For years, the working assumption in local search was that third-party directories and review platforms held the real power. Yelp, Tripadvisor, aggregator sites - these were the gatekeepers. That assumption needs revisiting. Data from Gemini's local citation behaviour shows that nearly 60% of its local recommendations point directly to business websites. Not directories. Not review platforms. The brand's own site.
That is a significant signal. It tells you something concrete about how Gemini is making its decisions, and it has direct implications for where businesses should be investing their AI visibility effort.
What 60% Actually Means in Practice
When Gemini answers a local query - say, 'best Italian restaurant in Manchester' or 'accountants near me in Bristol' - it is pulling citations from somewhere. The finding that business websites account for the majority of those citations is not a minor footnote. It means Gemini's extraction model is treating owned web content as the most reliable source of local information for a significant proportion of queries.
This matters because the instinct in local marketing has often been to prioritise Google Business Profile completeness, aggregator listings, and third-party review volume. Those still matter. But if Gemini is favouring the business website itself, then the quality and structure of that owned content becomes a direct ranking input - not a secondary signal.
Think about what Gemini needs to extract: business name, location, services, opening hours, pricing context, and ideally some signal of quality or relevance. If that information is buried in dense paragraphs, hidden behind JavaScript rendering, or simply absent from the page, no amount of directory presence will compensate for it.
The Inconsistency Problem: Same Query, Different Recommendations
Here is where the data becomes genuinely uncomfortable for anyone trying to build a repeatable AI visibility strategy. Citation patterns in Gemini can shift dramatically even when identical search queries are run. The same business might appear in one response and be absent from the next. This is not a small variance. It is the kind of inconsistency that makes measurement feel pointless and planning feel speculative.
It reflects something practitioners in this space already know: AI responses are probabilistic, not deterministic. Gemini is not returning a fixed ranked list. It is generating a response, and citations are part of that generation process. The implication is that a business earning citations is not guaranteed to earn them consistently - and a business missing from one response might appear prominently in the next.
This is why measuring AI search visibility through spot checks is unreliable. A single query run at a single point in time tells you almost nothing about your actual citation frequency. You need repeated sampling across query variants, across time, and ideally across devices and locations. Without that, you are building strategy on anecdote.
What Your Website Needs to Be Extractable
If Gemini is favouring business websites, the question becomes: what makes a business website extractable? The answer is structured, accessible, unambiguous content. Gemini needs to be able to identify what you do, where you do it, and why you are credible - without inference, guesswork, or having to follow multiple internal links to piece the picture together.
LocalBusiness schema is the obvious starting point. Name, address, phone, opening hours, service area, price range - all of it marked up correctly and kept current. But schema alone is not sufficient. The natural language content around it matters too. A page that leads with clear service descriptions, includes specific location signals in the copy, and addresses common local queries in plain English gives Gemini far more to work with than a page that relies purely on structured data.
Crawler access is a basic prerequisite that still catches people out. If your site blocks Googlebot or has aggressive bot filtering that prevents Gemini's crawlers from accessing key pages, you are invisible regardless of content quality. Check your robots.txt, review your server-side bot management rules, and make sure your local landing pages are not sitting behind authentication or JavaScript walls that prevent extraction.
Citations Vary - But Your Content Should Not
The instability in Gemini's citation behaviour does not mean the work is futile. It means the goal should be raising your citation probability across a wider sample of queries, rather than optimising for a single response. That distinction shapes how you approach the work.
Consistency across signals is what raises that probability. If your website says one thing, your Google Business Profile says another, and third-party mentions describe your services differently again, you are introducing ambiguity that AI models resolve by picking a source - which may not be yours. Entity consistency across all touchpoints - the same name, address format, service descriptions, and category signals - reduces that ambiguity and improves the chance that Gemini draws from your content rather than a competitor's.
The same principle applies across other AI surfaces. ChatGPT and Perplexity also surface local business information, and both reward the same fundamentals: clear, accessible, consistently structured content that a language model can extract and verify against multiple corroborating sources. Treating local AI visibility as a Gemini-specific problem misses the broader pattern.
How to Measure Citation Performance Without Losing Your Mind
Given the inherent variability, measurement needs to be designed for probability, not position. That means running batches of queries - not single checks - across a representative set of local search intents relevant to your category. Log citation presence or absence across each run, and track that over time. You are looking for trend lines, not snapshots.
On the traffic side, AI-referred sessions from Gemini will show up with referral sources you need to actively identify and tag in GA4. Direct traffic absorbs a meaningful proportion of AI-referred visits because users click a citation, land on a page, and the referrer is not passed cleanly. Setting up custom channel groupings in GA4 to capture known AI referrer strings helps separate this traffic from genuine direct visits. Then connect those sessions to lead events and, where your CRM allows it, to downstream revenue.
The Gemini citation data reinforces something that has been true across AI search platforms: owned content quality and accessibility are the controllable inputs. Citation frequency is the output, and it will vary. The businesses that earn citations consistently are the ones that have made their websites the most reliable, extractable source of accurate local information available - and have the measurement infrastructure to know when that is working.