Does AI search answer Swedish questions from Swedish sources?

Which AI search surfaces do Swedes actually use?

Google’s AI Overviews reach the most Swedes, followed by ChatGPT. In Internetstiftelsen’s survey, run in January 2026, 71% of Swedes said they use or read Google’s AI Overviews, 62% use AI services, and 81% use AI in one of these forms. ChatGPT went from 34% to 52% of Swedes between 2025 and 2026; Copilot is used by 18%, Gemini by 13% and Claude by 2%. One in seven, 14%, say they google less often than a year ago, and half of those say they use AI instead.

These are self-reported figures from a Novus panel of 3,776 people. Of AI users, 76% say they usually double-check whether an AI answer is correct, most often by comparing it with other sources such as traditional media, government sites or reference works. Google expanded AI Overviews to more than 40 languages on May 20, 2025, and Dagens PS reported that Swedish was included. AI Mode started rolling out in Swedish in Sweden on October 8, 2025.

What it means. For a Swedish audience the AI surface most people see is Google’s, and ChatGPT is second. That ordering matters below, because the evidence on language differs most between those two.

Does ChatGPT research in English when you ask in Swedish?

Often, according to one vendor, though it published no figure for Swedish. Peec AI, which sells AI-visibility tracking, reported in February 2026 that it had analyzed “over 10 million prompts and 20 million query fan-outs” from its own data. A fan-out is one of the background searches an assistant runs to answer a single question. Peec kept only cases where the user’s location matched the query language, such as Polish queries from Poland. In about 78% of the non-English sessions, ChatGPT added English-language searches, and 43% of the research steps for non-English prompts ran on the English-speaking web. In Peec’s words, ChatGPT “starts its research in the user’s language, but then switches to English”.

Four limits apply. First, the prompts are ones Peec’s customers chose to track, run daily by browser automation, not a panel of consumer ChatGPT sessions; Search Engine Journal adds that the report did not detail what the prompts were. Second, the post is inconsistent about the denominator: one sentence says that in all non-English sessions “at least one sub-search is performed on the English web”, another says “nearly 78% of cases”. Third, the reason Peec gives, that native-language sources “aren’t enough”, is Peec’s interpretation; Search Engine Journal notes that “OpenAI’s documentation does not describe how language is chosen for rewritten queries”. Fourth, the chart gives no per-language sample sizes.

What it means. If your customers ask ChatGPT, part of the research behind a Swedish answer probably happens in English, so your English pages, and English pages about you, can be read even when the question was in Swedish. How large that share is for Swedish is unknown.

Which AI engines cite Swedish-language pages for Swedish questions?

Most of them, most of the time, and Google AI Overviews most of all. Temso AI, which sells AI-visibility tools, matched 7,058,891 citations from ChatGPT, Microsoft Copilot, Grok and Google AI Overviews to the language of each cited page, for commercially oriented prompts in seven languages including Swedish. Temso’s study was last updated April 20, 2026 and uses data from early 2026.

Swedish does worse than most languages in Temso’s data. Of the six non-English languages, Swedish prompts got the lowest share of same-language pages on Copilot and on ChatGPT, and the second lowest on Grok, after Dutch; on Google AI Overviews Swedish was third of six, behind Italian and German. That ranking is our reading of Temso’s per-language table. On Grok, Swedish prompts got 43.7% English sources against 47.1% Swedish, “nearly equal” in Temso’s words. Temso offers the size of each language’s web as “the likely explanation”; it does not test it. On ChatGPT, the pages that were not in Swedish add up to 42.6% of citations; that is our subtraction from Temso’s 57.4%, not a figure Temso reports.

Two details shape how to read this. Temso counts a page as local when its language matches the prompt’s language, so a Swedish page on a .com domain counts as local. And its prompts are commercial and can name the country; one example in the study is “Bästa försäkringsbolag i Sverige”. Peec’s documentation points the same way for Google AI Overviews, without a number: “ChatGPT mixes English and local-language sources even for non-English prompts, while AI Overviews stays closer to the query language”.

In Temso’s data, industry matters about as much as engine. Across non-English prompts, Temso found local-language citation at 76.9% for K-12 education and 35.5% for hotels, a gap of 41.4 percentage points, not broken out for Sweden.

