DeepSmith

Nov 26 · AEO & AI Visibility

16 min read

Personalization and Location in AI Answers: Why Two Users See Different Results

Avinash Saurabh
Avinash Saurabh · CO-Founder & CEO
Two identical search bars each lead through location and chat icons to different stacked answer cards, under the line Same Question, Different Answers.

Two people can ask an AI engine the exact same question and get different answers, because the engine may be using different account memories, search history, location signals and language settings when it reads the prompt and builds a reply. If you track your brand in AI search, this matters a lot, since the answer you collect is only one version of what your buyers might see. This article walks through what changes the answer, how that works on ChatGPT, Google, Gemini and Perplexity, and how far one tracked result can safely be trusted. If you have noticed ChatGPT personalized results that don't match what a colleague sees, this is the reason.

What personalization in AI answers actually means

People tend to use "personalization" for everything that makes two answers differ, but it covers several different things, and it helps to pull them apart before looking at any one engine.

The first is account or conversation history. This is information from earlier chats or searches that an account setting allows the engine to reuse, like a preference you mentioned once or a topic you keep coming back to. The second is explicit profile information, where you tell the engine something directly, such as your location or the language you want replies in. The third is inferred location, where the service guesses a general area from your network connection, or uses your device location if you gave permission. Network location is usually approximate, while device location can be much more precise, so they are not the same signal.

Language is its own thing too. The language you type the prompt in, the response language saved in your profile and the display language of the app can all be different settings. And there is plain contextual relevance, which is not personalization at all. Your location, device and current question can shape a result even when every personalized recommendation setting is turned off.

One more distinction is worth keeping in mind. A signal can change how an engine searches and which sources it finds, or it can change only how the answer is worded and framed. The documentation for most engines does not say exactly which of those is happening, so it is safer to say an answer changed than to say you know why it changed.

How ChatGPT personalizes what it finds

OpenAI says ChatGPT Search may rewrite your prompt into one or more targeted search queries, and that it may use relevant Memory information when it does that, if Memory is turned on. Their example is a restaurant request that gets refined using a remembered dietary preference and a remembered location. That is a real route to different sources, because two people who type the same words can end up sending different searches out to the web.

Location works in a similar way. ChatGPT may estimate your general location, such as country, state or city, from your IP address and use it to improve search results. OpenAI says that general location may be passed to third-party search providers, but the IP address itself is not shared with them for this purpose. Precise device location is a separate and optional signal. If you allow it, it can help with very local recommendations, and Search still works without it, just with results that are less specific to your area.

Memory itself is not a full recording of everything you have ever said. It can include relevant past chats, saved memories, custom instructions, saved files or connected-app information, depending on your plan, region and workspace settings, and ChatGPT looks for relevant context when it thinks it will help. What gets used can change from one question to the next. That is part of why ChatGPT personalized results can feel inconsistent even for one person.

Temporary Chat is also less simple than it sounds. Before you start one, you can pick a Personalized or an Unpersonalized mode. Unpersonalized leaves out Memory, custom instructions and plugins, and a temporary chat does not create or update memories while it stays temporary. The choice has to be made before the conversation starts, and account or workspace restrictions take priority. OpenAI also notes that limited, safety-related use of earlier context may still happen. This matters if you run checks by hand, since a quick temporary chat may not be as clean a test as you expect. We cover that more in our guide to checking ChatGPT visibility manually.

How Google Search and AI Mode use your context

Google's general Search documentation separates personalized recommendations from the other reasons results can differ. When personalized recommendations are on, Search can use your account activity, including Search services history, to help find what you want. That can change the order of results or the order of content blocks, for example putting videos above websites for someone who watches a lot of videos.

Google also says results can vary because of language settings, localized results, location, device type and the search itself. Even with personalized recommendations off, Google may still use location, device and your current search. So switching off personalization does not make results identical across places or devices, and it does not erase stored activity or saved preferences. It's a good idea to read that as a statement about general Search, though. It does not prove that every Google AI feature uses each signal in the same way.

For AI Mode, Google documents a feature called Personal Intelligence, which can use your previous searches and activity saved in Search Services History, past interactions, and information you have shared to tailor its suggestions. You can ask AI Mode to remember or forget details. The feature needs you to be at least 18 and to have history and personalized recommendations turned on. AI Mode still works without those settings, but you can't pick up where you left off with earlier searches. Google also warns that personalization may not always be correct.

