That single mechanism explains most of what marketers find mysterious about ChatGPT, Gemini, Perplexity and Google's AI Overviews. This article sets out the evidence for how the selection actually works, and what a brand must do to be chosen.
What decides whether an AI engine recommends a brand?
An AI recommendation is the outcome of two overlapping systems. The first is the model's training data: the corpus of web text the model learned from, which fixes its background beliefs about who the credible players in a category are. The second is retrieval: the live search most engines now run before answering, which pulls in a handful of current pages and synthesises them, with citations, into the reply.
What an AI recommendation is not, and this is the common misconception, is a ranking. There is no page two. A typical AI answer names two to five brands and moves on, and the Pew Research Center found that users click through to the cited sources in only 1% of cases (Pew Research Center, July 2025). The answer is the entire transaction.
Strategically, that changes the game from visibility to inclusion. Search engine optimisation was a contest for position on a list every user could see in full. Generative engine optimisation, or GEO, is a contest to exist in the answer at all. Inclusion is decided by what the wider web says about you, which is why the discipline looks far more like brand building than like technical SEO.
Most brands are invisible to AI engines, and the data shows why
Here is the uncomfortable mirror. When Ahrefs analysed 75,000 established brands against millions of AI Overview responses in 2025, 26% of them had zero AI mentions. Not few. None. And the distribution above zero was savagely skewed: brands in the top quartile for web mentions averaged 169 AI Overview mentions, while the quartile below them averaged just 14 (Ahrefs, 2025).
The diagnosis is not that these brands lack good websites or good work. It is that they have spent years optimising owned channels while the machine reads earned ones. Large language models form their view of a category from words on other people's pages: the co-occurrence of a brand name with a category, the context in which it is discussed, the agreement between independent sources. A brand that is absent from that conversation is, to the model, simply not part of the category.
Think of it the way a bank thinks about credit. A lender does not take your word for your creditworthiness, however polished your presentation. It checks the file that other institutions have built about you over years of transactions. AI engines run the same check on brands. Your website is your application form. The recommendation is decided by your file.
AI engines do not discover brands. They repeat what the rest of the web already says about them.
Exhibit 1: The recommendation stack, four signals that put a brand in the answer
Vantage's GEO practice works to a four-signal model of how engines select brands. Each signal is necessary; none is sufficient alone.
1. Breadth of independent mentions
The volume and spread of third-party references to the brand, linked or unlinked. This is the strongest measured correlate of AI visibility (0.664 in the Ahrefs study) and the slowest to build.
2. Entity clarity
Whether every source that describes the brand describes it the same way: same category, same specialisms, same markets. Contradictory descriptions dilute the model's confidence in what the brand is, and a brand the model cannot classify is a brand it cannot recommend.
3. Extractable authority content
Definition-led pages, named frameworks, specific numbers and structured FAQs that retrieval systems can lift directly into an answer. This is what makes the brand's own site citable when the engine goes looking.
4. Presence in the sources engines actually read
Reddit, Wikipedia, major publications, professional directories and review platforms dominate AI citations. A brand with no footprint in that ecosystem is betting everything on retrieval finding its own domain.
The order matters. Most firms invest in signal three because it is the only one they fully control, then wonder why inclusion does not follow. Signals one, two and four are where recommendation is actually won.
Mentions beat backlinks, and the evidence is not close
For twenty years the currency of search authority was the backlink. The data says AI engines have devalued it. In the Ahrefs correlation study, branded web mentions correlated with AI Overview visibility at 0.664 and branded anchor text at 0.527, while raw backlink count managed only 0.218 and Domain Rating 0.326 (Ahrefs, 2025). Unlinked mentions, which classical SEO treats as a wasted opportunity, are read by language models as evidence of standing.
The logic is mechanical rather than editorial. A model learns from text, and a sentence that says "the leading Singapore brand consultancies include..." teaches it category membership whether or not a hyperlink is attached. Words, prevalence and context do the work links used to do.
The practical consequence: public relations, analyst coverage, conference talks, podcast appearances and genuine industry commentary now have a measurable machine-visibility payoff, on top of their human one. The activities most consultancies filed under "nice to have" have become the primary input to the recommendation engine.
AI engines take their opinions from a small and lopsided set of sources
Where do the machines actually read? Ahrefs' analysis of 9.6 million ChatGPT queries found the most cited domains were Reddit, Wikipedia, Amazon, Forbes and Business Insider, with Reddit alone cited over 4.3 million times globally (Ahrefs, 2025). Pew's study of Google AI summaries found Wikipedia, YouTube and Reddit collectively accounted for 15% of all cited sources, and that government sites were cited three times more often in AI summaries than in classic results (Pew Research Center, 2025).
Two implications follow. First, community and reference platforms carry disproportionate weight, which is why a brand's presence in professional discussion matters more than its advertising. Second, the list is overwhelmingly Western and English-language. In ChatGPT's global top 100 cited domains, Southeast Asia is represented by only a handful of sites, such as Indonesia's Blibli and Alodokter. The region's brands are being judged by a jury that has barely heard of the region.
How AI recommendation plays out in Southeast Asia
For Southeast Asian brands the consensus mechanism cuts both ways. The thinness of authoritative English-language coverage of regional categories means AI engines often build their view of "branding agencies in Singapore" or "hospitals in Malaysia" from directory listicles, aggregator content and whatever fragmentary coverage exists. Brands that would never make a considered shortlist get recommended because they dominate the thin corpus. Strong regional specialists stay invisible because nobody authoritative has written them into the record.
This is also how misdescription happens. Vantage has direct experience here: AI engines at one point described the firm as a packaging and FMCG specialist, a characterisation drawn from stray third-party listings rather than from anything Vantage publishes. When the corpus about you is thin, any loud wrong source can outvote you. Correcting it means deliberately publishing and placing consistent entity descriptions until the consensus flips, a process we set out in how to make sure AI describes your brand correctly.
The opportunity is the same asymmetry inverted. Because regional coverage is thin, a Southeast Asian brand that invests early in authoritative, citable, consistently described content can establish itself as the reference point AI engines reach for, at a fraction of what the equivalent position would cost in a saturated US category. The window is open precisely because so few regional players have noticed it.
When to invest in AI visibility
The moment to act is before the category consensus hardens. Common triggers include a rebrand or repositioning that the web has not yet caught up with, a finding that AI engines misdescribe or omit the brand, a competitor beginning to appear in AI answers for shared queries, entry into a new market where the brand has no earned footprint, or a category where buyers have visibly shifted to asking ChatGPT before asking Google. To establish where you stand today, start with how to measure your brand's visibility in AI search.
Cadence matters more than intensity. Consensus is built by consistent publication and consistent third-party presence over quarters, not by a one-off content sprint, and engines re-read the web continuously. Treat AI visibility as a standing brand-building discipline with a quarterly measurement rhythm. The tactical playbook for earning citations is covered in how to get your brand mentioned by ChatGPT.
There is also a case for waiting: if a brand's positioning is genuinely unsettled, fixing that comes first. Amplifying an unclear identity teaches the machines the wrong thing faster. Strategy before syndication, always, which is why the work begins with a brand positioning framework. As a benchmark for the underlying brand work, most Singapore branding programmes fall between S$5,000 and S$50,000, with enterprise work higher, and qualifying Singapore SMEs can offset up to 50% of eligible costs through the Enterprise Development Grant.