There are no local AI-search statistics worth quoting. This page is graded Broken chain — a structural gap rather than a degraded citation chain: we located no disclosed-method study comparing AI citation behaviour on local-intent queries against an informational baseline, on any engine. Local intent is a large share of all search volume, so the absence is a blind spot in the field rather than a niche omission. What follows is the shape of the gap, why it persists, and the measurement spec that would close it.
- No public study of AI citation for local-intent queries with a disclosed method was located. The honest figure for "how AI answers treat local businesses" is not a number, it is an absence.
- Local intent is recognised in retrieval design: "locations" is 1 of the 8 sub-query types named in Google's granted fan-out patent, US11663201B2.
- The likely reason engines diverge is data access, not model behaviour. An engine with a maps dataset can build a candidate set directly; one without has to infer it from ordinary crawled pages.
- The one recommendation that survives every version of this uncertainty: a plain service-area page on your own site. It is the single source every engine can reach without special data access.
- Everything on this page about how engines differ is inference from architecture, tiered as hypothesis. None of it is a measured finding.
The gap: there are no local AI-search statistics
Every research field has blind spots that reflect what is easy to study rather than what matters. AI-search citation research is heavily weighted toward broad informational and commercial queries, because those map cleanly onto a fixed query set run through a chat interface. The large public studies — Semrush's AI Overviews work, Ahrefs' overlap analysis, Profound's citation patterns — are all built on query sets of that first kind, as are the engine-level pages here on Claude, Gemini and Bing Copilot.
Google's documented query fan-out mechanism explicitly includes "locations" as one of eight sub-query type categories — set out in the granted patent, US11663201B2, and described publicly in Google's own comments on fan-out. Local intent is architecturally recognised in at least one major AI surface's retrieval design.
How citation patterns for local-intent queries actually differ from informational queries — on any engine — has not been measured with a disclosed method that we located.
sub-query types in Google's fan-out patent is 'locations' — local intent is recognised in the retrieval design itself.
How local-intent citation patterns differ from informational ones, on any engine, with a disclosed method.
| Local query category | Google AI Overviews / AI Mode | ChatGPT | Perplexity | Claude | Copilot |
|---|---|---|---|---|---|
| Restaurants and cafes | None located | None located | None located | None located | None located |
| Emergency trades | None located | None located | None located | None located | None located |
| Regulated professions | None located | None located | None located | None located | None located |
| Retail with stock | None located | None located | None located | None located | None located |
| Services without premises | None located | None located | None located | None located | None located |
"None located" means what it says: we found no study with a disclosed method reporting local-intent citation behaviour for that engine and category. It does not mean nobody has measured it privately. It means nothing checkable is public, which is why this page is graded Broken chain — the chain from claim to evidence has a structural gap in it, not a weak link.
Local-intent queries are a large share of all search volume. AI-search citation research on them is almost entirely absent. That's a genuine structural blind spot, not a niche afterthought.
Share on XWhat a real local measurement would require
Part of why local is under-researched is that it defeats the standard method. A fixed query set run once through a chat interface produces a result that describes one place, one moment and one way of supplying a location. Any study that wants to say something general has to control for all three. Here is the minimum specification.
| Requirement | What it means in practice | Why omitting it invalidates the result |
|---|---|---|
| A geo-varied query set | The same customer questions, repeated across multiple metro areas — ideally including at least one market outside the engine's home country | Two researchers running an identical query in different cities get different answers and neither is wrong. A single-metro result is a case study, not a statistic. |
| Explicit location controls | Record how location was supplied on every run: typed place name, account setting, or inferred device position — and hold it fixed within a comparison | Typing a neighbourhood and letting a system infer your position may reach different data entirely. Mixing the two produces an uninterpretable blend. |
| Category stratification | Report restaurants, emergency trades, regulated professions, stocked retail and premises-free services separately | The dominant signal differs per category. A blended "local" average describes nothing that exists. |
| Repeated runs, clean sessions | Several runs per question, each in a fresh unauthenticated session, with variance reported alongside the mean | These systems are non-deterministic and personalised. One run per query measures noise. |
| Null answers recorded | Log runs where no business is named, or the engine declines | Dropping nulls inflates every rate computed from the sample. |
| Timestamps and a short cycle | Date and time on every observation, with re-runs inside weeks rather than quarters | Opening hours, closures and seasonal demand move local answers within a single week — a much faster cycle than the general freshness effects claimed for AI search. |
| Mention and citation counted separately | Named in prose is one column; linked as a source is another | They are different outcomes with different value, and collapsing them is the most common way a local claim gets oversold. |
There is one more problem no protocol solves: the ground truth is contested. There is no correct list of the best coffee shops in a neighbourhood, so accuracy cannot be scored the way it can for a factual question. That is a reason to measure source composition and engine agreement rather than correctness. A local-intent subset of the Citation Index's query set, run to this spec against the informational baseline already collected under the measurement standard, is registered as a candidate expansion rather than a running study.
