How AI engines decide whether to cite your brand, and why ChatGPT, Claude and Perplexity don't agree

What is AI visibility, and why is it different from ranking on Google?

AI visibility is whether an AI answer engine names your business, describes it correctly and links to you when someone asks it for a recommendation. On Google a buyer sees ten options. In an AI answer they often see three.

Someone in Centurion types "who's a reliable accountant for a small construction company near me?" into ChatGPT or Perplexity. They don't get a page of results. They get a short paragraph naming a handful of firms, with a sentence on why each one fits. If you're not in that paragraph, you weren't considered, and nothing shows up in your analytics. The buyer never visited your site, so there's no bounce to see.

Sight AI's guide to brand visibility in large language models describes this as a silent loss. You don't see it drop because nothing happened on your side. It also raises a second risk that gets less attention: when an engine does mention you, it can repeat an old price, a complaint from years ago, or an unflattering comparison.

How do AI systems interpret, summarise and cite a brand?

In three stages. The engine works out what kind of business you are, squeezes what it knows about you into a sentence or two, then decides which URL, if any, gets the credit. You can lose at each stage for a different reason.

1. Interpret: is the engine sure what you are?

Before an engine can recommend you, it has to connect your name to a category and a place. A line like "Mokoena & Associates is a Pretoria accounting firm for construction SMEs" gives it something to work with. "We deliver excellence across the financial journey" doesn't. If your homepage, your Google Business Profile and your LinkedIn page each describe you differently, the engine can't tell which description is true, so it tends to leave you out.

2. Summarise: which sentence makes it through?

Engines compress. Out of a 1,500-word services page, maybe one sentence ends up in the answer, and it's usually a plain factual one: who you serve, where, for how much. Put that sentence directly under a heading worded as a question. Don't leave it for the engine to piece together from a hero banner and three paragraphs of brand story.

3. Cite: which URL gets the credit?

When an engine retrieves pages live, it links to whichever page stated the answer most directly and could actually be fetched. That's often a competitor's comparison article or a directory listing, not your own site, simply because that page answered the question and yours didn't.

How does the routing layer decide which model answers which query?

A routing layer classifies each question by type and difficulty, sends it to whichever model is best at that kind of task, then combines the results. A brand question can be answered by different models depending on how it was phrased.

AI products are moving away from one fixed model. Digital Applied's guide to Perplexity Computer describes an orchestrator that coordinates up to 19 models and does three jobs:

1. Task classification. Before any model runs, it decides whether the question needs a web search, document analysis, computation or writing.
2. Model selection. It picks the model that performs best on that kind of task, taking speed and availability into account.
3. Result synthesis. It combines what its sub-agents return into one answer with inline citations.

According to the same guide, factual research questions go to models built around retrieval plus live web search, while writing, coding and maths each go to different specialists. The details change as new models arrive, but for marketers the direction is clear, and it has three consequences:

• You can't optimise for one model. Next month the router may send "best accountant in Pretoria" to a different model than "compare accountants in Pretoria for a construction business". What stays the same is the web those models retrieve from.
• Recommendation questions go to retrieval. Questions like "best X", "X near me" and "X vs Y" count as research, and research gets routed to models that fetch live pages and cite them. So the page a crawler can fetch today counts for more than anything a model memorised last year.
• Citations get checked more closely. A system designed to tie every claim to a source will prefer a page where the claim is stated plainly and is easy to confirm.

Which kind of engine answers which kind of question?

The same logic applies to how you check your visibility, and to which AI tools your team uses for which job:

Why does one AI surface cite you cleanly while another does not?

Usually because the engines are reading from different places. The one that cites you fetched a page it could read. The one that doesn't is answering from memory, is blocked, or found someone else's page first.

If your team is building a wider AI-driven strategy, this is the table to keep handy. Each row is something you'll see in the answers, what usually causes it, and what to change:

The second row isn't hypothetical for us. While writing this post, we checked our own blog and found every post was loaded by JavaScript after the page opened. Most AI crawlers don't run JavaScript, so to GPTBot or PerplexityBot the blog was an empty "Loading…" screen. That's why this guide is published as plain HTML at its own address.

