AI Overviews and Local Search: What Actually Gets Cited (and How to Be the Citation)

"Will AI Overviews kill my website traffic?" is the wrong question for a local business to ask. The question that actually matters is narrower, and we can answer it: when an AI assistant recommends a plumber, a dentist, or a landscaper near you, whose website does it point to?

We ran our own test instead of repeating another industry stat with no sample size attached. VeloRank probed eleven live, web-grounded AI queries, phrased the way a real customer would ask them — "Who is the best plumber in Eagan, Minnesota?", "Best HVAC repair company in Rosemount MN" — and read every source cited in each answer. This post is what we found, and what a business can do about it.

One thing to be upfront about. We didn't query Google's AI Overviews panel directly; there's no public API for that. We tested a live, web-grounded frontier model instead, the same family of grounded answer engine that AI Overviews, AI Mode, and tools like ChatGPT and Perplexity all draw from. Treat this as a close stand-in for how AI recommends local businesses, not a screenshot of Google's own feature. We come back to that distinction, and to what else this data can't tell you, in the limitations section below.

How we counted

Each query got one AI-generated answer, and we pulled every source it cited. If the same URL showed up twice in one answer, we counted it once — the unit here is unique cited sources per response, not raw citation mentions, which would inflate the count for whichever directory happened to get name-dropped three times in a paragraph. We then read each source and sorted it into one of three buckets: the recommended business's own website, a third-party directory or listing platform, or "other," a small catch-all for things like manufacturer pages. No source got counted twice, and no query got asked twice.

Who gets cited when AI answers a hiring question

Across all 11 queries, we counted 41 unique cited sources and classified each one.

Where the citation pointed	Share	Count
Third-party directories (BBB, Angi, HomeAdvisor, listing portals)	68.3%	28
The recommended business's own website	24.4%	10
Other (manufacturer sites and similar)	7.3%	3

n = 41 unique cited sources across 11 queries, south-metro Minnesota cities and trades.

More than two out of three citations went to a directory instead of the business itself. That doesn't mean the business disappeared from the AI's answer. It means the recommendation got filtered through someone else's page, someone else's layout, someone else's ad slot sitting above the listing that's actually you.

The two queries not in the trade breakdown

Two of the eleven queries don't appear in our trade-by-trade breakdown, so here's the first look at them. The gap between these two groups isn't abstract. An own-site citation usually comes with a link straight to the business — its phone number, its actual photos, its own words about the job. A directory citation routes the same customer through a listing page built by someone else, with someone else's ads and someone else's competing options one scroll away.

"Best hair salon in Rosemount, Minnesota" returned 3 of 6 unique sources pointing to the salon itself. "Best butcher shop in Dakota County, Minnesota" returned 2 of 3 own-site citations.

At the other end, four queries returned nothing but directory citations:

Query	Own site	Directory	Other
"Recommend a good landscaping company in Burnsville, MN"	0	3	0
"Who is the best roofing company in Lakeville, Minnesota?"	0	2	0
"Who is the best handyman in Eagan, Minnesota?"	0	4	0
"Who is the best real estate agent in Rosemount, Minnesota?"	0	2	0

Why directories keep winning

The likely reason isn't complicated, and it's the same pattern we found looking at this from the other direction. In a separate crawl of 922 live Minnesota business websites, 87.7% had no LocalBusiness schema markup — the structured data that tells a search engine, or an AI model reading the page, what the business is, where it operates, and what it does.

An AI model assembling an answer grounds itself in whatever text it can read and trust fastest. A directory listing is built entirely from structured fields: business name, category, address, hours, review count. A business's own homepage, more often than not, is a hero photo and a phone number with nothing machine-readable describing the actual service. Given a choice between a page built for exactly this and a page that isn't, the AI takes the easier read.

There's probably a second, quieter effect too, though we can't measure it directly from this probe. A directory aggregates dozens of businesses under one domain with a long history of being crawled and cited, which can function like a shortcut for trust even when the model has no real way to verify any single listing on it. A single local business competing against that may start from behind on trust signal alone, before content quality even enters the picture. Schema markup and specific content won't necessarily erase that gap, but they're the two levers an individual business actually controls.

This isn't really a visibility problem, and that's worth sitting with. In the same research, the median Minnesota business website already ranks for 707 keywords in Google (any top-100 position, 138-site sample) — mostly brand terms and long-tail phrases, not necessarily meaningful visibility, but not nothing either. These businesses aren't buried. They're indexed, ranked for something, and then passed over at the exact moment an AI model has to decide who to actually name in its answer.

How to become the citation instead of the footnote

1. 1

   Add LocalBusiness schema. It's the single factor most directly tied to what an AI model can read and quote. VeloRank's free schema generator builds the markup in a few minutes — the 87.7% figure above is the gap it closes.

2. 2

   Write for the actual question, not the service category. "HVAC repair" is a category page. "Furnace repair Rosemount same-day" is something an AI model can quote directly in an answer. Match page content to how customers actually phrase the question, not how you'd label a menu item.

3. 3

   Name the city on the page, not just in a service-area list. Every business shut out in our worst-performing queries had thin or generic service pages. A dedicated page per city you serve gives the AI something specific to ground its answer in, instead of one paragraph covering ten towns at once.

4. 4

   Check yourself. VeloRank's free AI visibility checker runs a live query for your trade and city and shows whether the AI cites your site or the directory sitting above it.

AI Overviews and AI Mode are two different problems

This post is about one narrow question: when AI cites a source, is it you? For the wider shift — how a single local search now returns three different sets of businesses instead of one, and what that does to the person searching — see our companion piece on Google AI Mode and what it means for Minnesota local search. That post covers the searcher's experience. This one covers the citation itself.

Both point at the same fix, from different angles. Our 2026 local ranking factors analysis rates traditional signals like GBP and reviews "Critical" and AI-era signals "Medium and rising," but they aren't competing systems. The clean, structured, specific site that wins the citation is largely the same site that wins the local pack.

Limitations, stated plainly

Eleven queries is eleven data points. We're calling this a probe, not a survey, because that's what it is: no query was repeated, most cities have one or two trades tested (Rosemount has three: HVAC, hair salon, real estate), and the whole batch clusters around six south-metro Minnesota locations, not the state. A different model, a different day, or a different phrasing could return different citations for the same question, since these answers are generated fresh each time rather than pulled from a stable index.

We also didn't test Google's AI Overviews panel directly, for the reason stated above: no public API exists for it. What we tested is a proxy — a live, web-grounded frontier model answering the same kind of question. It's a reasonable stand-in for how a grounded answer engine picks sources. It isn't a guarantee that Google's specific AI Overviews implementation behaves identically.

Classification is also a judgment call in a handful of cases — a franchise location page, for instance, sits somewhere between "own site" and "directory," and reasonable people could sort it differently. We made a call on each one and moved on rather than pretending the line is always clean.

The full methodology, including how sources were classified and counted, lives in our Minnesota local SEO statistics study, the parent research this probe is drawn from.

Your next move

Run the free AI visibility checker to see who gets cited for your trade and city right now. If it's a directory instead of you, the schema generator and the ranking-factors guide above are the next two stops.

Questions about this research? Reach out at hello@velorank.net.