How AI Decides Which Brands It Cites — and How to Become One of Them

For twenty years, the goal of brand visibility was simple to state: rank #1 on Google. Get the blue link at the top, win the click. Everything in SEO — backlinks, keywords, page speed — was in service of that one position.

That goal is quietly dissolving. When someone asks ChatGPT "who are the best project-management tools for agencies?" or types a question into Google and gets an AI Overview, there often is no click. The answer arrives pre-assembled, with two or three brands named inside it and a handful of sources cited underneath. The user reads the answer and moves on. Nobody visits your homepage. Nobody sees your carefully optimised landing page.

The new question isn't how do I rank #1? It's how do I become the source the AI quotes? That's a different problem with different mechanics, and most of what worked for classic SEO only partly applies. This piece explains how large language models actually decide which brands and facts to surface, where they get those facts, and — honestly — what you can and can't do about it.

We sell Wikipedia and structured-data work, so we have an obvious interest here. We've tried to write this so it's useful even if you never hire anyone. Several sections below will tell you where our services don't help.

The shift: from ranking to being quoted

"Zero-click" isn't new — Google's featured snippets and Knowledge Panels were eating clicks years before ChatGPT existed. But generative answers accelerate it dramatically. Instead of a snippet pulled verbatim from one page, you get a synthesised paragraph that blends several sources, names specific entities, and rarely needs the user to leave.

This changes what "visibility" means in three concrete ways.

First, the unit of visibility is the entity, not the page. Google ranked URLs. An LLM reasons about things — companies, people, products, concepts — and the facts attached to them. If the model doesn't have a clear, consistent understanding that your company exists and what it does, no amount of on-page optimisation will get you named.

Second, citation is probabilistic, not deterministic. You cannot guarantee a model mentions you for a given query the way you could (roughly) target a keyword. The same prompt can yield different brands on different days, across different models, even at different "temperature" settings. The realistic goal is to raise the probability that you're surfaced accurately — not to lock in a slot.

Third, the work happens upstream of the answer. You're not optimising the output; you can't touch it. You're shaping the source material the model was trained on or retrieves from. That's a slower, more indirect lever than buying an ad or tweaking a title tag — and it's the entire game.

This emerging discipline has a few names — Answer Engine Optimisation (AEO), Generative Engine Optimisation (GEO), or just "AI visibility." The labels matter less than the underlying shift: you're optimising to be referenced, not clicked. Our AI visibility work is built entirely around that distinction.

Where LLMs actually get their facts

To influence what an AI says about you, you have to know where it's pulling from. There are three distinct mechanisms, and they behave very differently.

1. The training corpus. This is the giant snapshot of text the model learned from — a large crawl of the public web, books, and licensed datasets, frozen at some cutoff date. Facts baked in here are "remembered" by the model itself. They're powerful because the model treats them as background knowledge, but they're slow to change: if your company rebrands or pivots, the training corpus won't know until the next model is trained. Training data also skews toward sources that are large, heavily linked, and frequently duplicated across the web — which is a big reason encyclopedic and reference sites punch above their weight.

2. Live retrieval (RAG). Retrieval-Augmented Generation means the system runs a search at query time, pulls a few fresh documents, and feeds them to the model as context before it answers. This is how a tool can tell you something that happened last week despite a year-old training cutoff. Perplexity is built around this; ChatGPT and Gemini do it when they decide a query needs current information. RAG is where fresh, well-structured, easily-retrievable content matters most — because the system is actively going to look for sources in the moment.

3. Grounding indexes. Some systems are wired directly into a structured knowledge layer — Google's models can lean on the Knowledge Graph; many tools cross-check facts against Wikidata or similar entity databases. Grounding is how a model resolves "which 'Apple' do you mean?" and attaches a stable identity to an entity. It's less about prose and more about machine-readable facts: founding date, headquarters, industry, key people, official identifiers.

Most real answers are a blend. A model might recall your industry from training, retrieve a recent funding announcement via RAG, and ground your company's identity against a knowledge base — all in one response. The practical takeaway: you need to show up in all three layers, because you never know which one a given answer will lean on.

The four engines, compared

The major answer engines don't cite the same things. They have different architectures, different source preferences, and different appetites for live retrieval. Analyses published through 2026 paint a rough but consistent picture — directional, not precise, and shifting month to month as these products change fast.

A few honest caveats about that table. The percentages floating around the industry vary widely between studies because methodology differs — what counts as a "citation," which queries were sampled, which country. Treat any single number as a rough order of magnitude. What's durable across the studies is the relative pattern: ChatGPT is unusually Wikipedia-heavy; Google's AI surfaces lean on community platforms; Perplexity exposes and favours retrievable discussion. That pattern is what you plan around.