Google also documents connecting its content apps, such as Workspace and Photos, to give its AI experiences more personal context. That is one more possible source of difference between two people, but it only applies to people who connected those apps, and it would be a mistake to assume every answer uses it. If you want the practical side of watching this surface, see how to track whether your pages appear in Google AI Overviews.

Gemini and Perplexity add their own memory settings

Gemini Apps can use memory of your past Gemini chats to personalize replies. The documented feature needs a user aged 18 or older who is signed in to a personal Google Account, and it is not available for work, school or supervised accounts. It is also not available in some experiences, including Gems and Live chats. You can turn past-chat memory on or off and delete past chats, though Google notes there may be a short delay before a deleted chat stops being used.

Gemini can also use connected Google apps through Personal Intelligence. The connected-app settings are separate from the setting for memory of past Gemini chats, so turning off one does not necessarily turn off the other. On language, Google's documentation says Gemini can understand and reply in any supported language when you type or speak in it, no matter which display language you picked. The evidence we found does not show that your display language changes the search sources behind every answer.

Perplexity says it uses Memory and previous searches depending on the question and whether they help produce a better response. Memory keeps key information and preferences between conversations, and search history is configured separately. Its Profile settings can include location, interests, values, age, dietary needs and preferred response language. Location is inferred from the network by default, but you can set it by hand, and if your profile has no location, answers are based on the network location. By default the response language matches the language of your question, and you can override that in your profile.

Perplexity also says Memory and search history can be switched on and off independently, and that both are off in Incognito mode. Individual memories can be deleted in Settings. What its help pages don't do is promise that every profile preference affects every answer, or that Memory always changes retrieval instead of only the wording. So with Perplexity, as with the others, you can see the settings that exist, but you can't see exactly how they change a given result. If you track that engine, our guide on how to monitor your brand on Perplexity goes into the sampling side.

What a location study found about AI answers by location

Most of the engine documentation above describes controls, not effects. For the effect of place on what gets recommended, there is a 2026 arXiv preprint by Mahe Chen and Nick Koudas called "Location, Location, Generation: Where You Ask Matters for Generative Engine Optimization." It is one study, and it is worth being careful about how far it reaches, but it gives a concrete look at AI answers by location.

The authors wrote 300 English-language questions across ten commercial topics: home services, health care, finance, insurance, telecom, dining, retail, cars, travel and public services. The questions were split into "near me," place-dependent and not-local groups, and none of them named a location in the question text. They ran two experiments. In the first, they used the ChatGPT Responses API with web search, mainly gpt-5.4-mini, and either handed the search tool a declared location or none. The declared locations were 15 U.S. cities plus London, Toronto, Sydney and Dublin, and the paper reports 24,000 ChatGPT answers across its conditions. In the second, they sent requests to ChatGPT and Gemini with no declared location from different network sessions, including VPN exits in New York, Los Angeles, London and Sydney plus Canadian sessions, for 7,200 answers.

The main finding is that the two engines responded to different location channels. ChatGPT followed the location declared to the application, down to the city, and did not follow the tested network location. Gemini did the opposite, following the network country instead of the declared location. Those are findings about the APIs, models, questions and places the authors tested, not guarantees about every consumer app session.

A few of the reported numbers show why this matters for anyone who tracks one location:

  • For near-me questions, 79% of ChatGPT answers in U.S. cities and 92% in the other tested countries cited a city- or state-specific source.
  • Identical ChatGPT requests shared 23% of their recommended businesses, while requests with different declared locations shared 4%.
  • Across the 15 U.S. cities, 96% of businesses recommended for near-me questions showed up in only one of those cities.
  • Gemini's recommended names differed across network countries even for questions classed as not local, with a reported index of 0.07, while ChatGPT's recommendations did not change with network location in the tested conditions.

The authors list their own limits. They used APIs and small models that need not behave like consumer apps, and their robustness check with a full-size model covered ChatGPT only. The VPN exits were data-center addresses, the network test covered just five locations, and Gemini may localize more finely on home connections. The questions were English only and written by the researchers, and the measurements describe the engines as they were in October 2026. The paper is a preprint, and its bibliographic details do not state whether it has been peer reviewed.