How do you become the default answer for your category?

Be the source that defines the category in plain words, and get independent sites to repeat that definition. Engines lean towards the answer that most sources agree on.

• Publish the explainer. If you're the first business in your niche to write a clear, honest page answering "what is X and how do I choose one?", you become the source other people, and engines, borrow from.
• Say the same thing everywhere. Use the same one-line description, service list and prices on your site, Google Business Profile, LinkedIn page, directories and review profiles. Every mismatch gives an engine a reason to doubt you.
• Build local proof. For South African businesses, being mentioned by SA publications, industry bodies, chambers of commerce and review platforms like HelloPeter and Google tells engines where you operate, which a global brand can't easily copy.
• Publish comparisons. Write the honest "us vs the alternatives" page yourself. If you don't, the engine will use somebody else's.
• Update regularly. Retrieval engines prefer current pages. Sight AI notes that because of retrieval, new content can affect AI answers within weeks, compared with the months traditional SEO usually takes.

How should you measure AI visibility and citation quality?

Use a fixed set of 20–30 real buyer questions. Run each one several times on every engine your customers use, every month, and record more than whether you were mentioned.

The same question can get a different answer each time you ask, so a single check tells you very little. Track:

• Mention rate: the share of runs where your business is named.
• Position: first recommendation, or last in a list of five.
• Accuracy: whether the price, location and services are right.
• Sentiment: a clear recommendation, a hedged mention, or a warning.
• Citation quality: whether it linked to you, and to the right page, or to a directory or a competitor's round-up.
• Share of voice: which competitors appear alongside you, and how often.

Split the questions into broad category questions, specific use cases and feature or price questions. Most businesses find they appear for one type and not at all for another, and that gap tells you which page to build next.

Can you make sure your brand appears in every AI answer, every day?

No, and be wary of anyone who says they can. You don't control anyone else's model. What you control is how likely you are to be chosen, and whether you're described correctly when you are.

A realistic target: every page that matters can be fetched by every major AI crawler; every core fact is stated once, plainly, and marked up; independent sources back you up; and each month you measure whether your mention rate and citation quality are going up. Make that routine and you'll turn up in more answers, more often, with the right link.

A checklist to start this week

1. Check robots.txt allows GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot and Google-Extended.
2. Load your key pages with JavaScript switched off. If the content disappears, most AI crawlers can't see it either.
3. Write your one-sentence definition (who you are, who you serve, where) and use it on every profile.
4. Add Organization or LocalBusiness JSON-LD, plus FAQPage for questions buyers really ask.
5. Give each service and each area you serve its own page, and open each one with a direct answer.
6. Write down 20 buyer questions, run them through ChatGPT, Claude, Gemini and Perplexity, and record the results as your baseline.

Common questions

What is AI visibility?
How often, how accurately and how favourably AI answer engines such as ChatGPT, Claude, Gemini and Perplexity mention and cite a business when someone asks them a question instead of searching. You measure it by running a fixed set of real buyer questions across the engines and recording mentions, position, accuracy and the URL cited.

What is an AI routing layer?
An orchestrator that sorts each incoming question by type and difficulty, picks the model best suited to it, and combines the results into one answer. Digital Applied reports that Perplexity Computer routes across up to 19 models this way. For a brand, it means the same question can be answered by different models, so your facts need to be readable on the live web rather than relying on any one model's memory.

Can a business guarantee it appears in every AI answer every day?
No. AI answers vary with phrasing, engine and model version, and nobody controls another company's model. You can control whether your pages are readable by AI crawlers, whether your core facts are stated plainly and consistently, whether you have structured data, and whether independent sites back you up.

How do I get cited correctly instead of just mentioned?
Give each important fact its own page that doesn't depend on JavaScript to show its text, answer the question in the first sentence under a question-style heading, mark prices and FAQs up with JSON-LD, keep facts identical everywhere, and allow AI crawlers in robots.txt.