Steady Demand Local Citations Study (2026)

What the study counted, and one choice made before counting

The Steady Demand local citations study is a single-snapshot audit of what Google's two generative surfaces put in their citation lists when the query is about a local service business. Ben Fisher published it on 22 August 2026 as AI Overviews and AI Mode both cite local businesses in almost opposite ways. It completed 7,990 queries across both surfaces from a grid of 50 US metros, ten home-service and professional verticals, and eight query templates split evenly between local and informational intent. It classified 86,645 citations into eight source categories and crawled 25,898 cited pages.

That is a census of what was cited. Nothing was manipulated and no uncited business was measured. Every finding has the form among the pages that were cited, X% had property P, and that shape decides which conclusions survive.

One classification choice carries more weight than any number in the paper. The taxonomy puts Google Maps and Business Profile URLs in a bucket of their own, separate from the business's own site. Whether a Maps URL counts as the business or as Google is what the headline claims to answer, and the taxonomy answers it by definition before a citation is counted. The choice is defensible, but part of AI Mode cites Google, not you is a restatement of it.

48.8% against 100% is a definition, not a measurement

The figure that travels furthest is that AI Overviews produced an answer on 48.8% of queries and AI Mode on 100%. Read as a comparison of how aggressively the two surfaces generate, it is an artifact of framing.

AI Overviews is a module Google's systems decide whether to insert into a results page that exists either way, so 48.8% measures a real selection decision. AI Mode is a surface a user deliberately enters, and answering is what it is for, so a 100% appearance rate is close to a restatement of what the surface is. The two percentages answer different questions and are printed side by side as though they answered one.

The result worth keeping is inside AI Overviews. On local-intent queries it appeared 21.8% of the time; on informational queries, 75.8%. One surface, one selection mechanism, a 54-point swing driven by nothing but query intent.

That split is also the one finding here with independent support. Whitespark, a different vendor, collected 540 queries by hand in May 2025 and reported AI Overviews on 15% of local-intent queries and 92% of informational ones: a fraction of the sample, and the same shape.

The finding that holds: Google Maps takes four fifths of the slots

On local-intent queries in August 2026 the two surfaces drew from different pools. AI Overviews sent 73.5% of its local-intent citations to the business's own website, out of 3,869 citations. AI Mode sent 79.8% to Google Maps and Business Profile URLs and 19.7% to business sites, out of 40,883. That is a large, well-sampled descriptive finding about an undocumented system, published as the Steady Demand local citations study, with a blunt operational consequence: for those queries the Google Business Profile is the surface being read, not the website.

It is unreplicated: no other published study classifies AI Mode's local-intent citations by source type. It is consistent with what does exist. SE Ranking's February 2026 measurement of 1,321,398 AI Mode citations found google.com the most-cited domain at 17.42%, of which 36.1% were Business Profiles, close to 6.3% of all citations on a general keyword set, which is what you would expect if Maps citation concentrates in local intent. Google self-citation is growing fast, so a snapshot of it has a short shelf life.

One more number deserves attention. Of 6,006 business-site citations checked against the domain Google's own Places record lists as official, 95.5% matched. The other 4.5%, roughly one in twenty-two, pointed elsewhere, attributed to old domains, franchise-naming quirks or genuine errors with no breakdown published. Those cost different amounts: a dead domain costs a click; a franchise citation routed to the national site loses the lead inside the same brand. The check is agreement between two Google systems, not ground truth.

Three gaps of exactly 5.3 points

The page-signal table compares cited pages on five attributes: schema markup present, 64.0% for AI Overviews against 58.7% for AI Mode; mobile-friendly, 87.2% against 81.9%; concrete statistics, 59.5% against 54.2%; visible phone number on local intent, 68.9% against 58.4%; social profiles linked, 39.2% against 34.3%. Five signals all favoring AI Overviews reads like a preference for better-built pages. It almost certainly is not.

Start with the arithmetic. Three of the five gaps are exactly 5.3 percentage points, across signals whose base rates run from 54.2% to 87.2% with no mechanical relationship to each other. Five independent preferences do not land on one identical value. What produces that fingerprint is one block of pages present in one column and absent from the other, failing all three together, since pages tend to carry the whole modern-build package or none of it. No counts or dataset were published to rule out an artifact of assembly.

The composition problem is visible without the arithmetic. The method restricts technical signals to business-site citations where noted, and this table carries no note, no sample size and no caption, so a reader cannot tell whether google.com URLs sit inside it. Read charitably as business sites only, the AI Overviews column describes pages chosen out of 73.5% of its citations while the AI Mode column describes the residue, the 19.7% still cited after Maps took four fifths of the slots — two selection processes at very different selectivity, not two judges scoring one population.

