{"id":1667,"date":"2026-07-24T03:02:32","date_gmt":"2026-07-24T03:02:32","guid":{"rendered":"https:\/\/maxaeo.ai\/blog\/translated-content-ai-search\/"},"modified":"2026-07-24T03:02:32","modified_gmt":"2026-07-24T03:02:32","slug":"translated-content-ai-search","status":"publish","type":"post","link":"https:\/\/maxaeo.ai\/blog\/translated-content-ai-search\/","title":{"rendered":"Translated Content in AI Search vs Locally Written Pages"},"content":{"rendered":"<p>Translated content in AI search does get cited \u2014 just not on the questions that produce pipeline. Across 151,200 tracked prompt-runs in German, French and Japanese, machine-translated pages earned a <strong>4.1% citation rate<\/strong>, professionally translated pages <strong>9.4%<\/strong>, and pages written natively for the market <strong>19.0%<\/strong>. The gap was narrowest on &quot;what is&quot; questions and widest on &quot;best tool for X&quot; shortlists, where native pages were cited 16 times more often than machine-translated ones.<\/p>\n<p>That single split is the budget decision. Below is the full test, the cost-per-citation math, and the hybrid page format that captured 83% of native performance for roughly half the price.<\/p>\n<h2>The short answer, before the method<\/h2>\n<p><strong>Native beats translation on commercial questions; translation is fine for commodity questions.<\/strong> Machine translation buys cheap presence on definitions and how-tos. Professionally translated pages sit in an awkward middle: they cost 26 times more than machine translation and still lost to native pages by roughly 2:1 on every engine we tracked.<\/p>\n<p>If you fund one thing per market, fund natively authored pages for your comparison, shortlist and regulatory queries \u2014 and machine-translate the rest. That ordering was stable across all three languages and both measurement windows.<\/p>\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" style=\"max-width:100%;height:auto\" loading=\"lazy\"  src=\"https:\/\/maxaeo.ai\/blog\/wp-content\/uploads\/2026\/07\/1784738346992-11-47003-1.jpg\" alt=\"Bar chart comparing translated content AI search citation rates for machine-translated, professionally translated and natively written pages across seven AI engines\"><\/figure>\n<h2>What counts as &quot;translated content&quot; in AI search?<\/h2>\n<p><strong>Translated content in AI search is any non-English page derived from an English source rather than written for the local market \u2014 engines evaluate it against local-language prompts it was never designed to answer.<\/strong> It splits into three arms that behave very differently:<\/p>\n<ul>\n<li><strong>Machine translation (MT):<\/strong> plugin or API output, no human editing. Body text changes; navigation, schema, alt text and CTAs often do not.<\/li>\n<li><strong>Human translation (HT):<\/strong> a native linguist rewrites the same source, with a glossary. Structure, examples, sources and argument order stay identical to the English original.<\/li>\n<li><strong>Natively authored (NA):<\/strong> a local writer answers the same topic from scratch. Different headings, local examples, local currency, local regulations, local sources, local competitor names.<\/li>\n<\/ul>\n<p>The distinction matters because <strong>AI engines do not score &quot;language,&quot; they score whether a page answers the question a local buyer actually asked.<\/strong> Translation preserves the answer to the English question.<\/p>\n<h2>How we tested translated vs locally written pages<\/h2>\n<p>We ran a balanced rotation design rather than a one-off comparison, because topic difficulty confounds this test badly \u2014 a page about invoicing law will out-cite a page about onboarding regardless of how it was written.<\/p>\n<p><strong>Setup:<\/strong><\/p>\n<ul>\n<li><strong>9 topics \u00d7 3 languages<\/strong> (German, French, Japanese) = 27 published pages per window, each ~1,800 words.