What it means. The simple version of the worry, that a Swede’s question is answered from English pages, is not what Temso found on most engines. A Swedish page is usually what Google AI Overviews cites for a Swedish commercial question. ChatGPT and Grok cite a large share of non-Swedish pages for the same questions, so a Swedish-only site leaves part of their research to other people’s pages.

Why do studies disagree about which engine is most local?

Because they measure different things, on different prompts, in different industries. When the definition changes, the engine ranking changes with it.

In that preprint, “Google” means Google’s classic top-10 web results, not AI Overviews or AI Mode, so it says nothing about how local those two are.

MaxAEO’s averages across non-English markets put Google AI Overviews third of seven: Perplexity 44%, Copilot 41%, AI Overviews 38%, AI Mode 36%, Gemini 33%, Claude 29%, ChatGPT 27%. Its own caveat is that 200 prompts per market “cannot resolve differences below roughly 1.5% share”, and that consumer categories “will show a different source mix”. MaxAEO also states that in the Netherlands and Sweden “engines fall back to English sources 80%+ of the time”. That is MaxAEO’s statement: its table gives no Swedish English-share figure. MaxAEO leaves out companies’ own sites and counts only third-party citations, while Temso matched every cited page, own sites included, to its language. A low MaxAEO local share and a high Temso Swedish share can both be true. In Sweden, the top local sources for B2B software in MaxAEO’s data were tech and business press: Computer Sweden / IDG 5.9%, breakit.se 4.4%, di.se 3.1% and nyteknik.se 2.4% of non-vendor citations.

Temso’s industry pages count country domain endings, and Temso warns that “The country ending is a proxy for local sourcing, not a measurement of language” and that local pages on .com domains fall outside the count. In its telehealth data, Sweden had the highest local-domain rate of the non-English markets, at 95%.

What it means. There is no stable answer to “which AI engine is most local”, so don’t plan around one. Google AI Overviews comes out highest for Swedish prompts in Temso’s language study, the only study we found that reports the language of cited pages for Swedish prompts on each engine, and lower in studies that define “local” differently. The ranking in your own category, in your own market, is the only one that applies to you, and it has to be checked engine by engine.

Do AI models assume the user is in the US?

When they answer from memory, yes, on average. The best evidence is “Location Not Found”, a peer-reviewed paper accepted at ACL 2026, most of whose authors list Google Research. It tested 32 models on 2,156 questions in 12 languages about facts that depend on where you are, such as laws, dates and measurements. Models favoured US answers “even when models are asked in languages other than English”. US answers made up 26% of the answers in the data but 50% of the models’ outputs, an average US-bias score of 0.24. The score ranged from about 0 for Falcon 3 to 0.42 for Grok 4. In 6.5% of responses, the authors write, “local reality is completely overwritten by US norms”. The models included versions of GPT, Claude, Gemini and Grok.

Four details matter before applying this to what a Swede sees in ChatGPT or AI Overviews:

• No Swedish. The 12 languages are English, Spanish, French, German, Hebrew, Hindi, Indonesian, Italian, Japanese, Korean, Portuguese and Chinese.
• No web search. Models got “only the question as input”. The paper does not show whether search reduces the US default.
• Law questions showed little US bias. The bias score was 0.052 for law and 0.304 for questions about state and country, such as government and infrastructure.
• Saying where you are does not remove the bias. When the locale was written into the prompt (“Locale: Mexico. What is the currency?”), models still sometimes gave the US answer, and the more accurate a model was overall, the larger the share of its remaining errors that were US answers. The authors call US-centricity “a sticky and proportionally larger residual failure mode”. This tests the user’s prompt, not text on a company’s website.

A narrower preprint points the other way for legal and administrative questions. Across seven models, in English and Mandarin only, 74.5% of English-input responses to prompts that required a single answer adopted a US framework, and 53.3% of Chinese-input responses adopted a China framework.

What it means. When an assistant answers from memory, without searching, it leans toward US answers to location-dependent questions, so a Swedish company’s context can lose to the US default. How much depends on the topic: the bias was small for law questions and largest for questions about state and country. The research tests the user’s prompt, not your page, so it cannot show that stating your location on your site fixes this.

Do AI models prefer sources in the question’s language?

Yes, moderately, in lab settings, and they fall back to big languages when nothing matches. “Faux Polyglot” (NAACL 2025, peer-reviewed) found that models “displayed systemic bias towards information in the same language as the query language in both document retrieval and answer generation”. When no information is in the query’s language, models “prefer documents in high-resource languages during generation”.

