Generative Search SEO Depends on Location Signals (Not Just Keywords)

The Short Answer #

Location signals change what generative engines cite and which businesses they recommend—so generative search SEO isn’t reliably measured with keywords alone. The study shows ChatGPT follows declared location down to the city, while Gemini largely follows the network’s country.

Practically, your GEO monitoring must test both declared location and network location, or your dashboards will reflect only one market’s citation behavior. For multi-market businesses, this prevents “healthy” reports that miss real visibility gaps.

The caveat is that results are engine-dependent: the location channel each model uses differs, so you must control the same variables when comparing performance across markets or platforms.

Generative Search SEO Depends on Location Signals (Not Just Keywords)

Introduction: Your GEO Audit Might Be Testing the Wrong Place #

If you’ve been experimenting with generative search optimization (GEO), you’ve probably assumed that “the results” are mostly about the prompt. But new research from Mahe Chen and Nick Koudas makes a pretty uncomfortable point: where the user “is” can change not just the answer, but the sources and the businesses recommended. This paper is a controlled audit of how location changes what generative engines cite, based on 300 questions across ChatGPT and Gemini, run 31,200 times. The authors systematically shift location signals and measure what changes. The headline isn’t “location matters” (we already knew that in classic search). It’s that generative engines read location through different channels—and audits need to control those channels or you’ll measure the wrong thing.

Why This Matters #

This is significant right now because GEO monitoring tools and dashboards are often built around a simple assumption: “we’ll send the same queries from one location.” That used to be enough when personalization was mostly about the user’s region and the search engine’s geo tweaks. But generative search adds a twist: the model doesn’t just rank sources—it rewrites the query, decides what to search, and then the engine’s localization logic chooses which citations and local entities show up.

Here’s a scenario you could run today: you manage a network of local service businesses (plumbers, clinics, home repair, credit unions). Your monitoring system sends “best plumber near me” from one fixed city. If it only sets a “declared location” for ChatGPT (or only relies on network location for Gemini), you might get a “healthy” visibility report for one market while actually missing what happens in the other markets—because the cited sources (and recommended businesses) may change drastically with the location signal.

This research also builds on earlier AI search studies that looked at citation behavior, source selection, and how answers relate to traditional search. What’s new is the measured mechanism: ChatGPT follows the declared location down to the city, while Gemini largely follows the country of the network, even when the question doesn’t mention a place. That means GEO isn’t just prompt engineering anymore—it’s also geo-signal engineering.

How the Study Tested Location Effects Without Guesswork #

The researchers designed an audit that’s closer to a lab experiment than a typical “SEO monitoring” approach.

The question set: 300 prompts across 10 commercial topics

They wrote 300 questions in 10 commercial topics: home services, health care, finance, insurance, telecom, dining, retail, cars, travel, and public services. For each topic, they used 100 questions, split across:

• Near-me questions (explicitly asking for something nearby): e.g., “best plumber near me”

• Place-dependent questions (answers depend on where you are): e.g., prices, laws, providers, weather, “how much is sales tax on a car”

• Not-local questions (should not depend on location): e.g., “best robot vacuum” Importantly: none of the questions named a specific place.

Two kinds of location: “declared” vs “network”

The audit carefully separates location signals, because generative engines can take location from different places:

• A declared location: set by the application/tooling you use to call the model (for example, “approximate user location”)

• A network location: inferred from where the API request comes from (your connection, VPN exit, etc.)

They then used two engines with different behaviors and routing rules.

How they ran ChatGPT vs Gemini

For ChatGPT, they used gpt-5.4-mini (via the Responses API with a web search tool) and ran:

• E1 (main run): all questions twice, at 19 declared locations, with all answers coming from one network location

  - Total: **24,000 answers**

- **E1 robustness run:** repeated a sample with the full-size model `gpt-5.4`

  - Total: **720 answers** (60 questions × 2 × 6 locations)

For **Gemini**, they ran:

• E2 (main run): every question twice per engine call, in six network sessions from different regions

  • Network locations included the authors’ connection in Canada, VPN exits in New York, Los Angeles, London, Sydney, and the own connection again three days later (to measure drift)

  • Total: 7,200 answers

• Note: Gemini’s API also accepts declared coordinates, but in a pilot those had no effect, so the explanation is that Gemini ignores that declared location signal and uses network country instead.