Why local queries are architecturally different
Local intent likely deserves its own research track because of a genuine data-access asymmetry between engines. Google's AI surfaces can plausibly draw on Business Profile listings, Maps data and location-specific review corpora built over more than a decade. An engine without that infrastructure has to rely on whatever local information happens to be described in content reachable by its ordinary retrieval bots — OpenAI's and Perplexity's are documented, and all are catalogued in the bot registry. That is a structurally different and probably much thinner source base for the same query.
- 01 Locate the user Device signal, account setting, or the words of the question. Each route gives different precision.
- 02 Build a candidate set A maps dataset supplies it directly. Without one, it has to be inferred from ordinary web pages. This is where we expect the sharpest divergence.
- 03 Rank or filter Distance, rating, hours, relevance. Which inputs and in what order is undocumented for every engine.
- 04 Write and attribute The named source may be a listing, a review site, a local publication, or the business itself.
Hypothesis Step two is where we expect engines to diverge most sharply, because it depends entirely on data access rather than model behaviour. That is untested, and step three may matter more. Nothing in the chain runs at all if crawlers never arrive, which puts robots.txt policy and the training-versus-retrieval split upstream of all four steps.
Why category changes the question
| Category | Likely dominant signal | Why it complicates measurement |
|---|---|---|
| Restaurants and cafes | Reviews, photos, opening hours | Answers change with season, day and time of day |
| Emergency trades | Availability, coverage radius, licensing | Urgency queries may bypass discovery entirely |
| Regulated professions | Credentials, registration, jurisdiction — see YMYL constraints | An engine may be cautious and name institutions rather than firms |
| Retail with stock | Inventory and price | The useful answer is a live fact no static page carries — the problem agentic commerce runs into, as Previsible describes |
| Services without premises | Service-area description on the site itself | There may be no listing to find at all |
A single blended "local" figure would average across all five rows and describe none of them. Category stratification is not a refinement here; it is a precondition for reporting anything.
Which local tactics do not transfer between engines
The central claim of this page is a transfer failure, so it is worth being precise. Classic local SEO has an established playbook — claim and optimise a Business Profile, accumulate reviews, keep name, address and phone data consistent — built around a well-understood Maps ranking system. Much current "local GEO" advice extends that playbook to AI citation by assumption. It may turn out correct for Google's own surfaces. It says nothing about an engine without that data access.
| Local tactic | Assumes the engine can… | Fails when |
|---|---|---|
| Business Profile optimisation | Read a maps listing directly | The engine has no such dataset — no mechanism for the tactic to work through |
| Review accumulation | Reach the review corpus | It only sees review pages it happens to crawl |
| Proximity targeting | Know the user's position precisely | The query arrives in a desktop chat window with no location signal |
| NAP consistency across directories | Crawl those directories at all | Rarely — which is why this one survives most cost-benefit arguments |
| A plain service-area page on your own site | Crawl an ordinary web page | Almost never — the reason it is the one recommendation this page makes without hedging |
Open question Whether any local tactic produces the same effect on two different engines is untested. Until it is, treat every local GEO recommendation as engine-specific by default.
What a local business can do right now
Keep doing the classic basics well — an accurate, complete Business Profile, consistent name, address and phone data everywhere you are listed, and a steady flow of genuine reviews. None of that goes to waste even if AI citation turns out to work differently than expected.
Beyond that, make sure your own website describes your services in plain, specific language — close to what Onely calls LLM-friendly content — and is reachable by every crawler, which means a working sitemap and content that survives without JavaScript.
Consider the two-van heating company whose site has a home page, a contact page and a gallery, and names no towns anywhere. An engine with maps access can still find it; an engine that only reads web pages has no crawlable sentence saying where it operates. Hypothesis Adding a short service-area paragraph, emergency hours and a licence number should make it describable by the second kind of engine. Untested — but the test we would apply to any local tactic while evidence is missing is simply: would you still do it if the AI effect turned out to be zero? A page a customer can read passes that test.