The effect is real but not absolute. Answers came from native-language documents 42% of the time and from foreign-language documents 29.58% of the time; only 8.55% combined both. The study used 13 models, 170 documents, five languages (English, Hindi, German, Arabic and Chinese) and synthetic documents about fictional entities, with one retrieval design. It did not test Swedish or a live search engine.

What happens when two sources disagree matters as much. Robert Östling and Murathan Kurfalı of Stockholm University gave 12 models conflicting facts in five languages. All of them, GPT-5.2 included, “ignore the conflict and confidently assert only one of the possible answers in the large majority of cases”. Models tended to prefer the prompt’s language, but “factors other than language seem to determine which needle is chosen in at least 90% of cases”. This is a preprint, most models were in the 3B to 9B range, and Swedish was not one of the tested languages.

What it means. In lab tests, a page in the question’s language had an edge, and pages in large languages such as English were the fallback when nothing matched; Swedish was not tested. If your Swedish and English pages give different prices, dates or terms, an assistant may pick one and not mention the other, as the tested models did in most cases.

How often do Swedish AI answers cite machine-translated Reddit threads?

On Google’s AI products in Sweden, most Reddit citations point to Reddit’s machine-translated pages. Peec analyzed 64.77 million Reddit citations from March 1 to June 10, 2026, across 20 countries and four engines: ChatGPT, Google AI Overviews, Google AI Mode and Gemini. It counted a citation as translated when the URL carried Reddit’s ?tl= parameter, which it calls a proxy. Overall, 5.71% of Reddit citations led to a translated page.

Peec’s headline range, “between 40% and 73%”, spans these eight markets and both Google products; “English threads” is Peec’s description of the originals. In the US, UK, Australia and India translated pages stay under 1% of Reddit citations.

ChatGPT changed in May 2026. Its share of Reddit citations pointing to translated pages fell from 6.14% in April to 0.66% in May, and in Sweden from 20.51% to 1.62%. The 10.1% in the table pools all months, including before the change. Gemini cited translated Reddit in just under 1% of its Reddit citations in most months, and Perplexity, checked separately, never cited one.

Two facts limit how large this is. These are shares of Reddit citations only, and Peec does not give the number of Swedish Reddit citations. Profound, another vendor, which evaluated 3.25 billion citations across 7 models and 14 countries in March 2026, found that Google AI Overviews can nearly erase social content in Swedish. If Swedish AI Overviews rarely cite social content, a high translated share of a small number is a small effect.

Reddit says it wants the translated pages found in search. Its Q3 2024 shareholder letter says “Machine translated content is indexed and discoverable in search engines, driving incremental user traffic to Reddit”. By Q3 2025 it was “available in 30 languages”; the letters do not name Swedish. Google’s own “translated results” feature does not list Swedish and Google “doesn’t host any translated pages”, so these appear to be Reddit-hosted pages indexed like any other; Google has not said so.

What it means. On Google’s AI products in Sweden, an English Reddit thread about your category can be cited in Reddit’s machine-translated version, so what English-language Reddit says about you can reach Swedish answers. It is a narrow channel, and on ChatGPT it had almost disappeared by May 2026 in Peec’s data, which ends on June 10, 2026.

What do Google and the AI companies say about language and location?

Google documents language and location signals for classic Search and says AI Overviews and AI Mode have no extra requirements. Its AI features page says “There are no additional requirements to appear in AI Overviews or AI Mode”, and that both may use “query fan-out”. That page says nothing about language.

Google’s international guidance, written for Search, says:

• “Google Search tries to find pages that match the language of the searcher.”
• Google determines a page’s language from its visible content: “We don’t use any code-level language information such as lang attributes, or the URL.”
• Use one language per page and avoid “side-by-side translations”, with “different URLs for each language version of a page”.
• Other signals of the intended audience “can include local addresses and phone numbers on the pages, the use of local language and currency, links from other local sites, or signals from your Business Profile”.
• “Most, but not all, Google crawls originate from the US”, so content should not change by visitor IP.

Google’s spam policy lists scraping content to generate many pages, “including through automated transformations like synonymizing, translating, or other obfuscation techniques”, under scaled content abuse, where “little value is provided to users”. It is not a ban on translating your own site; it is a reason not to publish bulk, unchecked machine translation.