One number does keep recurring strongly enough to anchor on: analyses in 2026 consistently find Wikipedia is the single most-cited domain in ChatGPT's answers — in some studies roughly half of its top factual citations trace back to Wikipedia. Reddit is repeatedly the next tier, often cited as something like 10–12% of ChatGPT's US citations. Even allowing for measurement noise, the message is unambiguous: encyclopedic sources dominate, and community sources are the strong second act.

Why Wikipedia and Wikidata are over-represented

If you only fix one thing in your AI-visibility stack, it's almost always the encyclopedic layer. There are four structural reasons LLMs over-rely on Wikipedia and its sister project Wikidata — and none of them are accidental.

Neutrality. Wikipedia's house style is deliberately non-promotional, attributed, and balanced. That's exactly the tone a model wants to reproduce when it's trying to sound factual rather than salesy. Training on neutral prose teaches the model to speak neutrally, so neutral sources get reinforced.

Structure. Articles follow a predictable shape: a definitional first sentence, an infobox of key facts, sectioned body, references. That regularity makes Wikipedia unusually easy for a model to parse and for a retrieval system to extract clean facts from. Messy, idiosyncratic content is harder to mine reliably.

Open license. Wikipedia's content is freely licensed for reuse. That removes legal friction from including it in training sets and reproducing it — so it gets included, broadly and repeatedly. Duplication across the web amplifies its weight in the corpus.

Entity IDs. This is the quiet superpower. Wikidata assigns every entity a stable identifier (a "Q-number") and machine-readable statements — this company, founded this year, in this industry, led by this person. That's the connective tissue grounding systems use to know who you are and to disambiguate you from everyone with a similar name. A Wikipedia article gives the model prose; the linked Wikidata item gives it structured truth. Together they're the closest thing to an "official record" the open web has.

This is why a Wikipedia presence does double duty: it's a heavily-weighted training source and it usually creates or strengthens the Wikidata entity that grounding systems rely on. If you want to understand the structured-data half specifically, we wrote that up in Wikidata and the knowledge graph. And the honest prerequisite — covered in our Wikipedia page creation work — is that none of this is available to you unless your organisation genuinely meets Wikipedia's notability bar. No notability, no article, no shortcut. That's a feature of the system, and it's the same reason the citations are trustworthy in the first place.

The secondary sources: Reddit, Quora, YouTube, LinkedIn

Encyclopedic coverage is the foundation, but it's not the whole picture — and for some engines it's not even the dominant one. The community layer is where a different kind of signal lives: not "here are the verified facts about this entity," but "here is what real people say when they discuss it."

Reddit is the standout. It shows up heavily across ChatGPT, Google AI Overviews, and Perplexity. The reason is that Reddit threads contain exactly what a model needs for opinion-shaped and recommendation-shaped questions — candid, specific, comparison-rich discussion ("we switched from X to Y because…"). When someone asks an AI for recommendations rather than facts, community discussion is disproportionately influential. Our Reddit AI visibility work is about earning a genuine, non-spammy presence in the threads that matter to your category.

Quora appears prominently in Google's AI surfaces in particular, for the same reason: it's structured question-and-answer content that maps cleanly onto the kinds of questions users actually ask an answer engine. A well-answered question that ranks can become source material. We cover the specifics in Quora AI visibility.

YouTube is increasingly cited, especially by Google (unsurprisingly — same parent company). Transcripts are searchable text, and how-to or review content answers a huge share of practical queries.

LinkedIn skews toward Perplexity and B2B contexts, where professional profiles and company pages serve as identity and credibility signals.

A blunt caveat on this layer: it is not something you can or should try to fake. Astroturfing Reddit, planting Quora answers, or stuffing forums gets detected, downranked, and can damage the brand. The legitimate play is to be genuinely present and genuinely useful where your audience already talks — which is slower, but it's the only version that survives. Anyone promising to "flood Reddit so the AI picks you up" is selling a liability.

What you actually control

Here's the part nobody likes, stated plainly: you cannot inject content into ChatGPT, Gemini, Perplexity, or Google's AI. There is no dashboard, no paid placement, no API that lets a brand insert a sentence into a model's answer. Anyone who tells you they "control how AI talks about your brand" is selling vaporware. We say this to prospects regularly, and it disqualifies a chunk of what the market wants to buy.

So if you can't touch the output, what can you do? You influence the inputs. Three of them, specifically.

Entity existence. Does a machine-readable record of your organisation exist, and is it correct? This is the single highest-leverage thing for most brands, because it's binary in a way the others aren't — either the grounding layer knows you exist as a distinct entity, or it doesn't. A Wikidata item, a Wikipedia article where notability supports one, a complete Google Business Profile, consistent presence in industry databases.

