Gemini's local AI search cites business websites most

When Google’s Gemini answers a local query, it cites the business’s own website far more often than directories or review platforms. A new study from the local SEO agency Steady Demand analysed 14,472 AI citations across 1,487 queries in 50 of the largest U.S. metros and found that business websites account for 60% of all citations, and 59.9% of Gemini’s. The same study found AI recommendations are strikingly inconsistent: ask the identical question twice and the cited sources overlap only about 40% of the time. For multi-location brands, the signal is plain. Your own location pages are the asset AI reads first.

What happened

Steady Demand, a US agency led by co-founder Ben Fisher, a Google Business Profile Diamond Product Expert, published its AI Citation Ledger: an analysis of 14,472 citations drawn from 1,487 queries across 50 U.S. metros and 10 local-service categories. The headline finding is how much weight AI places on a business’s own domain.

Business websites drew 60% of all citations across both engines studied. Gemini leaned on them hardest at 59.9%, while ChatGPT cited business websites only 15.9% of the time and relied more on Reddit and directories. Reddit alone accounted for 13.7% of citations, more than every local-service directory combined at 10.3%. The two engines rarely agreed with each other: their cited domains overlapped just 8% of the time, and they named the same top business in only 4.2% of cases.

The study’s second finding is about consistency, which Steady Demand calls Grounding Drift. Asking Gemini the same question twice, back to back on the same day, produced overlapping sources 46.3% of the time, falling to 41.0% a few hours later and 26.5% the next day. Gemini named the same top business on roughly 7.9% of repeat queries, about one in thirteen. Google’s traditional local pack, used as a control, returned the same top listing 90.2% of the time.

“Asking the same question twice doesn’t run the same search twice. Each call is its own independent, evidently stochastic decision about what to type into Google on your behalf.”

Steady Demand, AI Citation Ledger

Why it matters

Two things follow for anyone responsible for local visibility at scale. First, owned web content, not just a Google Business Profile, is what AI reads when it composes a local answer. A brand’s location pages are doing more work than ever, because the engine is quoting them directly. Second, AI visibility is volatile in a way traditional rankings are not. When the same query yields different sources hours apart, a one-off spot check tells you almost nothing, and any measurement has to be built on repeated sampling over time.

There is a quality signal buried in the data too. Businesses that AI recommended averaged 4.75 stars, close to the 4.84-star average of plain Google results, and 97% sat at 4.0 stars or higher. AI is not surfacing weak performers; it is drawing from strong, well-reviewed businesses with credible owned content.

What this means for multi-location brands

For a brand running hundreds or thousands of locations, the takeaway is a governance one, not a per-query one. If business websites win most AI citations, then every local landing page and store-locator page across the estate has to be accurate, crawlable, and specific enough to be worth quoting. A thin or duplicated location page gives the engine nothing distinctive to cite. This is the same discipline that underpins generative engine optimization: correct data and genuinely local content, maintained consistently everywhere.

The practical response is to control the inputs rather than the black box. Keep a single source of truth for location data across every profile with a local business listing platform, keep each location’s pages content-rich and current through a store locator built for scale, and feed AI-ready profiles through Places AI. Because AI answers drift query to query, do not measure success on a single check: track presence and outcomes across locations over time, and treat the goal as being the obviously credible answer for a real customer in a real place, whatever the engine returns on any given run.

The bottom line

The engines that increasingly stand between customers and local businesses are quoting those businesses’ own websites above almost everything else, and they are doing it inconsistently. For multi-location operators the answer is not to chase a moving target but to own what AI reads: accurate data and distinctive local pages across every location, measured over time rather than in a single snapshot.