What makes AI engines recommend one brand over another?

An AI engine recommends a brand by assembling a consensus from text it can access, through two routes that behave differently. The first route is model memory: training crawlers such as GPTBot collect web content that shapes what a model knows without searching [2]. The second route is live retrieval, where the engine searches the web at answer time, reads what it finds, and grounds its recommendation in those sources. Google documents that its AI Overviews and AI Mode may use a query fan-out technique, issuing multiple related searches to build one response [1], and OpenAI documents that a dedicated crawler, OAI-SearchBot, controls whether a site can appear in ChatGPT search answers at all [2]. In both routes the recommendation is a summary of what independent sources say rather than a lookup in some brand database, which is why the drivers examined below are mostly about what third parties publish, not what a brand says about itself [7].

The evidence behind those drivers comes in three grades. A small set of behaviors is documented by the platforms themselves: eligibility gates, crawler rules, and statements about how shopping results are assembled [1][2][13]. One driver set has been tested experimentally, in a Princeton-led benchmark that measured how specific content changes affect visibility in generative answers [3]. Everything else, including the headline statistics about brand mentions, listicles, and reviews, comes from correlational studies that measured what recommended brands have in common without proving cause [4][5][6]. This answer labels each driver accordingly.

The engines also differ from each other more than most coverage suggests. A Profound analysis of 680 million citations found ChatGPT's most-cited domain was Wikipedia at 7.8%, while Google AI Overviews and Perplexity both leaned on Reddit [8], and Muck Rack found the platforms cite at different rates entirely, from 96% of ChatGPT responses down to 55% for Claude [7]. A brand can be the default recommendation in one engine and absent from another, so every driver below is a tendency, not a universal rule [6][8].

Mention volume is the strongest single predictor measured to date. Ahrefs studied 75,000 brands with established websites and found branded web mentions correlated with AI Overview brand visibility at 0.664, ahead of branded anchor text at 0.527, branded search volume at 0.392, Domain Rating at 0.326, and referring domains at 0.295 [4]. The authors flagged the obvious caveat themselves, writing that correlation is not causation and that even the top factors showed only moderate strength [4]. Semrush's 2026 study of 1,094 ChatGPT categories points the same direction with a weaker signal: brands that owned a topic had higher branded search volume than their runner-up in 55.7% of pairs, while organic traffic predicted ownership no better than a coin flip at 48.4% [6].

The plausible mechanism is co-occurrence. Language models learn associations from repeated proximity in text, so a brand that appears next to its category terms across many independent pages becomes the statistically expected completion when someone asks for a recommendation in that category. That mechanism also explains why mentions work without links: the model reads the sentence, not the anchor tag. The practical move is to count and grow unlinked mentions in the places each engine demonstrably reads, which differ by platform. Profound's citation data puts Wikipedia and Forbes high for ChatGPT, and Reddit, YouTube, and Quora high for Google AI Overviews and Perplexity [8].

Earned media supplies the bulk of what AI engines cite. Muck Rack's Generative Pulse study, which analyzed more than 25 million links from ChatGPT, Claude, and Gemini responses across 17 industries, found 84% of citations pointed to earned media, with journalism alone accounting for 27%; paid and advertorial content accounted for 0.3% [7]. Across the study's three editions since July 2025, the earned share has held between 82% and 89%, which makes it one of the most stable findings in this field [7]. The implication for recommendations is direct: when an engine answers a "best X for Y" prompt, it is largely paraphrasing what editors, reviewers, and communities have already written.

Ranked comparison articles are the most concentrated version of this driver, because a listicle hands the engine a pre-built recommendation set. Seer Interactive measured 2 million citations from November 2025 through February 2026 and found listicles made up 16.6% of total ChatGPT citations, a large share for a single content format [11]. The same study is a caution flag: listicle citations fell 30% from December 2025 to January 2026, declining in 13 of 16 industries, as ChatGPT became more selective [11]. The durable version of this driver is earning inclusion in the specific roundups and press coverage already being cited for your target prompts, rather than treating any one format as permanent. Publishing your own "best of" page on your own domain rarely substitutes, since owned content is competing for the 16% of citations that earned media does not already occupy [7].