Source authority. When the model retrieves or recalls facts about you, where do they come from? Independent, reputable, editorial sources carry far more weight than your own marketing pages. This is where classic earned media and PR still matter enormously — they're not just for humans anymore; they're the high-trust substrate the models learn from. A brand with substantive coverage in reputable outlets is a brand the AI can cite confidently.

Consistency across the web. Models cross-reference. If your founding year, headquarters, leadership, and core description say one thing on your site, another on LinkedIn, a third on an old press release, and a fourth on a directory, you've introduced ambiguity — and ambiguity makes a model hedge, generalise, or get it wrong. Consistency is unglamorous and it's one of the most common reasons AI answers about a company are subtly off.

Notice what all three have in common: they're about building a reliable source base, not gaming an algorithm. That's the honest core of AI visibility. You're not tricking the model — you're giving it accurate, consistent, well-attributed material so that when it does talk about you, it gets you right and is more likely to name you.

The AI-visibility stack

It helps to think of all of this as a layered stack, built bottom-up. Each layer makes the one above it more effective, and skipping the foundation undermines everything else.

Layer 1 — Entity. The machine-readable identity: Wikidata item, knowledge-graph presence, stable identifiers, a clean Google Business Profile. This is the bedrock. Without it, the model isn't sure you exist as a distinct thing, and everything above is built on sand. Highest leverage, usually the first thing to fix.

Layer 2 — Encyclopedic. The neutral, authoritative reference layer — chiefly Wikipedia, where notability allows it. This is the heavily-weighted, high-trust source that engines (ChatGPT especially) lean on hardest. It both feeds training corpora and reinforces the entity layer beneath it.

Layer 3 — Community. Reddit, Quora, YouTube, LinkedIn — the discussion and opinion layer that drives recommendation-shaped answers and is disproportionately important for Google's and Perplexity's surfaces. Earned genuinely, never faked.

Layer 4 — Owned. Your own website, blog, documentation, and structured data (schema markup). This is the layer you control most directly and, somewhat counterintuitively, the least independently trusted — a model knows your site is your marketing. Owned content matters for RAG retrieval and for feeding clear facts into the layers below, but it can't carry the whole load on its own. The classic SEO instinct to pour everything into owned content is exactly backwards for AI visibility.

The mistake most brands make is starting at Layer 4 (publish more blog posts!) and ignoring Layers 1–2. The stack works bottom-up: fix your entity, earn your encyclopedic and authoritative coverage, build genuine community presence, then let owned content amplify. A great blog on top of a non-existent entity is a great blog the AI can't attribute to anyone.

How to audit your current AI visibility

You can get a rough read on where you stand in an afternoon, without buying anything. Here's a practical starting sequence.

1. Ask the engines about yourself. Open ChatGPT, Gemini, and Perplexity and ask each the questions a customer would: "What is [your company]?", "Who are the leading companies in [your category]?", "Is [your company] a good choice for [use case]?" Note three things: Are you mentioned at all? Are the facts correct? Which sources get cited? This is your baseline, and it's often sobering.

2. Check your entity layer. Search Wikidata for your organisation — is there an item, and is it accurate? Look at whether a Google Knowledge Panel appears when you search your brand name. These tell you whether the grounding layer knows you exist.

3. Audit consistency. Pull your core facts — founding year, HQ, leadership, one-line description — as they appear across your site, LinkedIn, Crunchbase, directories, and any old press. Flag every discrepancy. Each one is a small reason for a model to hedge or err.

4. Map your source base. List the genuinely independent, reputable coverage of your brand from the last couple of years. Be strict: your own blog, sponsored posts, and press-release syndication don't count. This is the material the trustworthy layers are built from — and if the list is thin, that's your real constraint, not your SEO.

5. Find your community gaps. Search Reddit and Quora for your category and your brand. Are the relevant conversations happening without you? Is the existing discussion accurate?

Where you start depends on what the audit reveals. If the engines don't know you exist, start at the entity layer — that's foundational and binary. If you exist but the facts are wrong, fix consistency and shore up authoritative sources. If you're accurate but invisible in recommendation queries, the community layer is your gap. And if your independent source base is genuinely thin, the honest answer is that no AI-visibility tactic substitutes for earning real coverage first — the same truth that governs whether a Wikipedia article is even possible.

None of this is fast, and none of it is a trick. AI visibility is the slow, compounding work of becoming a brand the internet describes accurately and consistently — so that when an answer engine reaches for a source, yours is the reliable one it finds. That's not a hack you buy. It's a base you build.

WikiBusines builds the encyclopedic and structured-data foundation that AI answer engines rely on. If you want an honest read on your current AI visibility, email team@wikibusines.com and we'll run a baseline audit.