Four independent studies published in the past three weeks converge on an uncomfortable finding for marketers: getting cited by an AI answer engine has almost nothing to do with the search rankings brands spent two decades building, and the citations that do land are far less stable than anyone assumed.
Between late July and this month, Steady Demand, the B2B research firm 2X, Similarweb and Yext each published separate research into how ChatGPT, Gemini and Perplexity decide which brands to mention. None of the four coordinated with each other. All four arrived at a version of the same conclusion: AI search visibility is not a smaller, weirder cousin of SEO. It runs on different inputs, rewards different behavior, and right now it is being won largely by brands that do not realize they are playing.
The citations that will not sit still
Steady Demand’s contribution is the most granular. The local SEO agency logged 14,472 citations across 1,487 identical local-search queries run through Gemini in 50 U.S. metro areas, then repeated the same queries hours and days apart to see whether the answers held steady. They mostly did not. Back-to-back identical queries produced only 46.3% overlap in cited domains. By the next day, that overlap had fallen to 26.5%. Gemini recommended the exact same top business only 7.9% of the time, compared with 90.2% for Google’s traditional local pack.
Steady Demand co-founder Ben Fisher and the firm’s research team framed the mechanism behind the swings this way: “each call is its own independent, evidently stochastic decision about what to type into Google on your behalf.” In other words, Gemini is not retrieving a fixed answer and re-reading it back. It is regenerating the underlying search strategy every time, which means the citation a brand earns this morning is no guarantee of the citation it earns this afternoon. Steady Demand’s own published dataset shows the pattern held across metros and query types, and a separate SISTRIX study of 82,619 prompts in May found weekly domain churn of 5% to 74% across platforms, which suggests the instability is structural rather than a Gemini quirk.
For a category that has spent years building dashboards around stable, auditable rank positions, that is a hard floor to build on. A brand cannot optimize its way to a fixed slot in an answer that is being partially reinvented on every request.
Who is actually winning, and why
If citations are unstable request to request, they are not random in aggregate, and Similarweb’s Generative AI Brand Visibility Index is the clearest look yet at what separates the brands AI models cite from the ones they skip. Similarweb ranked 113 brands across finance, travel, consumer electronics, beauty, fashion and news, then flagged the “overachievers”: brands that show up in AI answers more often than their traditional search demand would predict. In consumer electronics, Apple leads outright, but specialists like B&H, Adorama and iFixit are gaining faster than their search volume alone would explain. In news, Reuters ranks first, while Science Direct, a far smaller property by scale, qualifies as an overachiever.
Similarweb’s Adelle Kehoe, director of product marketing, said the pattern points to a different kind of authority than SEO rewards: “Authority, not just scale, is emerging as the differentiator.” She added that the effect compounds for brands with a long track record in a narrow lane: “Years of building trust, deep specialism, and recognizable positioning are now compounding in AI search.” Read against Steady Demand’s instability data, the two findings fit together. Scale does not buy a fixed citation slot the way it bought a stable page-one ranking. Depth in a specific domain appears to survive the noise better than breadth does.
The B2B blind spot
The picture gets worse once the lens shifts from consumer categories to B2B. The AI Innovation Lab at 2X benchmarked 70 B2B companies across the discovery, evaluation and purchase stages of the buyer journey and found 96% are effectively invisible in early-stage AI answers: the moment when a buyer asks a generative engine what to consider before they know which vendors exist. Only 4.3% of the companies studied maintain what 2X calls a healthy discovery funnel, appearing when buyers are still forming the question rather than only once they already know the company’s name.
2X’s Lisa Cole, chief marketing, product and AI officer, put the stakes in blunt terms: “CMOs are waking up to a hard truth: you can’t manage what you don’t show up for.” Will Waugh, executive director of the 2X AI Innovation Lab, pointed to the underlying mechanics: “AI models don’t care about org charts or market caps. They respond to clarity, consistency, and corroboration.” The firm’s audit tied the gap to concrete, fixable issues: missing or incomplete structured data, AI crawlers that are blocked or simply unmanaged, thin third-party review coverage, and community sentiment on forums like Reddit that nobody at the company is tracking.