Reviews are the one recommendation input a platform has described on the record. When OpenAI launched shopping results in ChatGPT search, it stated the results are not paid ads and that OpenAI takes no commission, with products assembled from structured metadata such as price and availability plus third-party content [13]. OpenAI's product lead described the selection logic in plain terms: the system is "trying to understand how people are reviewing this, how people are talking about this, what the pros and cons are" [13]. That is a documented statement that review text itself, beyond star averages, feeds the recommendation.

The citation data backs this up at scale. In Profound's 680-million-citation analysis, the review platform G2 accounted for 1.1% of all ChatGPT citations, tied with Forbes, while Gartner at 1.0% and Yelp at 0.8% ranked among Perplexity's most-cited domains [8]. Those look like small percentages until you consider they are shares of every citation across every topic; within software or local-service recommendation prompts, review platforms are heavily concentrated. Sentiment is the honest gap in the evidence. No published study has isolated how strongly positive versus negative review sentiment changes an engine's recommendation, and vendor claims about sentiment weighting are unmeasured. What can be said is mechanical: because engines summarize pros and cons from review text [13], the substance of your reviews, including the specific complaints, is likely to surface verbatim in answers, so the review corpus is both a visibility driver and a message the engine repeats.

Search visibility works as a documented gate and a correlated strength, and the distinction matters. On the documented side, Google states that to appear as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to be shown in Google Search with a snippet, and that no additional optimizations or special markup are required [1]. OpenAI documents an equivalent gate: sites that opt out of OAI-SearchBot are not shown in ChatGPT search answers, though they can still appear as navigational links [2]. Failing these gates removes a brand's own pages from the retrieval pool entirely, no matter how strong every other driver is.

Beyond the gates, the correlations are real but uneven. Seer Interactive ran 10,000 buyer-style questions through GPT-4o and found brands ranking on page 1 of Google correlated with LLM mentions at roughly 0.65, with Bing rankings slightly weaker at 0.5 to 0.6 [5]. The same study found backlinks had weak or neutral impact, which the Ahrefs brand study corroborates: Domain Rating correlated at only 0.326 and raw referring domains at 0.295 [4]. Semrush's topic study found its Authority Score predicted ChatGPT topic ownership in just 52.5% of pairs, statistically indistinguishable from chance [6]. The reasonable synthesis is that ranking well for the sub-questions an engine fans out to still feeds retrieval [1][9], but accumulating domain authority as an end in itself shows little measurable payoff in AI recommendations [4][5][6].

AI engines cite measurably fresher content than classic search rankings surface, but the effect varies by platform and topic. Ahrefs analyzed roughly 17 million citations across ChatGPT, Perplexity, Gemini, Copilot, and AI Overviews and found the average AI-cited URL was 1,064 days old versus 1,432 days for URLs in organic Google results, a 25.7% freshness gap [10]. ChatGPT showed the strongest preference for recent pages, while Google's AI Overviews actually cited content slightly older than organic results, by about 16 days [10]. So a freshness push aimed at ChatGPT visibility may do nothing for Google's AI surfaces.

Within a single retrieval set, recency behaves more like a tiebreaker than a primary driver. Ahrefs' separate study of 1.4 million ChatGPT prompts found the median cited page was around 500 days old, and that freshness became decisive mainly for news-flavored queries, where cited pages skewed clearly younger than non-cited ones at equal relevance [9]. For a brand competing on recommendation prompts, the evidence supports keeping comparison-relevant pages visibly maintained and dated, and expecting the payoff to concentrate in engines and topics where freshness is rewarded, rather than treating updates as a universal lever [9][10].

Entity consistency is the most widely recommended driver with the least direct evidence, so it belongs in the inferred tier. No platform documents a consistency factor, and no large study has isolated it. The inference runs through retrieval mechanics that have been measured. Ahrefs' 1.4-million-prompt study found cited URLs showed substantially higher semantic similarity to the engine's internal fan-out queries than non-cited ones, with cited pages scoring 0.602 against the original prompt versus 0.484 for uncited pages, and that search results with natural-language URL slugs had an 89.78% citation rate compared to 81.11% without [9]. Engines match meaning across names, titles, and descriptions, so a brand described the same way everywhere is easier to match and consolidate than one that appears under three names and four category labels.

The co-occurrence evidence points the same direction: if mention volume drives visibility [4], mentions fragmented across inconsistent names dilute the very signal being counted. The low-cost application is uniformity: one canonical brand name, one short category descriptor, and the same core facts repeated across the website, directories, review profiles, and press materials. Treat anything more elaborate sold under the entity-optimization label as unproven, because that is what the current evidence supports [9].