The link data settles the direction. Far from citing weaker domains, AI Mode's cited domains carry 428 median referring domains against 296 and 2,824 median backlinks against 1,562. If AI Mode's surviving business-site slice skews toward multi-location brands, which is what survives when a Maps URL is the default, their corporate pages are less likely to carry a local phone number, and the phone gap is the largest in the table. A single difference in the kind of page cited explains every figure.

Every business in the sample had already been cited

The reputation block reports a 4.77 average star rating against 4.75, 1,129 reviews against 1,266, sentiment 0.71 on both, median domain age 21.5 years on both, and Business Profile completeness above 98%. The conclusion drawn is that neither surface cares much about your reputation.

That does not follow, for structural rather than arithmetic reasons. Every business measured was already cited, and the sample narrows further to the 1,000 most-cited businesses. The table shows only that cited businesses on one surface look like cited businesses on the other. Whether reputation influenced selection cannot be read from it, because how uncited businesses score was never collected.

The standard names apply. Selecting on the dependent variable means the sample is defined by the outcome, so the outcome cannot be explained by it. Restriction of range is the mechanism: conditioning on a high-reputation subset collapses the variance, and collapsed variance guarantees a null difference. The design supports the probability of a rating given that a business was cited; the claim requires the probability of being cited given a rating. Those are different quantities, and no base rate was collected to bridge them.

Fisher's own sister study makes the competing reading live: the AI Citation Ledger reports about 90% of AI picks at 4.5 stars or better, which is what a cleared threshold looks like rather than an irrelevant variable. The same objection applies to the identical 21.5-year median domain age. This is a common and easy error, not a dishonest one, and the same Places pull against uncited businesses would settle it either way.

The opposition is conditional on intent

On local-intent queries the headline earns itself: AI Overviews at 73.5% business sites against AI Mode at 79.8% Google properties is close to a mirror image. On informational queries it collapses. AI Mode's Google self-citation falls from 79.8% to 22.7%, a drop of 57 points, and its business-site share rises from 19.7% to 47.5%. On an informational query AI Mode behaves less like an opposite than like a citation engine pointed at the open web.

The precise statement is that the surfaces diverge sharply on local-intent queries and converge on informational ones. That points to one system with a local-intent override rather than two philosophies of citation: when the query wants a business, AI Mode reaches for the structured Places record; otherwise it reaches for documents.

Credit where due: the study publishes the 22.7% and 47.5% figures itself and hedges the headline with almost. The qualification is simply not in the title, and the title is what gets quoted. One gap keeps this from being fully checkable: the AI Overviews informational-intent split is not published, so convergence rests on AI Mode's numbers moving toward where AI Overviews sat on a different query population.

Google documents none of the mechanism being measured

Checked in September 2026, Google documents nothing about how AI Mode selects local business citations or whether it uses Maps or Business Profile data at all. Search Central's page on AI features and your website, last updated 10 December 2025, describes neither as an input; Business Profile appears only as generic hygiene advice. The AI Mode help documentation gives no selection criteria for local results.

Silence is itself a finding. Every behavioral claim in the study is Observed at best and can change without notice. It also bears on the signal table: the same Search Central page states there are no additional requirements to appear in AI Overviews or AI Mode, and that publishers need not create new machine-readable files or markup for these features. That sits squarely against reading a 64.0%-versus-58.7% schema gap as a Google preference.

Who published it, and what the funding decided to measure

Steady Demand sells Google Business Profile management, Local Services Ads management and AI optimization services for local businesses. The study concludes that you are now running two campaigns, a content campaign for AI Overviews and a Business Profile campaign for AI Mode, and those map onto two of the publisher's service lines. No funding statement appears on the study or its research index. Worth stating once, plainly, and reading for what it implies rather than as disqualifying.

Set against it: the limitations section is unusually candid for agency research. Six are named, two of which undercut the study's own headline table: no JavaScript rendering in the crawl, and sampling rather than exhaustive coverage on the signals the reputation conclusion rests on. The method is published in enough detail to be checked, which is the only reason the criticism above is possible.

The useful question is what the funding decided to measure and what it left unmeasured. The design measured what is cited, not what changes citation: no intervention, no before-and-after, no control group. Collecting Places data on uncited businesses would have cost one more API loop and tested whether reputation and profile completeness discriminate. That is the study that is missing.

Frequently asked questions