When AI Recommends a Local Business, It Reads the Website, Not the Reviews.

Short answer: your website. A new study of 14,472 citations found that when Gemini recommends local businesses, nearly 60% of its citations are business websites. Directories, review sites, and forums combined make up the rest. If your site is thin, you're handing the recommendation to whoever wrote more down.

What the Study Actually Found

Local SEO firm Steady Demand ran 1,487 local search queries through Gemini across the 50 largest US metro areas and 10 service categories, then tracked all 14,472 sources the AI cited in its answers (Search Engine Land, August 19, 2026). Three findings matter for a business owner.

Your website is the primary source. Business websites made up nearly 60% of everything Gemini cited. The runner-up wasn't Yelp or a directory. It was Reddit, at 13.7%. Directories and review platforms trailed behind that (Search Engine Land, August 19, 2026).

The answers are wildly inconsistent. The study calls it "Grounding Drift." Running the identical query twice produced overlapping cited sources only about 40% of the time, and the same top business recommendation just 7% of the time. For comparison, Google's traditional local pack returned the same top listing about 90% of the time (Search Engine Land, August 19, 2026).

Different AIs barely agree with each other. When the same queries ran through ChatGPT, the two platforms cited the same domains only 8% of the time and named the same top business 4.2% of the time (Search Engine Land, August 19, 2026).

Steady Demand co-founder Ben Fisher summed up the volatility plainly: "A single AI search should be treated as a snapshot, not a reliable measure of long-term visibility" (Search Engine Land, August 19, 2026).

Added October 7, 2026: for Google's own AI answers there is now a steadier measure than a single search. A free Search Console report counts how often your pages appear in them, and here is how to see if your site shows up in Google AI results.

Where This Is Heading

The money is already moving, even if the playbook hasn't. In a separate Search Engine Land survey analysis published August 24, 82% of marketers said they've committed budget to AI search visibility, averaging 24% of their budget, yet 81% still just call the work "SEO" (Search Engine Land, August 24, 2026). Translation: everyone knows customers are asking AI assistants for recommendations, and almost nobody has changed what they actually do about it.

Put the two studies together and the strategy writes itself. You cannot chase any single AI answer, because the answers reshuffle constantly. What you can do is strengthen the one source every answer keeps drawing from. That's not a directory profile. It's your website.

What to Actually Do This Week

Ask the AI about your own category, more than once. Try "best [what you do] in [your town]" in Gemini a few times over a few days. Don't panic over any single answer; at 7% repeat consistency, one bad result means almost nothing. Instead, look at which businesses get cited and open the pages the AI is reading. That's your competition's homework, visible for free.

Put the facts in plain text on real pages. Services, prices or ranges, service area, hours, who you're for. AI assistants cite pages they can read. They can't recommend what you never wrote down, and they can't parse what only exists inside an image, a PDF, or an Instagram post.

Keep your listings accurate, but stop over-investing there. Directories earned a shrinking slice of citations in the study. Keep your name, category, and description consistent everywhere so the models connect it all to one business, then put the real effort into the site itself.

One Thing to Skip

Don't pay anyone who guarantees you a spot in AI answers. With 7% repeat consistency inside one platform and 8% agreement between platforms, nobody can promise placement, and the survey data backs up your skepticism: marketers named buzzword-heavy pitches without explanation as the number one vendor red flag, and ranked case studies far above terminology fluency for credibility, 34% to 9% (Search Engine Land, August 24, 2026).

One more thing, since this blog practices what it preaches: this post was drafted by Ninth Degree's agent pipeline and edited by a human before publishing, and building websites that AI assistants can actually read, cite, and recommend is a service we sell.