Three grades of evidence exist in this field, and they justify different levels of confidence. The documented grade covers what platforms publish about their own systems: Google's index-and-snippet eligibility rule and query fan-out description [1], OpenAI's crawler controls and the statement that opting out of OAI-SearchBot removes a site from ChatGPT search answers [2], and OpenAI's description of shopping results as organic, metadata-driven, and review-informed [13]. These are the only claims that describe cause. They are also narrow: none of them explains why one eligible brand beats another.

The experimental grade currently rests on one major study. The GEO benchmark, presented at KDD 2024, tested content modifications against generative engines and found the best-performing changes boosted visibility by up to 40%, with adding citations, quotations, and statistics among the strongest tactics, while effectiveness varied enough by domain that the authors recommend domain-specific approaches [3]. It is the only public evidence where researchers changed content and measured a response.

Everything else is the correlational grade: the Ahrefs mention study [4], Seer's rankings analysis [5], Semrush's topic ownership work [6], Muck Rack's earned-media shares [7], and Profound's domain rankings [8]. These studies are large and honest about their limits; Ahrefs states correlation is not causation and Seer describes its work as a first phase of spotting correlations [4][5]. Read the documented items as gates to clear, the experiment as the best guide to content changes worth making, and the correlations as a map of where recommended brands already tend to be present. Direction of causality stays unresolved: well-known brands earn more mentions and more recommendations, so some of the measured correlation is fame, not tactics [4][6].

The drivers shift faster than the studies that measure them. Semrush tracked 230,000 prompts across thirteen weeks in 2025 and watched Reddit fall from appearing in close to 60% of ChatGPT responses in early August to around 10% by mid-September, while Wikipedia dropped from roughly 55% of responses to under 20%; the shift was isolated to ChatGPT, suggesting a deliberate algorithmic change [12]. Seer's listicle data shows the same pattern at the format level, with citations down 30% in a single month [11]. Any specific tactic derived from a citation study inherits that volatility, so mention-building in a currently favored channel is a position to monitor, not a durable asset.

Recommendation instability also runs deeper than most dashboards show. Semrush found only 15.2% of 1,094 ChatGPT categories had a clear brand owner, and two related prompts within the same buying decision routinely surfaced different brands [6]. Measuring your brand on a handful of prompts therefore produces noise; topic-level tracking across many phrasings is the minimum credible measurement. Platform divergence cuts the same way, since a freshness investment helps on ChatGPT but not on AI Overviews [10], and a Reddit presence matters more for Perplexity and Google than for ChatGPT's citations [8][9]. Finally, keep the effect sizes in proportion: the strongest correlation on record is 0.664 [4], which leaves most of the variance unexplained. The honest posture is portfolio-shaped: clear the documented gates, make the experimentally supported content changes, spread earned-media effort across channels, and re-measure monthly [1][3][7].

It is not paid placement

No major engine currently sells recommendation slots in organic answers. OpenAI states its shopping results are not paid ads and that it earns no commission on sales [13], and Muck Rack's 25-million-link analysis found paid and advertorial content at 0.3% of AI citations [7]. Advertising may sit adjacent to AI answers, but the recommendation itself is assembled from earned and organic material.

It is not a single algorithm you can reverse-engineer

ChatGPT, Gemini, Perplexity, and Google's AI surfaces retrieve from different sources, cite at different rates, and favor different domains [7][8]. A driver map like this one describes tendencies across platforms; it cannot be reduced to one ranking formula, because the platforms themselves publish no such formula and behave differently from each other [1][8].

It is not classic SEO under a new name

The overlap is real at the gate level, since indexability and snippet eligibility still control whether Google's AI features can use a page [1]. The strength signals diverge, though: backlinks and domain authority show weak correlation with AI recommendations [4][5][6], while unlinked mentions, review discussion, and earned coverage carry the strongest measured associations [4][7].

It is not a platform-specific playbook

This answer maps the evidence for why recommendations happen across engines. Turning it into tactics for one engine requires that engine's own source preferences and documentation, which differ enough that a Gemini plan, a ChatGPT plan, and a Perplexity plan are genuinely different documents [8][10][12].