The costs worth naming. Effort on the wrong surface is pure waste, so check which engines actually send you anything using your own referral segment before the population-level market share figures. Aggressive review chasing creates certain compliance and reputation risk against an uncertain AI benefit. Vendor guarantees of an AI listing deserve exactly the scepticism a guaranteed ranking does. And the largest risk of all is spending on a speculative channel while your opening hours are wrong on half your listings.
Measuring this for your own business
A small project, not a research programme — one owner, an afternoon a month, applying a cut-down version of the spec above.
- 01 Write ten customer questions Mix categories. Include one price, one availability and one comparison question.
- 02 Fix the location wording Same phrasing every time, written down. Change one thing at a time later.
- 03 Three runs per engine Fresh session each time. Non-determinism is the default, not the exception.
- 04 Record four things Named? Linked? Which competitors appeared? Which sources were named?
- 05 Study the sources A directory or local publication that keeps appearing may be easier to reach than the engine.
- 06 Repeat monthly The trend needs at least three months before it means anything.
- 07 Change nothing for two months You need a baseline before you can attribute anything to a fix.
Confounds in any local result
If your appearance in local answers changes, several ordinary things could explain it before any AI effect does. A competitor opened or closed, so the candidate set moved underneath you. Your listing data changed — hours, category or address edits, including ones you did not make. Seasonality shifted the question mix.
A third-party page changed: if a local roundup that named you was updated, an engine relying on it may stop naming you, the same single-point-of-failure dependency that makes Reddit and Wikipedia so consequential elsewhere. The citation aged out — persistence is itself unmeasured, and the half-life study takes up that end of it. Or the location signal moved a fraction and crossed a boundary.
Common misreadings
"Local does not matter for AI search." The opposite. The volume is large and the research is missing. Those are different statements.
"My Business Profile is my AI strategy." It may be most of it on one surface and none of it on another. Assuming one surface is the whole world is exactly the error this page describes.
"I was named once, so it works." A single appearance tells you almost nothing given how variable these answers are between runs.
"An AI answer replaces the map pack." Both may appear, and they may draw on different data — AI Mode and AI Overviews do not even agree with each other, as far as anyone has been able to check. Watching one and ignoring the other gives a partial picture.
Start with the one thing every engine can reach: check what AI Overviews currently return for your main service query, record the date and the exact location wording, and repeat it next month. If the local slice of the Citation Index runs, results — including nulls — ship through the newsletter.
Null results we would publish
- No difference between local and informational citation patterns. If local queries behave like everything else, the case for a separate research track disappears and we will say so.
- No engine gap. If non-Google engines name the same businesses as Google despite lacking a maps dataset, the data-access argument on this page is wrong.
- No category effect. If stratifying by category changes nothing, the table above is decoration and should be removed.
- No listing effect. If businesses with complete listings appear no more often than those without, that would be an uncomfortable and highly publishable result.
Open questions
- Does an AI answer for a local question reduce clicks to businesses, to directories, or to neither?
- How does an engine handle a business that is open now versus one that is closed, and does it check?
- Do local answers name a smaller set of businesses than a map result would, and is that set stable?
- Does language change the source mix for the same place, and by how much?
- Are new businesses systematically absent from AI answers for longer than they are from classic local results?
- Do open-source crawling tools, of the kind surveyed in open-source SEO agent tooling, handle location input consistently enough to build a shared local dataset on?
Limitations
- This page identifies a gap; it does not close it. No first-party data on local AI-search citation exists on this site today, and the grade here reflects an absence of research rather than a broken citation chain.
- The architectural argument is inference. No operator has documented whether or how a local dataset feeds its answers. Google's AI optimization guide and crawler documentation are the closest first-party material and neither addresses it.
- No figure is offered for the size of the local AI opportunity, because no defensible one exists. Terminology and grading conventions are in the glossary and about.
Namdev, R. (2026). Local AI search statistics: the gap where the numbers should be (v1). Retrieved from https://ritiknamdev.com/blog/local-ai-search-statistics Published under CC BY 4.0 — reuse freely with attribution.
Complements Google AI Mode statistics, which documents "locations" as one of the eight patent-defined sub-query types fan-out can generate.