How does AI decide which brands to cite?
It does not run a ranking formula you can look up. An assistant composes an answer from whatever it recalls from training, whatever its search step returns, and any structured identity data it can match, then writes fluent text from that mix. Brands that appear consistently across reliable, readable sources are easier for that process to name correctly.
The vendors do not publish the weighting. What follows separates what is documented from what is our working interpretation, and labels each. For the practical plan, see how to improve brand visibility in AI search.
What does training data contribute?
Training data gives the model background knowledge about well-documented entities, frozen at a cutoff date. If your company is described in many reliable places, the model has some memory of it; if not, it may say nothing or guess.
Documented: Wikipedia is a deliberate part of training mixes. The GPT-3 paper lists Wikipedia at about 3% of the training mix (Brown et al., 2020), and Meta's LLaMA paper lists about 4.5% (Touvron et al., 2023). Those are shares of a very large corpus, so they show deliberate inclusion, not dominance. In January 2026 the Wikimedia Foundation announced AI-company partners for its Wikimedia Enterprise licensing, among them Google, Microsoft, Perplexity, Amazon, Meta and Mistral.
Consequence: training changes slowly. A rebrand or a corrected fact will not reach a model's memory until it is retrained, which is why a fix at the source can take months to show up in some assistants.
What happens when an assistant searches the web?
When a product has a search step, it runs queries, reads a handful of pages and writes the answer from them, often with links. This is retrieval, and it is where fresh, clearly written, crawlable pages matter.
Documented: Perplexity and ChatGPT's search mode show their sources next to the answer, and Google states that AI features in Search are built on its existing indexing and quality systems; see Google's AI features and your website page. So the basics of being indexable, readable and trustworthy still apply.
Working interpretation: retrieval favours pages that answer the question in one clear passage, from sources the system already treats as reliable. A thin company site competes poorly with a comparison article or a reference page that mentions the company. Which sources are chosen can vary run to run.
What is grounding, and where do Wikidata and knowledge graphs fit?
Grounding is matching a name to a specific real-world entity, so that "Apple" means the right one. Structured records such as Wikidata items and Google's Knowledge Graph help with that.
Documented: Wikidata holds stable identifiers and referenced statements and is openly licensed; in October 2025 the Wikimedia community and partners announced a Wikidata Embedding Project to make it easier for AI developers to use. Google builds knowledge panels automatically from sources across the web, does not publish which sources feed a given panel, and does not let anyone buy one; see Google's knowledge panel help.
Not documented, so we do not claim it: that Wikidata statements directly produce a knowledge panel or ChatGPT's wording. What we can say is that a clear, referenced entity record reduces the chance of being confused with a namesake. Mechanics in Wikidata and the knowledge graph.
Why do ChatGPT, Gemini and Perplexity name different brands?
Because they differ in how much they rely on memory versus search, which sources they trust, and how they phrase answers. Even one product gives different answers across runs.
| Assistant | How it commonly answers | What to check for your brand |
|---|---|---|
| ChatGPT | Training memory, with a search step when the question needs it | Whether you appear without search, and with it |
| Google (AI Overviews, Gemini) | Built on Google's indexing, with links shown | Whether your pages and independent coverage rank and are readable |
| Perplexity | Search-first, sources visible | Which pages are cited, and whether they are yours or third parties |
Treat this as orientation, not a spec. Products change monthly, and published studies of citation shares use different methods, so we do not quote a single percentage. Test your own category.
Which signals make a source more likely to be cited?
In our practice, five qualities show up repeatedly. These are not documented ranking factors.
- Independence. A trade publication or encyclopedia describing you carries more weight than your own page. Wikipedia's rule of verifiability works the same way: it needs published reliable sources.
- Consistency. The same name, date and leadership across profiles. Contradictions force a model to hedge.
- Clarity. Plain statements near the top of a page, headings phrased as questions, definitional first sentences.
- Accessibility. Pages not blocked in robots.txt and readable without scripts. Check what you allow crawlers to read.
- Recency where it matters. Retrieval-heavy products favour current pages for time-sensitive questions.
Notability sits underneath all five: where little independent coverage exists, there is little to cite. Our Notability Index shows how uneven that is by sector (in our own assessments of 2018 to 2026, 98% of Crypto / Web3 brands rated low notability against 19% in Finance / Fintech / Investment).
What do you control, and what don't you?
You control inputs; you do not control outputs.
- Cannot control: the wording of an answer, whether you are named, the date a model is retrained, which sources a product picks on a given day. No agency can promise these.
- Can control: the accuracy of your own pages and profiles, consistency of facts, access for crawlers, structured data on your site, and your effort to earn independent coverage.
- Partly influence, through legitimate routes: Wikipedia and Wikidata, which volunteers run and which require notability and disclosed conflicts of interest. See paid editing and COI disclosure and why Wikipedia is ChatGPT's top source for what that does and does not mean.
How do I check what AI says about my brand?
Run a fixed panel of prompts across assistants and log mentions, accuracy and cited sources.
- Write 10 to 15 prompts: what is [brand], who leads [category], is [brand] good for [use case].
- Run each in ChatGPT, Gemini and Perplexity, with and without search where the product allows.
- Record: named or not, facts right or wrong, sources cited.
- For each wrong or missing item, find the page the claim most likely came from.
- Repeat monthly and read the trend.
If you find an error, ChatGPT wrong about your company covers the repair. For a wider view of Wikipedia's role, read Wikipedia AI visibility. Tools that automate the panel are compared in AI brand monitoring tools.
Frequently asked questions
It draws on what it learned in training, on pages its search step retrieves when it searches, and on entity information it can match. The exact weighting is not published and answers vary between runs, so mention is a matter of probability.
Wikipedia is neutral in tone, consistently structured, openly licensed and sourced, and it is a documented part of training data for major models. That does not mean it is the only source, and it does not mean every company qualifies for an article.
No. There is no paid placement inside assistant answers, and anyone selling guaranteed mentions is selling something that does not exist. You can improve the sources behind the answers.
It depends on the product. Assistants that retrieve live pages can update in weeks after the sources change; model memory updates when models are retrained. Nobody can give a firm date.
Yes, for retrieval and as the place where your facts live, but it is the least independent source. Models treat it as a claim to be checked against other sources.
Schema markup helps machines read your facts and is good practice. For llms.txt we are not aware of documentation from the major vendors confirming that it changes citations. Neither replaces independent sources.
Then focus on independent coverage, consistent entity data and community presence, and re-check notability as coverage grows. A notability audit tells you where you stand source by source.
How WikiBusines helps
We work on the source layer. Paid editing is always disclosed on Wikipedia in WikiBusines' name, and we never guarantee that an article will be published or what an AI will say.
- AI visibility audit, €490: what the main assistants say about you, which sources they lean on and which layer to fix first. Not credited toward packages.
- Notability audit, from €490, credited toward page creation from €2,400 (English, 4 to 5 weeks), if ordered within one month; Wikidata from €680.
- Annual support, from €720/year: up to four updates and a 12-month guarantee.
To start, write to us via contact.
Authors and editorial policy
Bohdan DubylovskyiCo-founder and CBO, WikiBusines→
Roman MelnykCo-founder & Chief Editor, WikiBusines→The article was prepared by the WikiBusines co-founders. We sell paid Wikipedia services, so we disclose this interest: paid editing is always declared on Wikipedia in WikiBusines' name, and claims about the rules are backed by links to primary sources.
Editorial policy: payment disclosure, conflicts of interest, sources