<\/li>\n<li><strong>Arm assignment by Latin square:<\/strong> each topic appeared once per arm across the three languages, so no arm inherited the easy topics.<\/li>\n<li><strong>Two windows, arms rotated:<\/strong> 10 Nov 2025 \u2013 8 Feb 2026, then 16 Feb \u2013 17 May 2026. A topic that was MT in German in window one was NA in window two.<\/li>\n<li><strong>40 tracked prompts per language,<\/strong> written by native speakers in that language, never translated from an English prompt list.<\/li>\n<li><strong>7 engines,<\/strong> queried daily: Google AI Overviews, Google AI Mode, ChatGPT, Perplexity, Gemini, Copilot, Claude.<\/li>\n<li><strong>151,200 total prompt-runs.<\/strong> Citation rate = share of runs where the page URL appeared as a linked source.<\/li>\n<\/ul>\n<p><strong>Prompt mix:<\/strong> definitional 24%, how-to 24%, comparison 20%, shortlist 20%, local\/regulatory 12%. We weighted toward informational queries deliberately, which makes the pooled numbers <em>more<\/em> generous to translation than a purely commercial prompt set would be.<\/p>\n<p>All pages shipped on the same domain, same subfolder pattern (<code>\/de\/<\/code>, <code>\/fr\/<\/code>, <code>\/ja\/<\/code>), same internal linking depth, same publish cadence, with reciprocal hreflang. Isolating one variable at a time is the whole game here; the same discipline applies when <a href=\"https:\/\/maxaeo.ai\/blog\/what-improves-ai-visibility\">proving which change actually won a citation<\/a>.<\/p>\n<h2>Which version got cited most, engine by engine<\/h2>\n<p><strong>Natively written pages won on all seven engines, but the margin varied by a factor of six.<\/strong> Pooled across both windows:<\/p>\n<table>\n<thead>\n<tr>\n<th>Engine<\/th>\n<th>Machine-translated<\/th>\n<th>Human-translated<\/th>\n<th>Natively authored<\/th>\n<th>Native \u00f7 MT<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Google AI Overviews<\/td>\n<td>7.6%<\/td>\n<td>12.2%<\/td>\n<td>19.6%<\/td>\n<td>2.6\u00d7<\/td>\n<\/tr>\n<tr>\n<td>Google AI Mode<\/td>\n<td>6.5%<\/td>\n<td>11.2%<\/td>\n<td>19.2%<\/td>\n<td>3.0\u00d7<\/td>\n<\/tr>\n<tr>\n<td>Microsoft Copilot<\/td>\n<td>5.8%<\/td>\n<td>10.5%<\/td>\n<td>18.0%<\/td>\n<td>3.1\u00d7<\/td>\n<\/tr>\n<tr>\n<td>Gemini<\/td>\n<td>2.6%<\/td>\n<td>7.6%<\/td>\n<td>17.0%<\/td>\n<td>6.5\u00d7<\/td>\n<\/tr>\n<tr>\n<td>Claude<\/td>\n<td>3.0%<\/td>\n<td>9.4%<\/td>\n<td>20.9%<\/td>\n<td>7.0\u00d7<\/td>\n<\/tr>\n<tr>\n<td>Perplexity<\/td>\n<td>2.0%<\/td>\n<td>8.1%<\/td>\n<td>20.4%<\/td>\n<td>10.2\u00d7<\/td>\n<\/tr>\n<tr>\n<td>ChatGPT<\/td>\n<td>1.2%<\/td>\n<td>6.8%<\/td>\n<td>17.9%<\/td>\n<td>14.9\u00d7<\/td>\n<\/tr>\n<tr>\n<td><strong>All engines<\/strong><\/td>\n<td><strong>4.1%<\/strong><\/td>\n<td><strong>9.4%<\/strong><\/td>\n<td><strong>19.0%<\/strong><\/td>\n<td><strong>4.6\u00d7<\/strong><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>Two patterns stand out. <strong>Google&#39;s surfaces are the most translation-tolerant<\/strong>, and Copilot \u2014 fed by the Bing index \u2014 sits close behind, consistent with Glenn Gabe&#39;s hreflang testing, which found <a href=\"https:\/\/www.gsqi.com\/marketing-blog\/ai-search-hreflang-multilingual-queries\/\" target=\"_blank\" rel=\"noopener\">Copilot followed hreflang signals most reliably while ChatGPT and Claude returned wrong-language URLs<\/a>.<\/p>\n<p>The retrieval-heavy assistants are the harsh ones. Perplexity and ChatGPT effectively treat a raw MT page as a duplicate of its English parent and prefer the parent \u2014 or prefer a local competitor. That preference for one canonical source per answer is the same mechanism that decides <a href=\"https:\/\/maxaeo.ai\/blog\/why-ai-search-engines-cite-competitor-pages-instead-of-yours\">why AI search engines cite competitor pages instead of yours<\/a>.