Measuring “what changed”: citations and business recommendations

For each answer, they recorded:

• Whether the model searched at all

• The search queries it used

• The cited URLs/domains

• Whether any query named a location

To quantify how location changed sources, they compared domains using a Jaccard index and computed a “localization index.” They also extracted the businesses/brands/products each answer recommended (excluding directory/review list links unless a business was explicitly recommended). In a manual check, 29 of 36 extractions were fully correct.

If you’re building GEO monitoring, that’s an important detail: they weren’t just counting citations—they were measuring the commercial recommendations you actually care about.

What the Study Found: Location Changes Citations (Sometimes a Lot) #

Let’s get to the core findings: location affects which sources a model cites, and the effect depends on query type and which engine you’re using.

ChatGPT: city-level declared location changes sources strongly for local queries

With declared locations:

• For near-me questions, ChatGPT’s cited sources differed between two US cities by 0.14, and between cities in different countries by 0.28.

• For place-dependent questions, the differences were 0.17 (within US) and 0.43 (between countries).

• For not-local questions, the effect was essentially nil: 0.00–0.03.

They also measured a second behavior: sometimes ChatGPT didn’t cite sources at all depending on whether a declared location was provided. Without a declared location, it often asked for ZIP code/city. With a declared location, it cited sources in 64–73% of answers, compared to 46% without.

Gemini: network country drives citations—even for not-local questions

Gemini behaved very differently:

• For near-me questions, source differences between network locations in different countries averaged 0.43.

• For place-dependent questions, differences were also around 0.40.

• Surprisingly (and crucially for GEO auditors), for not-local questions, citations still differed across network countries by 0.23.

So even when the question shouldn’t depend on place, Gemini’s network-based localization changes what it cites.

Localization Happens Through Different Channels (And That Changes Everything) #

This is the “gotcha” part for GEO: it’s not just that location matters. Each model uses location differently, and even within a model, localization can happen in two ways.

ChatGPT uses location twice: in query rewriting and in backend localization

For ChatGPT, location entered via two routes:

1. The model writes the place into its searches - For near-me questions , ChatGPT put the place into100% of its searches

• For place-dependent questions , it did so in48% of cases

2. For
3. The search backend localizes even if the query itself doesn’t name the place - If neither of the two answers’ search queries named the place, citations still differed with localization index 0.17

• If both queries named the place, the index rose to 0.37

4. If neither of the two answers’ search queries named the place, citations still differed with localization index

They also found ChatGPT’s localization doesn’t “fade with distance.” For near-me questions:

• Localization index was 0.13 between US cities < 1,000 km apart

• It was 0.14 between cities > 3,000 km apart

And the distance-vs-effect correlation was basically 0.01.

In plain language: for ChatGPT, every city can be its own market. A “nearby” city doesn’t reliably stand in for your target city.

Gemini ignores declared location and follows network country

For Gemini, the story is simpler:

• It ignores declared coordinates (declared location had no effect in their pilot)

• It follows the country of the network

• It didn’t localize down to city in their setup

They tested drift and found it didn’t explain country-level differences:

• Repeating the same connection 3 days apart still produced low variation (values like 0.03, 0.01, 0.05 across pairs)

• And London vs Sydney differed as much as other country pairs

Also, Gemini’s queries named the country (in 56% of near-me answers) and never a city.

What Changes for Businesses: The Recommendations Are Not Stable Across Location #

If you only care about citations, you might stop here. But GEO is about visibility—and visibility includes which brands and businesses get recommended.

ChatGPT recommendations are city-specific when location is declared

With declared location, ChatGPT changed recommended businesses:

• A near-me answer from ChatGPT named 4.3 businesses on average.

• Two identical requests in the same city shared 23% of business names.

• Two different declared locations shared only 4% of business names.

• They report that 96% of businesses recommended for near-me questions appeared in only one of the 15 declared cities.

That means if you monitor from a single city, you’re not “covering” nearby markets—you’re missing most of the distinct business mentions.

