A new analysis of local AI search finds a clear signal and an inconvenient reality: Google’s Gemini most often cites businesses’ own websites, but its recommendations shift so frequently that a single query can’t be treated as reliable evidence of visibility.
What the dataset shows
Steady Demand ran 1,487 local queries across 50 major U.S. metro areas and 10 service categories, collecting 14,472 citations returned by Gemini. Nearly 60% of those citations pointed directly to a business’s website — more than the combined total from directories, review platforms and forums.
Reddit emerged as the second-most-cited domain category, accounting for 13.7% of citations and outperforming the combined local-service directory category (examples: Angi, Thumbtack, HomeAdvisor).
Repeatability — or lack of it
Steady Demand flagged inconsistent output as the study’s most consequential finding. The firm’s co-founder Ben Fisher called the effect “Grounding Drift.” Repeating the same Gemini query produced overlapping cited sources in only about 40% of cases.
For comparison, a control test using Google’s traditional local pack returned the same top listing roughly 90% of the time. Gemini recommended the same top business only about 7% of the time.
How Gemini compares with ChatGPT
The researchers repeated the same queries in ChatGPT to compare citation patterns. Across the two platforms, both cited many of the same domains only 8% of the time, and they recommended the same top business in just 4.2% of cases.
The platforms showed different source preferences: Gemini leaned toward business websites, while ChatGPT cited Reddit and business directories more heavily.
Why this matters for marketers
Many local marketers use AI outputs as a proxy for visibility or reputation. The study shows that approach is risky: different AI models surface different evidence and rarely agree on a single top result. Treat any individual AI answer as a snapshot, not a stable measure of long-term visibility.
Actionable takeaways for local businesses and agencies
1) Treat AI search outputs as one input, not the final metric. Use AI citations to flag potential visibility gaps or content issues, then verify across additional platforms.
2) Prioritize crawlable, authoritative content on your own website. Because Gemini favored business sites in this dataset, accurate service pages, location pages and structured data (schema) increase the chance AI models will find and cite your business.
3) Maintain third-party listings and monitor forums. Directories, review sites and social platforms still influence how some models ground answers. Keep NAP (name, address, phone) consistent across listings and watch forums where customers discuss services.
4) Monitor multiple AI engines. Because models pull evidence from different sources, tracking only one provider can give a misleading picture of overall AI visibility.
5) Run controlled, repeatable tests. Capture citations over time and under controlled conditions to detect trends rather than relying on one-off queries.
Limitations and what to watch next
Steady Demand’s dataset is substantial but represents a sample of metros, categories and queries. The public summary does not include full methodological details such as the query list, time windows or exact repeat counts, which limits how broadly the results can be generalized.
Practitioners should watch whether models shift toward more owned-site citations or greater use of third-party content, and for any improvements in AI transparency or APIs that make citation auditing easier at scale.
Short term, the practical implication is simple: AI will reshape how users discover local businesses, but it increases the number of places you must manage. Start by diversifying measurement, strengthening authoritative website information, and building repeatable monitoring across multiple AI models to understand where your business is actually being cited.