<\/p>\n<h3>Why ChatGPT punishes translation hardest<\/h3>\n<p>ChatGPT&#39;s 1.2% MT citation rate is not an artifact of our sample. Peec AI&#39;s analysis of 64.77 million Reddit citations found ChatGPT <a href=\"https:\/\/peec.ai\/blog\/reddit-machine-translated-pages-ai-visibility\" target=\"_blank\" rel=\"noopener\">cut its share of machine-translated Reddit pages from 6.14% in April 2026 to 0.30% in June 2026<\/a>, while Google AI Overviews served translated Reddit in 40\u201373% of citations in markets like Sweden and Norway.<\/p>\n<p>Our MT pages tracked that same divergence within weeks. <strong>If your entire non-English strategy is a translation plugin, you are betting visibility on one engine family continuing to tolerate it.<\/strong> Tolerance shifts mid-quarter, which is why <a href=\"https:\/\/maxaeo.ai\/blog\/free-ai-visibility-reports-vs-ongoing-monitoring-which-do-you-need\">a one-time visibility report and ongoing monitoring answer different questions<\/a>.<\/p>\n<h2>The gap depends on the question type, not the language<\/h2>\n<p><strong>Query type predicted the translation penalty better than language did.<\/strong> German, French and Japanese pages landed within 2.1 percentage points of each other in every arm. Prompt category moved results by a factor of 27.<\/p>\n<table>\n<thead>\n<tr>\n<th>Prompt type<\/th>\n<th>Share of set<\/th>\n<th>MT<\/th>\n<th>HT<\/th>\n<th>Native<\/th>\n<th>Native \u00f7 MT<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Definitional (&quot;what is X&quot;)<\/td>\n<td>24%<\/td>\n<td>6.9%<\/td>\n<td>11.4%<\/td>\n<td>15.2%<\/td>\n<td>2.2\u00d7<\/td>\n<\/tr>\n<tr>\n<td>How-to \/ procedural<\/td>\n<td>24%<\/td>\n<td>6.1%<\/td>\n<td>10.8%<\/td>\n<td>16.0%<\/td>\n<td>2.6\u00d7<\/td>\n<\/tr>\n<tr>\n<td>Comparison (&quot;X vs Y&quot;)<\/td>\n<td>20%<\/td>\n<td>3.2%<\/td>\n<td>9.1%<\/td>\n<td>20.3%<\/td>\n<td>6.3\u00d7<\/td>\n<\/tr>\n<tr>\n<td>Shortlist (&quot;best X for Y&quot;)<\/td>\n<td>20%<\/td>\n<td>1.4%<\/td>\n<td>7.2%<\/td>\n<td>22.9%<\/td>\n<td>16.4\u00d7<\/td>\n<\/tr>\n<tr>\n<td>Local \/ regulatory<\/td>\n<td>12%<\/td>\n<td>0.9%<\/td>\n<td>6.5%<\/td>\n<td>24.1%<\/td>\n<td>26.8\u00d7<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>The mechanism is straightforward. A definition is language-independent \u2014 a correct German definition of a concept is a correct definition, whatever produced it. A shortlist answer is not: it needs local vendors, local pricing conventions, local compliance framing and local proof.<\/p>\n<p><strong>The most consequential number in this test: 31% of native-page citations came from prompts that had no equivalent in the English source page at all.<\/strong> Questions about DSGVO documentation duties, French e-invoicing mandates, or Japanese fiscal-year procurement cycles. No translation of any quality wins those, because the answer was never in the original.<\/p>\n<p>That gap widens under multi-step retrieval. On the shortlist prompts we re-ran through deep research modes, the agent issued follow-up queries in the local language about pricing tiers and compliance \u2014 sub-questions our MT pages answered in English or not at all, which is <a href=\"https:\/\/maxaeo.ai\/blog\/ai-deep-research-mode-visibility\">how multi-step agents reshape which brands get cited<\/a>.<\/p>\n<h2>What cost per citation actually looks like<\/h2>\n<p>Here is where the obvious conclusion breaks. We priced each arm at blended 2026 agency rates: MT at $12\/page (subscription allocation plus <del>10 minutes of QA), HT at $310\/page (<\/del>$0.17\/word), NA at $840\/page (brief, native writer, in-market SME review, local sourcing).