The location and language settings we found in OpenAI’s, Anthropic’s and Perplexity’s documentation are for their developer APIs, not the consumer apps people use.

What it means. Google tells you what it uses to match language and audience in Search, and those signals are cheap to get right. Whether AI Overviews and AI Mode use the same signals, Google has not said beyond “no additional requirements”. For the chat assistants, the documentation we found covers their developer APIs, not their consumer apps.

What should a Swedish company publish, and in which language?

Publish the facts that decide a purchase in Swedish and in English, on separate pages that agree, with your location and currency stated in the text, and earn mentions in Swedish third-party sources. The steps below are in order, each with where its evidence stops. All of them are inference from the evidence: we found no study that tests any of them on a company’s own pages. Steps 3, 5 and 7 rest on the thinnest evidence.

1. Make sure the AI services can read both language versions. Language does not matter if the page is refused or the facts exist only in JavaScript. Run the free Can AI read my site? check on both the Swedish and the English URL.
2. Keep the key facts in Swedish, on Swedish URLs. Price or price range, what is included, who it is for, delivery or lead time, terms. Swedish pages were the majority of citations for Swedish prompts on AI Overviews (85.1%), Copilot (60.4%) and ChatGPT (57.4%) in Temso’s data, models prefer same-language sources in the lab, and Google matches pages to the searcher’s language. Use one language per page and its own URL, not side-by-side translation.
3. Publish the same key facts in English, on their own URLs, and link the versions with hreflang. This is for the share of research that happens in English: Peec’s fan-out data for ChatGPT, the roughly four in ten ChatGPT citations for Swedish prompts that were not Swedish pages (our subtraction from Temso’s 57.4%), and the lab finding that models fall back to high-resource languages when nothing matches. Inference: Peec published no Swedish fan-out figure, and its prompts are tracked prompts, not a consumer panel. hreflang helps Google pair the versions; Google does not use it to detect language.
4. Make the versions say exactly the same thing. Same prices, dates, names and terms in both languages, updated together. When sources conflict, models usually assert one answer and do not mention the conflict. Write the English version properly; Google’s spam policy covers scraped content mass-produced through automated translation with little value to users.
5. State where you operate and how you price, in the text. Country and city, the markets you serve, “prices in SEK” or kronor, and which country’s law applies to your terms. Google lists local addresses, phone numbers, local language and currency as audience signals for Search, and models answering from memory leaned toward US answers on average when the locale was left unstated. This is inference: the locale research tests the user’s prompt, not page text, and the US bias for law questions was small.
6. Earn mentions in Swedish third-party sources, and in English ones. For B2B software, MaxAEO’s top Swedish local sources were Computer Sweden / IDG, breakit.se, di.se and nyteknik.se. But MaxAEO’s Swedish local share was only 17% of non-vendor citations, so English-language coverage of your category matters too. The section Whose pages does the AI actually read? in our review of AI recommendations covers why third-party pages carry weight.
7. Know what English-language Reddit says about your category. On Google’s AI products in Sweden, most Reddit citations pointed to machine-translated English threads. An English thread about your category may be what a Swede is shown, in Reddit’s machine translation. Inference, and a narrow channel: it is a share of Reddit citations only, Swedish AI Overviews may rarely cite social content, and ChatGPT’s translated share fell to 1.62% in Sweden in May 2026. Do not plant threads; see what decides AI recommendations for why.
8. Check per engine, in both languages, with and without naming Sweden. Ask your real customer questions in Swedish without a place name, in Swedish with “i Sverige”, and in English, on AI Overviews, ChatGPT and Copilot at least, and note which pages and which languages are cited. Engine rankings flip between studies, industry changes the picture, and a country-named prompt is not the same test as an unnamed one. Run each prompt several times, because single answers vary; as one practitioner who tested hreflang put it, “testing for different languages across AI Search platforms is not a perfect science”.

What it means. None of these steps is expensive, and none is guaranteed. Steps 2 to 5 are the ones a company controls fully: the same facts, in both languages, on pages that agree, with your country and currency in the text. They are also useful to human readers whatever the engines do.

What does the evidence not tell us yet?

