When AI answers a question, your brand can show up two ways. As a citation, linked as the source the model drew from. Or as a mention, named in the answer with no link at all. They look the same on the page. They do different work on the model. Most marketing teams track one, ignore the other, and misread their whole position in AI search because of it.
Ask ChatGPT, Gemini or Perplexity for the best GEO agencies in South Africa, then read the answer slowly. Some brands are linked. There is a source pill or a footnote, and you can click it. Other brands are named in the prose with nothing to click at all. Both sit in the answer. Only one sends you anywhere.
That difference is not cosmetic. A citation and a mention are two different signals, built by different work and worth different things to your business. Part 1 covered how AI builds an entity profile of your brand. Part 2 covered the two pools it draws on. This part is about what lands in the answer, and why winning AI search means winning both forms of it, not just the one your dashboard counts.
A citation is a link. A mention is a name.
Let us be precise, because the whole argument turns on the definitions.
A citation is when the model attributes part of its answer to you and links to you as the source. You are the footnote. The reader can click through. In Perplexity it is the numbered source. In ChatGPT it is the cited link. In Google's AI Overviews it is the source card. A citation says: this came from this brand, and here is where to check it.
A mention is when the model names you in the answer without citing you as the source. Agencies worth considering include X, Y and Z. Your brand is in the recommendation, woven into the prose, but it is not the source the answer leans on. There may even be a link right beside your name, so the real test is not whether there is something to click. It is where the click goes.
That is the part most teams get wrong. The test is not whether there is a link. It is whose link it is. When your name appears, look at the source the model attaches to that line. If it points to you, you are cited: you are the evidence the answer is built on. If it points to a directory, a review site or a competitor's round-up, you are only mentioned, and that third party is the one being cited as proof. If there is no source at all, the model is naming you from memory. Same brand on the same page, three very different positions.
One more thing on that screen fools almost everyone, and it is probably what you are looking at. Sometimes the model attaches a business card to your name: your website, a call button, directions. That is not a citation either. It is an entity card, the model surfacing your structured business listing because it recognises you as a real, known business, the entity model from Part 1 made visible on the page. It is a mention, and a strong one, because the model is routing the customer straight to you. You can have that card and still not be the cited source behind a single claim in the answer. The recommendation, the citation and the business card are three separate wins, earned by three separate kinds of work.
A citation is the model telling you where it looked. A mention is the model telling you what it thinks. You want to be in both sentences, because the customer reads both, and only one of them is a recommendation.
Both move the entity model. They just move different parts of it.
Most coverage treats citations as the prize and mentions as a consolation. That is wrong. They build different parts of the same entity, and you need both.
01. A citation builds evidence
When the model cites you, it is treating you as a source worth attributing. Three things follow. You earn a click, the only direct referral traffic AI search hands out. You teach that engine your domain is one it can trust on the topic, which makes the next citation likelier. And you put fresh, dated proof into the live pool, the corroboration that keeps your entity current between training runs. A citation is how you prove the pattern is real, this week, on the record.
02. A mention builds association
When the model names you without a link, it is placing you. You belong in this category, next to these competitors, in answer to this kind of question. That is co-occurrence, the mechanism Part 1 named as the engine behind the entity model. A mention is proof the model already associates you with the topic strongly enough to recall you unprompted. And commercially it is often the bigger prize, because being in the recommendation is what wins the customer, link or no link.
03. One without the other is a half-signal
This is the cost in the headline. Cited but rarely mentioned, and you are useful, not chosen: the model quotes your market data in an answer that goes on to recommend your competitors. You are the reference desk for someone else's sale. Mentioned but rarely cited, and your position is fragile: you are recalled from memory with nothing fresh underneath, so the day a better-evidenced rival turns up in retrieval, you are dropped. Evidence without association makes you a source. Association without evidence makes you a rumour. You want both, so the model names you and backs you in the same breath.
How much the original GEO research found that adding citations, quotations and statistics lifted a source's visibility in AI answers.
From the 2023 paper that named the field, 'GEO: Generative Engine Optimization' (Aggarwal et al., Princeton University). Citing sources, adding quotations and adding statistics were the strongest tactics, each lifting a source's visibility by roughly 30 to 40 percent against the baseline. The citation-worthy moves are the measurable ones.
