How AI Answers Change by State: A Data Study

If you monitor AI search from a single national vantage point, you are measuring an answer that many of your customers never see. A dealership, law firm, or dentist can be the top ChatGPT recommendation in one state and completely absent in the next — and until recently, most AI-monitoring pipelines had no way to tell.

This is a question we kept hearing. Several teams building AI-visibility products asked for the same capability: the ability to target a specific US state or city when querying AI engines, because their own customers care about local recommendations. One described the exact failure mode — a prompt like “Where can I finance a Chevrolet vehicle near Lackawanna, NY?” returned different results in the ChatGPT UI (queried from New York) than through a national scraping vantage point. The local answer is the real answer, and they couldn’t capture it.

So we built state-level proxy targeting for cloro’s AI monitor endpoints, then ran a study to quantify how much it actually changes what the models say.

What state-level targeting does

Every cloro AI monitor endpoint takes a country parameter. State-level targeting adds an optional state parameter (a two-letter USPS code) that routes the request through residential infrastructure in that state. The AI engine then infers the user’s location from that egress point — exactly as it would for a real person in Sacramento or Buffalo — and localizes its answer accordingly.

curl https://api.cloro.dev/v1/monitor/chatgpt \
  -H "Authorization: Bearer $CLORO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "Where can I finance a Chevrolet vehicle near me?",
    "country": "US",
    "state": "NY"
  }'

State targeting is supported on the non-Google AI answer engines — ChatGPT, Gemini, Perplexity, and Copilot — and costs +2 credits per call. (Google AI Mode and AI Overview already localize via UULE, Google’s own location mechanism; we included AI Mode in the study as a Google-side comparison.) The state parameter is only valid when country is US — as one of our early tests confirmed, an unsupported combination returns a clear validation error rather than silently ignoring the location.

How we tested

We built a set of eight location-neutral but locally-answerable prompts — the phrasing never names a city, so the only geographic signal is the state proxy. This isolates what state targeting alone does to an answer. The prompts span the verticals our customers monitor:

Vertical	Prompt
Auto	”Where can I finance a Chevrolet vehicle near me?”
Legal	”Who is the best personal injury lawyer near me?”
Insurance	”What is the best homeowners insurance company near me?”
Home services	”Find a top-rated HVAC repair company near me”
Real estate	”Who is the best real estate agent near me?”
Healthcare	”Recommend a highly rated dentist near me”
Finance	”What is the best credit union to join near me?”
Fitness	”What is the best gym near me?”

Each prompt was run through five engines — ChatGPT, Gemini, Perplexity, Microsoft Copilot, and Google AI Mode — across six states chosen for regional and market spread: California, Texas, New York, Florida, Illinois, and Washington. We also ran a national baseline (country=US, no state) for every prompt/engine pair to measure the delta between “no location” and “state-targeted.” That’s 279 successful responses, collected on 2026-07-23.

For each response we extracted the set of cited-source domains and measured three things:

• Cross-state divergence — one minus the mean pairwise Jaccard overlap of the cited-domain sets across the six states. A score of 1.0 means no two states cited the same source; 0.0 means every state cited an identical set.
• Overlap with the national baseline — the mean Jaccard overlap between a state’s citations and the no-location answer. Low overlap means state targeting changed the answer a lot.
• Localization rate — the share of state responses that cited at least one state-exclusive domain (a source that appeared in exactly one of the six states), a conservative, objective proxy for “the model named something genuinely local.”

We deliberately scored the cited sources rather than trying to parse business names out of prose: domains are unambiguous, and citations are exactly what an AI-visibility team needs to track.

Finding 1: five engines, five very different local behaviors

The headline table. Every metric is computed over the state responses only, except “overlap with national,” which compares each state to the no-location baseline.

Engine	Localized	Cross-state divergence	Overlap with national	Avg sources
Microsoft Copilot	96%	1.00	0.01	5.9
Google AI Mode	85%	0.73	0.27	11.6
ChatGPT	52%	0.84	0.27	4.5
Gemini	52%	0.90	0.10	1.6
Perplexity	56%	0.47	0.56	12.1

Read the divergence and national-overlap columns together and the story is clear. Copilot’s state answers are essentially unrelated to its national answer (1% overlap) and unrelated to each other across states (1.00 divergence). ChatGPT and Google AI Mode change substantially by state (27% overlap with national). Perplexity barely moves: its state answers still share 56% of their sources with the no-location baseline — more than double any other engine.