It tells us little about Swedish questions specifically: the Swedish figures on this page all come from vendors’ citation data, and none of the lab studies tested Swedish. As of October 2026:

• No Swedish in the lab. Location Not Found, Faux Polyglot and the Stockholm University conflict study did not test Swedish.
• No test of page text. We found no study that checks whether stating location, currency or jurisdiction on a company’s page changes AI answers. The locale study changed the prompt and used no search.
• Mostly vendor data. Temso, MaxAEO, Peec and Profound all sell AI-visibility tools. Their prompt sets differ and none is a panel of real Swedish users.
• No consumer-app documentation found. The location settings OpenAI and Anthropic document are for developers, and according to Search Engine Journal, OpenAI’s documentation does not describe how language is chosen for rewritten queries.
• Things change monthly. In Peec’s data, ChatGPT’s use of translated Reddit fell by about 90% between April and May 2026.

What it means. Treat the steps above as cheap insurance, not as proven levers. Re-check your own category every few months, because the engines change faster than the studies.

FAQ

Should a Swedish company publish its website in English as well as Swedish?

For the pages that state what you sell and what it costs, yes, as separate pages that say the same thing. Swedish pages were the majority of citations for Swedish prompts on Google AI Overviews, Copilot and ChatGPT in Temso’s April 2026 data, but ChatGPT also adds English searches to most non-English sessions in Peec’s data. An English version covers that research. We found no study that tests this directly, so it is an inference from the evidence.

Does ChatGPT use Swedish sources when I ask in Swedish?

Partly. In Temso’s data from early 2026, 57.4% of the pages ChatGPT cited for Swedish prompts were in Swedish. In a February 2026 report, Peec says ChatGPT added English-language searches in about 78% of the non-English sessions it tracked, but it published no Swedish figure, and its prompts were ones its customers chose to track, run by browser automation, not a panel of consumer sessions.

Does Google AI Overviews prefer Swedish sources for Swedish questions?

In the only study we found that measured it, yes. Temso found that 85.1% of the pages Google AI Overviews cited for Swedish prompts were in Swedish, the highest of four engines. Other studies that define “local” differently rank AI Overviews lower, for example third of seven engines in MaxAEO’s B2B software data.

Will an AI assistant assume my company is in the US?

When it answers from memory without searching, it may. A peer-reviewed study of 32 models found they favour US answers to location-dependent questions on average, even when asked in other languages, though it did not test Swedish or web search, and the bias for law questions was small. Stating your country, city and currency on your pages is a sensible precaution, but we found no study that tests its effect.

Is machine-translating my Swedish site into English enough?

Not as a bulk job. Google’s spam policy names scraping content and running it through automated transformations such as translation to generate many low-value pages as scaled content abuse. That is not a ban on translating your own site, but it is a reason not to publish bulk, unchecked machine translation. Models that see two versions which disagree also tend to pick one answer without mentioning the conflict. Translate the pages that matter, have a person check them, and keep the facts identical in both languages.

Why do AI answers in Swedish cite Reddit threads written in English?

Reddit machine-translates threads and lets search engines index the translated pages. Peec found that 71.6% of Google AI Overviews’ Reddit citations and 65.4% of AI Mode’s in Sweden pointed to translated pages between March and June 2026. That is a share of Reddit citations only, and ChatGPT’s figure in Sweden fell to 1.62% in May 2026.

Related pages on this site

• What decides whether an AI recommends you, and what to do about it: the four gates and seven steps; this page adds the language and location layer.
• How to write content that LLMs retrieve and cite: how to write each language version so it is retrieved one section at a time.
• Can AI read your website?: the free crawler-access check from step 1.
• How AI agents choose products for consumers: for consumer products, which third-party sources move AI shopping agents.
• SEO and AEO for Swedish companies: how Copy for Originals produces checked content in Swedish and English.

Sources

All accessed October 24, 2026. Several sell AI-visibility tools; they are labelled in the text where their figures are used.