The mistake of measuring only one
Most AI-visibility tracking counts citations, because citations are easy to count. They are links. They behave like the backlinks and referral traffic teams have watched for twenty years, so they drop neatly into the dashboard you already have. Mentions are harder. They are unlinked strings of text in the middle of prose, so you have to read the whole answer and spot your brand by name, not by URL.
So teams manage what they can see. They watch a citation count climb and miss that they are being mentioned less, or mentioned and then argued out of the recommendation. Or they see the brand named in a few answers, call AI visibility solved, and never notice that not one of those answers links to them, which means nothing fresh is holding the position up.
Both reads are half-blind. An honest measure of AI visibility captures all of it: is the brand named at all, is it linked, and in what role, recommended, warned against, or used only as a source. Track one number and you will manage the wrong half of your presence with total confidence.
AI citations Algorithm measured across ChatGPT, Perplexity and Claude in the South African market.
Every one was a link the engine chose to show. The mentions sitting alongside them, the unlinked recommendations, are a separate count and a separate job. Source: Algorithm Lighthouse, 2026.
The full per-engine breakdown is in our research piece on the 46,315 citations we measured locally.
The South African angle
The local market sharpens this. Part 1 made the point: South Africa is a thin-signal web. Outside financial services the model has few sources to cite for most categories, so citations are there for the taking. Publish the quotable stat, the clear definition, the piece of original research, and you can become the cited source for a whole topic far faster here than in a crowded market.
The thin web cuts the other way on mentions. Less third-party chatter means the model's recalled associations are weaker, and weaker associations move more easily. The brands building both right now, the cited proof and the named association, are compounding a lead in categories that are still open. The ones doing neither are not holding steady. They are absent, and absence is what the model defaults to when it has nothing to recall and nothing to cite.
In a thin-signal market, citations are easier to win and mentions are easier to lose. That is the South African GEO window in a sentence, and it closes as the local web fills up.
What this means for your brand: two jobs, run together
A citation and a mention are different signals, so they need different work. Run both, or accept a known hole in your position.
To earn citations, become quotable. The research is unusually clear on this. Publish what models attribute: original data, named statistics, plain one-line definitions, direct expert quotes, comparison tables, claims someone can source. Put it on pages an engine can crawl, parse and trust, structured so the model's job is easy. Done well, this is the part of GEO that measurably lifts how often you are cited. It is also the part you control most directly, because it lives on surfaces you own or place.
To earn mentions, become associated. This is the off-page half, and you cannot do it on your own site alone. Get named next to your category and your competitors across the third-party web the model reads: editorial in the publications each engine trusts, a place in the best-of and comparison pieces, expert positioning, partnerships, the same description of you wherever you turn up. Mentions are earned by being described the same way, by enough independent sources, that the model recalls you without going to look. You earn that across everyone else's pages, not your own.
The brands winning AI search in 2026 run these as one programme. The quotable assets that earn citations are the same assets third parties pick up and reference, which earns mentions, which strengthens the entity, which makes the next citation land harder. Together they compound. Apart, each one caps the other.
How Algorithm thinks about this
We measure both, separately, for every client we run through Lighthouse GEO, because the gap between them is usually the brief. Cited but rarely mentioned is an authority problem dressed as a visibility win: useful to the model, not yet chosen by it. Mentioned but rarely cited is a fragile position the next well-evidenced competitor can take. Which gap you are looking at decides the work.
That is the difference between counting links and reading your position. A citation count climbing on its own can hide a brand sliding out of the recommendation. A handful of mentions with nothing cited underneath can look like presence and be a draught. Read both, per engine, in context, and you get the real picture: not whether AI is talking about you, but whether it is backing you.
The next step
Part 4 turns this into a build. If citations and mentions are the two things AI does with your brand, the next part lays out the strategy for winning both: what AI already knows about you from training, what it can find about you in real time, and how to work the two at once. Sign up for the series to get it the day it lands, or book a Lighthouse visibility audit to see how often each engine cites you, how often it only mentions you, and which of the two is quietly costing you.