Different engines, different rules
The instability inside a single engine is only half of Steady Demand’s finding. The other half is that Gemini and ChatGPT barely agree with each other in the first place. Running the same 1,487 local queries through both platforms, Steady Demand found the two engines cited overlapping domains only 8% of the time and recommended the identical top business just 4.2% of the time. The two platforms are not drawing on the same evidence to answer the same question: Gemini cited business websites directly in 59.9% of cases, while ChatGPT leaned far more heavily on Reddit and business directories, at 15.9% for direct business-website citations. Perplexity, per Yext’s separate analysis, diversifies further still, pulling from sources like MapQuest and TripAdvisor that neither Gemini nor ChatGPT favors.
That cross-engine divergence is what makes a single-platform AI-visibility score misleading on its own. A brand tuned for Gemini’s preference for owned websites can still be nearly invisible on ChatGPT if its Reddit and directory presence is thin, and vice versa. The gap between what search ranks and what AI cites that MarTech covered earlier this year turns out to be a gap between what each AI engine cites, not a single, unified “AI search” surface at all. Marketers auditing visibility on one model and assuming it transfers to the others are working from a false premise.
Why brand-controlled sources win
Yext’s research supplies the mechanism that ties the other three studies together. Analyzing 6.8 million AI citations across ChatGPT, Gemini and Perplexity between July and August, Yext found 86% originate from sources the brand already controls or directly influences: its own website, its business listings, and its owned reviews and social profiles. Reddit and open forums, despite their outsized reputation as an AI-citation goldmine, accounted for only 2% of citations in Yext’s dataset. The split varies by model: Gemini leans toward websites at 52.1%, while OpenAI’s citations skew toward listings at 48.7%.
Yext chief data officer Christian J. Ward drew the conclusion directly: “The most impactful sources are the very ones they can already control or influence.” CEO Mike Walrath framed it as a return of agency to marketers who may have assumed AI search was a black box: “When brands control their data, they control their visibility.” Set against 2X’s finding that 96% of B2B brands are invisible in early-stage answers, the implication is that the gap is not mostly an algorithm problem. It is a data-hygiene and content-coverage problem that predates the AI layer and simply gets exposed by it.
The mix of which owned channel matters most also shifts by category, per Yext’s breakdown. In retail, first-party websites drove 47.6% of citations. In financial services, brand-owned websites accounted for 48.2%. In healthcare, business listings dominated at 52.6%, ahead of website citations, a reversal from retail and finance that Yext’s researchers tied to how heavily consumers lean on directory-style comparison when picking a provider rather than a product. A single, one-size-fits-all AI-visibility checklist misses that difference: the fix for a healthcare brand is closer to a local-listings audit than a content-marketing push.
What it means for the marketing leader
Taken together, the four studies describe a visibility system with different physics than organic search. A ranking earned in March does not guarantee a citation in April, and a citation earned this morning is not guaranteed this afternoon. The brands overperforming their scale are the ones with deep, structured, well-corroborated content in a specific domain, not the ones with the biggest media budget. And the raw material AI models draw on most heavily is not third-party buzz but the brand’s own structured data, listings and reviews, which is also the layer marketing teams have the most direct ability to fix.
That reframes the CMO’s AI-search problem from a ranking exercise into an information-architecture one. The shared measurement standard the industry adopted earlier this year gives marketers a common way to compare citation share across platforms, but a stable measurement standard does not fix an unstable underlying signal. Marketing teams are increasingly treating AI visibility the way they once treated technical SEO: a discipline with its own audit, its own owner and its own budget line, distinct from paid AI-visibility advertising and separate from the content team that writes for humans.
How to evaluate AI visibility now
Three things follow from the data rather than from vendor pitch decks. First, treat any single AI-citation snapshot as exactly that, a snapshot, not a KPI: Steady Demand’s own data shows a single query result can flip within hours, so a monthly or quarterly cadence of repeated, identical queries matters more than any one measurement. Second, audit the structured data, review coverage and crawler access that Yext and 2X both flag as the actual inputs models draw on, since that is the layer a marketing team can change without waiting on a platform’s black box to behave. Third, benchmark against category specialists, not just direct competitors: Similarweb’s overachiever data suggests the brands gaining ground are often narrower and deeper, not bigger. Marketers who spent the last two years treating AI visibility as a media-buying line item are now looking at four studies in one month that all say the harder, less glamorous work is the one that moves the number.
Source: Steady Demand