Vendor and press studies on language and citations

• Peec AI (Tomek Rudzki), “ChatGPT searches in English, even when you don’t”, byline February 12, 2026. Accessed October 24, 2026.
• Peec AI docs, “ChatGPT searches in English”. Accessed October 24, 2026.
• Peec AI docs, “Optimizing for Google AI vs ChatGPT”. Accessed October 24, 2026.
• Matt G. Southern, “ChatGPT Search Often Switches To English In Fan-Out Queries: Report”, Search Engine Journal, February 18, 2026. Accessed October 24, 2026.
• Temso AI (David Gregorian), “Lost in Translation: How AI Models Handle Local-Language Sources”, last updated April 20, 2026. Accessed October 24, 2026.
• Temso AI, “Which Sources Shape AI Answers About Universities? (Q3 2026)”, last updated October 1, 2026. Accessed October 24, 2026.
• Temso AI, “Print & Copy Shops in AI Search: Which Domains Get Cited (Q3 2026)”, last updated October 1, 2026. Accessed October 24, 2026.
• Temso AI, “Which Sources Shape AI Answers About Telehealth? (Q3 2026)”, last updated October 1, 2026. Accessed October 24, 2026.
• MaxAEO (Chris Han), “Local Sources AI Citations by Country”, July 24, 2026. Accessed October 24, 2026.
• Profound, “How query language reshapes AI citations”, April 21, 2026. Accessed October 24, 2026.
• Peec AI (Metehan Yesilyurt), “Translated Reddit Is Winning AI Citations.”, byline June 22, 2026. Accessed October 24, 2026.
• Glenn Gabe (GSQi), “AI Search, hreflang, and translated content. How do ChatGPT, Perplexity, Gemini, Copilot, and Claude handle multilingual queries?”, December 1, 2025. Accessed October 24, 2026.

Academic papers

• Mor-Lan et al., “Location Not Found: Exposing Implicit Local and Global Biases in Multilingual LLMs”, arXiv:2604.19292, ACL 2026 main conference. Accessed October 24, 2026.
• Wang and Suresh, “Which Institutional Frameworks Do Chatbots Assume? Auditing Jurisdictional Defaults in Multilingual LLMs”, arXiv:2606.00333 (preprint). Accessed October 24, 2026.
• Sharma, Murray and Xiao, “Faux Polyglot: A Study on Information Disparity in Multilingual Large Language Models”, NAACL 2025. Accessed October 24, 2026.
• Östling and Kurfalı, “Language Bias under Conflicting Information in Multilingual LLMs”, arXiv:2604.07123 (preprint, v2 June 2, 2026). Accessed October 24, 2026.
• Chen, Wang, Chen and Koudas, “Generative Engine Optimization: How to Dominate AI Search”, arXiv:2509.08919 (preprint, September 10, 2025). Accessed October 24, 2026.

Reddit (shareholder letters)

• Reddit, Q3 2024 Shareholder Letter (PDF). Accessed October 24, 2026.
• Reddit, Q3 2025 Shareholder Letter (PDF). Accessed October 24, 2026.

Google and AI platform documentation

• Google Search Central, “AI features and your website”. Accessed October 24, 2026.
• Google Search Central, “Managing multi-regional and multilingual sites”. Accessed October 24, 2026.
• Google Search Central, “Tell Google about localized versions of your page”. Accessed October 24, 2026.
• Google Search Central, “Spam policies for Google web search”. Accessed October 24, 2026.
• Google Search Central, “Translated results in Google Search”. Accessed October 24, 2026.
• Hema Budaraju, “AI Overviews expand to over 200 countries and territories, more than 40 languages”, Google, May 20, 2025. Accessed October 24, 2026.
• Åsa Wallenrud, “Google satsar stort på AI: Ny sökfunktion till Sverige”, Dagens PS, May 21, 2025. Accessed October 24, 2026.
• Google Sverige, “AI-läget lanseras i Google Sök i Sverige”, October 8, 2025. Accessed October 24, 2026.
• OpenAI, “Web search” (API guide). Accessed October 24, 2026.
• Anthropic, “Web search tool” (API docs). Accessed October 24, 2026.
• Perplexity, “Language filter” (Search API docs). Accessed October 24, 2026.

Swedish usage

• Internetstiftelsen, Svenskarna och internet 2026, AI chapter. Accessed October 24, 2026.
• Internetstiftelsen, Svenskarna och internet 2026, English edition. Accessed October 24, 2026.
• Internetstiftelsen, “Så använder svenskarna AI 2026”. Accessed October 24, 2026.

How to cite: Copy for Originals, “Does AI search answer Swedish questions from Swedish sources?,” October 24, 2026, https://tools.copyfororiginals.com/does-ai-search-cite-local-language-sources/.