The localization rate — how often an engine cited a source that appears in only one state — tells the same story from the other direction:

Finding 2: ChatGPT returns a per-state map pack

For the auto prompt — the exact use case a customer flagged — ChatGPT returned a local map pack of named Chevrolet dealerships, and the list was completely different in every state:

• Texas → Galleria Chevrolet, Jupiter Chevrolet, Clay Cooley Chevrolet Dallas, Huffines Chevrolet Plano, Sam Pack’s Five Star Chevrolet
• New York → Lester Glenn Chevrolet, Circle Chevrolet, Seacoast Chevrolet, Ciocca Chevrolet
• Florida → Ferman Chevrolet, Dimmitt Chevrolet, Maher Chevrolet
• Illinois → Advantage Chevrolet of Hodgkins, Marino Chevrolet, Currie Motors, Jennings Chevrolet
• Washington → Jet Chevrolet, Auburn Chevrolet, Chevrolet of Bellevue, Titus-Will Chevrolet

Nationally, the same prompt returns brand and lender pages — chevrolet.com, gmfinancial.com, lendingtree.com. The shift shows up in the source authority too (more on that below): ChatGPT’s mean cited Domain Rating drops from 85.5 nationally to 76.5 at the state level, and the share of citations going to the local long tail rises from 0% to 36%. The model swaps national brand pages for local dealer sites the moment it knows which state you’re in.

ChatGPT was not perfectly consistent — in a minority of calls it declined to localize and asked for a precise location instead (for the auto prompt in California, it responded “I need your location first”). Localization is the strong tendency, not a guarantee on every call.

Finding 3: Copilot goes hyper-local

Copilot was the most aggressive localizer by a wide margin. It didn’t just change sources by state — it routinely named a specific metro or neighborhood inside the state, inferred from the proxy:

• Dentist, Texas → “a highly rated dentist in Beaumont, TX, Dr. Angel Rivera at Gulfside Dental”
• Dentist, New York → “Dental Arts of Sayville (15 Foster Ave, rating 4.9)”
• Personal injury lawyer, New York → “the highest-rated personal injury lawyers near you in Brooklyn include Brooklyn Injury Attorneys, P.C.”
• HVAC, Texas → “Top-rated HVAC repair companies near you in San Antonio include Champion AC, Rosenberg Plumbing & Air”
• HVAC, New York → “a highly rated HVAC repair company in Manhattan, American HVAC Corp”

That behavior is why Copilot scores 1.00 on cross-state divergence and 0.01 overlap with its national answer: it is answering a fundamentally different, hyper-local question in each state. For anyone tracking whether a local business appears in Copilot, a national-only query is close to useless.

Finding 4: Perplexity mostly ignores the state proxy

Perplexity is built around citing authoritative aggregators, and that instinct dominated location. For insurance, legal, and real-estate prompts, Perplexity returned an identical set of sources across all six states (cross-state divergence of 0.00), leaning on national directories like NerdWallet, Justia, Avvo, and Insurance.com. It also asked for a city or ZIP more often than any other engine rather than acting on the proxy location.

When Perplexity did localize, it sometimes did so in its citations — a New York query surfaced Justia’s /lawyers/personal-injury/new-york page rather than the generic national one — but the effect was inconsistent. Its 56% overlap with the national baseline (highest of any engine) and 0.47 divergence (lowest) both confirm it is the least location-reactive engine in the set.

Gemini sits in between: it often localized in prose (a New York auto query volunteered “if you are located in the New York City metro area…”) but cited very few sources — 1.6 per response on average — so there is far less citation surface to track than with Perplexity (12.1) or AI Mode (11.6).

The citation-surface gap matters for tracking: an engine that cites a dozen sources gives you far more brand-visibility signal per query than one that cites one or two.

Finding 5: the source-authority shift

To characterize what kind of sources state targeting surfaces, we pulled Ahrefs Domain Rating (DR) for the 145 most-cited domains and grouped them into authority tiers — Tier A (DR 80+), B (50+), C (20+), D (below 20).