Gemini recommendations change across countries—even for not-local questions

For Gemini:

• Recommended business names differed between network countries even when questions were not local

  • index 0.07 (CI 0.04–0.10)

• And for a non-place question like “cheapest groceries,” Gemini recommended different national chains depending on network country.

The paper includes example prompts showing:

• For “credit unions,” ChatGPT names local institutions matched to each declared city.

• For “cheapest groceries” with no place in the question, Gemini recommends different chains across network countries—and repeats those same chains for New York and Los Angeles (since both are US networks).

GEO Monitoring Implications: Your Tool Can Measure the Wrong Engine-Location Pair #

Now for the part you can actually use.

Compare engine behavior: which location signal matters?

Here’s the “don’t mess this up” comparison, based on what they observed:

Engine	Location signal that matters most	Granularity in their tests	What happens for not-local questions
gpt-5.4-mini /gpt-5.4 (ChatGPT)	Declared location from the app/tool call	City-level (localization doesn’t fade with distance)	Mostly stable: localization index around 0.00–0.03

| gemini-3.1-flash-lite (Gemini) | Network country (declared coordinates had no effect) | Country-level in their setup | Still changes: localization index around 0.23 |

How to set up audits correctly (the paper’s practical warning)

The authors’ discussion leads to three concrete rules:

1. Set location in the channel the engine actually uses

• ChatGPT: set declared location
• Gemini: control network location

2. ChatGPT: set
3. Use the right granularity for each engine

• ChatGPT: a single “representative” city doesn’t work—even nearby cities can behave differently
• Gemini: one network per country may be enough for city-level differences (in their setup), but you should measure each country market

4. Hold location constant when comparing over time or across engines

If you change location signals between runs, you’ll see source and business changes that arenot due to model updates or optimization efforts.

What This Means for “Optimization”: Where Visibility Is Actually Won #

Finally, the study hints at how GEO should think about winning visibility.

ChatGPT visibility comes via place-specific sourcing

For near-me and place-dependent questions, ChatGPT:

• writes the place into searches

• cites sources that name city/state/country

That suggests local visibility for ChatGPT isn’t won through national “top lists.” It’s won through pages that appear relevant for place-specific queries—often:

• local directories

• local review sites

• local institutional pages (government offices, local organizations, etc.)

Gemini visibility is more like a country-level filter

For Gemini, the country of the network decides national chains and sources—even for questions that don’t depend on place. So your “brand visibility” claim may only hold within the same network country context.

If you’re using Gemini for monitoring, you shouldn’t assume that a measurement from one network location generalizes to another country.

Key Takeaways #

Key Takeaways #

• Location changes citations and business recommendations in generative search; it’s not just “better answers.”
• ChatGPT follows declared location down to the city and uses it both by rewriting search queries and by backend localization.
• Gemini follows the network country , ignoring declared coordinates in the study’s setup—even for questions that don’t mention a place.
• For ChatGPT, local effects don’t fade with distance: nearby cities behave like different places .
• GEO monitoring that runs from one place can seriously mis-measure visibility:
  • For near-me queries, the recommended businesses often show up in only one city (96% were city-unique across 15 cities).
• For near-me queries, the recommended businesses often show up in only
• Audit design matters as much as prompt design : you must control both the declared location channel (ChatGPT) and the network channel (Gemini) to get results you can trust.

If you want, tell me what GEO tool or workflow you’re using (and whether it calls ChatGPT, Gemini, or both). I can suggest a concrete audit setup that controls the right location signals for your use case.

Sources Used #

This article is a plain-English breakdown of the following peer-reviewed preprint. Read the original for full methodology and results:

- [Location, Location, Generation: Where You Ask Matters for Generative Engine Optimization](https://arxiv.org/abs/2610.24081) — arXiv
- Authors: Authors: Mahe Chen, Nick Koudas

source & further reading

thepromptindex.com — original article AI Changes the Value of Schooling vs Experience (Returns) Agentic AI may be reshaping arXiv submissions—here’s the signal ChatGPT Scores and Novelty Words: Do Abstracts Game Evaluation?