What happened: AI assistants failed to recommend 85.6% of real local businesses in a new census-style audit from Norly Research. The study enumerated every restaurant, cafe, and bar across two Bali markets, 4,776 venues in total, then ran 2,208 search-grounded queries through ChatGPT, Claude, Gemini, and Perplexity. Having a website more than doubled a venue’s odds of being recommended, roughly a 1.92x lift, while star ratings had no effect on whether a venue was recommended at all, only on its ranking once it made the list.
Why it matters: Most prior AI-visibility research works backward from a curated list of known brands and checks whether AI systems mention them. Norly’s approach inverted that: it started from a complete market census, including businesses no one thought to ask about, and found the miss rate was even higher for that full population than for brand-focused studies typically report. That is a materially different starting point for a marketing leader trying to size the actual scope of the AI-visibility problem rather than just checking whether their own brand shows up.
The original insight: The gap between having a website and getting recommended is the more actionable finding than the overall miss rate. It suggests AI assistants are still leaning on crawlable, structured web content as a baseline signal for local recommendations even as they draw on other sources, which means a venue’s own site remains a lever a business fully controls, unlike star ratings or third-party directory listings it does not. That reinforces a pattern MarTech has tracked since the grounding-drift research on AI citation instability, and lines up with earlier findings that AI overviews already cite a different web than organic search ranks. Norly discloses that it sells AI-visibility and review-management tools, a competing interest the firm states plainly rather than obscures.
Source: Norly Research