The aggregators that anchor these answers are overwhelmingly Tier A: NerdWallet (90), Yelp (94), Forbes (94), BBB (93), Angi (90), Justia (90), Realtor.com (91), Healthgrades (86). Median DR of cited, DR-covered domains is 84.

This study is a one-off cut of local queries. For the same source-side question asked weekly of software categories, What AI cites publishes every domain the engines linked, with the long tail included rather than truncated to a head list.

At the aggregate level, state targeting doesn’t crater the authority mix — Tier A share moves only from 57% (national) to 55% (state), and mean DR is essentially flat (66.4 → 65.8). The reason is subtle and worth stating plainly: even a national, no-location query already pulls in some local businesses. What state targeting changes is not the authority tier so much as which specific entities fill it — and there, the rotation is near-total (see the divergence and overlap numbers above).

The authority shift is clearest per engine:

Engine	Mean cited DR (national)	Mean cited DR (state)	Local long-tail share (state)
ChatGPT	85.5	76.5	36%
Copilot	49.5	54.9	94%
Google AI Mode	69.7	68.2	43%
Perplexity	63.9	63.3	35%
Gemini	49.2	49.7	28%

Copilot is the standout: 94% of its state-level citations point to domains outside the 145 most-cited — the low-DR long tail of individual dealer, clinic, and firm websites. It is the engine most likely to send a citation to a small local business, and the one where appearing in the index is least about traditional domain authority.

Finding 6: some questions change far more than others

Location sensitivity isn’t uniform across topics. Averaging cross-state divergence over all five engines, the most location-dependent verticals are the ones where the “right” answer is inherently a nearby storefront — dentists, auto financing, credit unions, HVAC repair — while categories dominated by national brands and comparison directories move the least.

Homeowners insurance (0.64) and real estate agents (0.55) diverge least — insurance answers lean on national carriers and comparison sites (Allstate, NerdWallet, Insurify), and real-estate answers on national portals (Zillow, Realtor.com, Compass) — so the brand set is stable even when the local agent changes. The lesson for tracking: weight your location sampling by how local the buying decision actually is.

What this means for AI visibility tracking

The clearest evidence of the national blind spot is how little a state answer resembles the no-location one — for Copilot, a 1% source overlap means a national-only query captures almost none of what a real in-state user sees:

Three practical takeaways for anyone measuring brand presence in AI answers:

1. A national vantage point has a local blind spot. If your AI-visibility tracking queries from one location, you are blind to the state-by-state reality that Copilot and ChatGPT already serve. For multi-location brands, local-service businesses, and the agencies that serve them, location-aware AI monitoring is now table stakes — the same way local rank tracking by city became standard for Google.
2. Engine choice matters as much as prompt choice. The gap between Copilot (1.00 divergence) and Perplexity (0.47) means a brand can look invisible in one engine’s national answer and dominate another’s local one. Track each engine on its own terms, and weight them by how location-reactive they actually are.
3. Local citations are winnable without Tier-A authority. Copilot’s 94% long-tail citation rate shows that AI local recommendations are not gated on domain authority the way national head-term rankings are. A DR-15 dealership site can be the cited source — if the model is querying from its metro.

Limitations

This is an honest snapshot, not a longitudinal panel. Each prompt/engine/state cell was a single call, and AI answers are nondeterministic, so exact businesses will vary run to run; we report structural patterns (divergence, overlap, localization), which are far more stable than any individual name. The study covers six states and eight verticals in the United States on one date — directionally representative, not exhaustive.

Proxy-based geolocation can also occasionally misfire: in one California call, Copilot’s answer resolved to out-of-region banks, a sign the request egressed from the wrong location. It’s a good reminder to validate geo-targeted results and to treat any single answer as a sample, not ground truth. The state parameter targets a metro within the state, so which city a model picks (Copilot chose Beaumont for Texas, not Houston) reflects the proxy’s specific location, not a “whole state” average.

Run it yourself

Every number here came from cloro’s public API. Add "state": "<US-STATE>" to any ChatGPT, Copilot, Gemini, or Perplexity monitor call, keep country set to US, and compare the cited sources against your national baseline. If you’re building AI-visibility tracking for local or multi-location brands, that one parameter is the difference between the answer you see and the answer your customers get.

The full methodology — every prompt, state, engine, and metric definition — is described above, and the snippet reproduces any single cell. Start with an API key and run your own verticals.