So the fair way to use it is as evidence that place can change sources and recommendations in tested generative search, and nothing more. It does not show that every engine behaves this way, that account history was tested, or that any one percentage applies to your prompts.

Language changes less than you might expect, and more than nothing

The documentation supports a few separate language claims. Google says language settings can be one reason general Search results differ between people. Perplexity says its response language defaults to the language of your query and can be overridden in your profile. Gemini lets you choose a display language and says it can reply in supported languages when you prompt in them.

Those facts support saying that language can change the language of the answer, and it may affect the localized search context behind it. They don't support a rule that flipping a language toggle changes which sources get picked or which brands get cited on every engine. We did not find a controlled study that isolates language as the only variable in consumer AI search, so that stronger claim stays unknown for now. If your buyers ask in several languages, treat each language as its own set of questions to check, rather than assuming the English results carry over.

What one tracked answer really represents

A tracked answer reflects the conditions it was collected under. It is not an average of what everyone sees. To read it properly you need to know the engine and feature, the exact prompt, the location signal, the language, the account or history state, any connected apps, and whether the engine was using web search or another mode. Those are context variables, and when two results differ it is usually not possible to say which one caused it.

The location study is a good reminder that a single location is a poor stand-in for local answers. In its near-me questions, the recommended businesses changed sharply from place to place. That is evidence about limited generalization in that setup, and not a benchmark you can apply to every program. It is also a different thing from run-to-run variation, where the same person asking the same question twice can still get different answers, which we cover in a separate piece.

This is where AI visibility location tracking runs into a real limit. A visibility result is useful, but how far it generalizes depends on how closely the collection conditions match the audience you care about. A brand that sells in one country and one language has less to worry about than one that sells across several regions. And you should not assume a tool offers location-segmented tracking unless its own documentation says so. For a sense of how tools gather answers in the first place, see how AI visibility tools collect data.

What to do with this as a marketing lead

You don't need to solve personalization to get value from tracking. A few habits keep you honest about what your numbers mean.

Write down the conditions next to every result. Note the engine, the prompt, the date, and the location and language you collected it under. When a colleague says they saw something different, that note is how you work out whether it is a real change or just a different setup.

Decide which audience your tracking should stand for, because AI visibility location tracking is only as useful as the match between your setup and your buyers. If most of your buyers are in one country and write in one language, one consistent collection setup can be a fair view of them. If you sell in several places, plan to treat each place as its own slice, and avoid averaging them into one number. A good way to start is the repeatable prompt set approach, where you check the same questions on a regular schedule so changes are easier to spot.

Keep manual checks and tool checks separate in your head. Your own logged-in account carries your own history, so what you see there is not a neutral baseline. A clean session, or a tool that collects answers the same way each time, gives you something steadier to compare against. DeepSmith tracks the prompts you choose on a schedule and keeps the full answer history for each one, which makes it easier to see how an answer moved over time, though it does not tell you why. If you want to try that on your own questions, you can start a free trial.

Finally, if you run a local or multi-location business, treat place as a first-class part of your content plan, and not only your tracking. Our look at local and multi-location AEO tools is a place to start, and so is the guide on getting a local business cited in near-me answers.

Frequently asked questions

Why do two people get different AI answers for the same question?

Their accounts may supply different memories or search histories, their profiles, locations or languages may differ, and the engine may retrieve different local or language-relevant sources for each of them. The mix depends on the platform and the feature, so a difference in the answer doesn't tell you which signal caused it.

Do AI answers still vary by location if personalization is off?

They can. Google says general Search may still use location, device type and your current search when personalized recommendations are off. ChatGPT Search may use an approximate location from your IP address, and precise device location is optional. What happens depends on the platform.

Does changing the language change the sources an AI cites?

It can change the language of the answer, and general Search can be localized by language. The evidence we reviewed does not show that changing language alone changes citations in a consistent way across all AI engines.

Is one tracked answer representative of what all users see?

Not necessarily. A result reflects its engine, prompt, account context, language and location conditions. The location study found large differences in recommended businesses for its tested near-me questions, but it does not set a universal rate for all engines or users.