a specific, checkable number that answered a question standalone. The engine needed a threshold and you were the source that supplied one. This is the outcome you can engineer deliberately, and it is why original data outperforms positioning copy.
For a mid-market legal team of your size, Clausewise is probably the best fit, mainly because its review workflow is built for teams without a dedicated legal ops function.
A recommendation. A judgement made on your behalf
Nothing on your own website produced that sentence. It came from the pattern across reviews, comparisons and discussion that consistently associates you with one specific buyer type. Which is why recommendation work looks less like content marketing and more like making sure the market describes you the way you would describe yourself.
One more pattern worth naming, because it is common and it is not on the ladder at all. You get mentioned, and the description attached to your name is wrong. Old pricing, a feature you removed, a market you left. That is a fourth outcome, it is worse than absence in a buying conversation, and it needs correcting rather than amplifying.
The mention is undervalued, and there is now evidence
A mention sends no traffic. Nothing appears in your analytics. For most marketing teams that makes it invisible, and invisible tends to mean worthless.
A study published in June 2026 by Michael Iannelli and Alan Ai measured what actually happens afterwards. They joined opt-in clickstream data to the same users’ conversations with ChatGPT, Claude and Gemini, then looked at behaviour off the platform. When an assistant recommended a brand to someone with no recent engagement with it, Google searches for that brand rose by 4.3 percentage points, visits to the brand’s own site rose by 2.4 points, and visits to brand-specific retailer pages rose by 1 point.
Two caveats belong with those numbers. The study is observational rather than a controlled experiment, and it did not observe transactions, so the authors describe the findings as purchase-adjacent. It also covers consumer brands rather than B2B software procurement.
The mechanism is what transfers. Somebody reads your name, does not click, and searches for you later. Your analytics records a branded search or a direct visit and credits it to nothing in particular. The authors put this plainly, noting that standard referrer-based and last-click measurement miss the upstream exposure entirely.
This is why branded search volume is the most useful proxy you already have. It is imperfect and it is free. If mentions are rising and you have no other explanation, branded search tends to move first, weeks before anything shows up in a channel report.
Branded search is the proxy. The three counts are the measurement.
Watching branded search volume costs nothing and tells you something real, and it will not tell you which of the three outcomes you are getting or which competitor is getting them instead.
If you want that split out properly across four engines, it is the first thing our visibility check reports, and you keep the raw answers.
The citation is the one you can actually engineer
Of the three, this is where deliberate work has the most direct effect, because a citation is awarded to a passage rather than to a company. The engine needed a fact, your paragraph supplied it cleanly, and it credited the source.
Which means the requirements are specific rather than reputational. A passage that answers a question without depending on the paragraph above it. A number in the same sentence as the claim it supports. A named framework or a piece of original data that has no equivalent elsewhere, so there is a reason to attribute rather than absorb.
The uncomfortable part is what a citation is worth in traffic. Research tracking nearly 69,000 Google searches found that when an AI summary appeared, users clicked a link cited inside it on just 1 percent of visits. So a citation is worth having for the authority and the occasional click, and if you are buying it expecting a traffic channel you will be disappointed.
The wider exchange runs the same way. Network-level data on how often AI platforms fetch pages relative to the visitors they send back has produced ratios in the tens of thousands to one for some platforms, though the published figures overstate the real position because app traffic arrives without a referrer. Being cited is not a fair trade. It is still better than being absent.
The recommendation is the prize, and it is barely yours to control
A recommendation is a judgement. The system is not reporting that you exist or that you said something useful. It is telling a buyer that you are the answer, which is the closest thing in this field to a sales call you did not have to make.
It also rests almost entirely on what other people have written about you. A model asked to recommend weighs corroboration, and your own site claiming you are the leading option in your category contributes close to nothing to that calculation. Reviews, independent comparisons, analyst coverage and community discussion do the work. This is the slowest of the three to move and the one no configuration change touches.
There is a question underneath this that nobody has settled. When a system recommends a product, is that an endorsement? The FTC’s endorsement guides were written for humans and organisations making claims, and they turn on whether a reader would understand a statement to reflect an independent opinion. An AI recommendation reads exactly like an independent opinion to most people. Whether that framing eventually attaches obligations to anybody is an open question, and if you are in a regulated category it is worth watching rather than assuming.
Why the distinction changes what you do on Monday
Decide which of the three you are actually short of, because the work diverges immediately.
If you are not being mentioned at all, you have a presence problem, and it is almost certainly technical rather than editorial. Crawler access, rendering and whether your name resolves to a distinct company. No amount of writing fixes a page nobody is permitted to fetch.
If you are mentioned but never cited, you have an extractability problem. Your content is reachable and understood and no single passage stands on its own well enough to be quoted. That is editing work, and it is narrow and quick.
If you are cited but never recommended, you have a corroboration problem, and it is the expensive one. You are useful as evidence and not yet credible as a choice. That takes quarters, and it takes work outside your own domain.
Which of the three are you short of
Answering that properly means running your category questions across several engines and classifying every result as a mention, a citation or a recommendation. Our free AI visibility check does the classification and reports the three counts separately, which is the part that tells you whether to call an engineer or an editor.
Run the same category question in four engines and write down which of the three you got. Most companies have never done this deliberately, and it takes about ten minutes to find out which conversation you should be having.
Record it as three separate counts rather than one score. Out of ten category questions, how many mentioned you, how many cited you, and how many recommended you. Three numbers with the same denominator, which anybody can recount and disagree with. A single composite visibility figure hides exactly the information you need, because it cannot tell you whether your problem is technical, editorial or reputational, and those three have nothing in common except the symptom.
Then run it again a fortnight later. The gap between the two runs tells you how much of what you saw the first time was real.
Three outcomes, three different fixes, and the only way to know which one you are short of is to run your own category questions and classify what comes back.
We publish our own measurement work as we run it, including the runs that contradict what we expected. The newsletter is where it goes out first.
Related reading
- The AEO glossary. Every term defined, each marked by whether it comes from a real standard or from marketing.
Shaban runs answer engine optimisation for B2B SaaS companies at Uncited Brands. He works the unglamorous end of the problem, mostly crawler access, retrieval diagnostics and measurement baselines, and publishes the tests that do not go his way alongside the ones that do.