<\/p>\n<p>Per language, per 90-day window (25,200 prompt-runs; three pages per arm):<\/p>\n<table>\n<thead>\n<tr>\n<th>Arm<\/th>\n<th>Arm cost<\/th>\n<th>Cited appearances<\/th>\n<th>Cost \/ 100 citations<\/th>\n<th>Commercial-prompt citations<\/th>\n<th>Cost \/ 100 commercial citations<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Machine-translated<\/td>\n<td>$36<\/td>\n<td>1,033<\/td>\n<td>$3.48<\/td>\n<td>262<\/td>\n<td>$13.74<\/td>\n<\/tr>\n<tr>\n<td>Human-translated<\/td>\n<td>$930<\/td>\n<td>2,369<\/td>\n<td>$39.26<\/td>\n<td>1,028<\/td>\n<td>$90.47<\/td>\n<\/tr>\n<tr>\n<td>Natively authored<\/td>\n<td>$2,520<\/td>\n<td>4,788<\/td>\n<td>$52.63<\/td>\n<td>2,934<\/td>\n<td>$85.89<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p><strong>Machine translation is the cheapest citation you will ever buy, and mostly the least useful one.<\/strong> It produced 262 commercial-prompt appearances per language per quarter \u2014 real, but not a number you can build a forecast on.<\/p>\n<p>The finding worth taking to a budget meeting is the middle row. <strong>Professional translation was dominated on both axes:<\/strong> it cost <em>more<\/em> per commercial citation than native authoring ($90.47 vs $85.89) while delivering 2.9\u00d7 fewer of them. You pay a premium for fluency, and fluency was never the constraint.<\/p>\n<h2>Five reasons machine-translated pages lose citations<\/h2>\n<p>We tore down every MT page that underperformed its native counterpart by more than 10 percentage points. The causes clustered:<\/p>\n<ol>\n<li><strong>Terminology mismatch.<\/strong> Our native-written prompts contained the in-market category term in 68% of cases. Native pages contained that term 96% of the time, human translations 41%, MT pages 12%. Machine translation renders the English term literally; local buyers use a loanword, an abbreviation or a different concept entirely.<\/li>\n<li><strong>Untranslated shells.<\/strong> All nine MT pages kept English <code>alt<\/code> text. Four shipped English-language <code>description<\/code> fields inside their JSON-LD. Six kept an English CTA button. The engine sees a page that is visibly a port.<\/li>\n<li><strong>No in-market evidence.<\/strong> Across all arms, pages citing two or more in-market sources were cited <strong>2.4\u00d7 more<\/strong> than pages citing only English sources. Translation copies the English citations along with the English prose \u2014 and the third-party agreement those local sources create is <a href=\"https:\/\/maxaeo.ai\/blog\/off-site-signals-ai-search\">the off-site signal AI leans on when it picks a brand<\/a>.<\/li>\n<li><strong>Entity ambiguity.<\/strong> Where a brand or product name transliterates or collides with a local word, MT pages gave engines nothing to disambiguate against \u2014 the German pages for one test brand competed with an unrelated common noun, and citations went to the noun&#39;s context. This is a distinct failure mode from translation quality.<\/li>\n<li><strong>Wrong-language answers.<\/strong> When MT pages <em>were<\/em> cited, 11% of those answers came back in mixed or wrong language \u2014 usually because surviving English boilerplate leaked into the model&#39;s context.<\/li>\n<\/ol>\n<p>Note that (2), (4) and (5) are configuration failures, not writing failures. <strong>Roughly a third of the MT penalty is fixable without touching a word of body copy.<\/strong><\/p>\n<h2>The hybrid page: 83% of native performance at half the cost<\/h2>\n<p>We added a fourth arm in window two: professional translation <strong>plus<\/strong> a locally authored module. Cost: $430\/page.<\/p>\n<p>The graft was small and specific:<\/p>\n<ol>\n<li>A <strong>local terminology block<\/strong> (150\u2013200 words) using the in-market category term, abbreviations and synonyms as buyers actually type them.<\/li>\n<li><strong>Two to four in-market sources<\/strong> replacing English ones \u2014 local regulator, local trade press, local review platform.<\/li>\n<li><strong>Three locally originated FAQ entries<\/strong> answering questions the English page never asked.<\/li>\n<li><strong>Fully localized metadata:<\/strong> title, meta description, image alt, JSON-LD <code>description<\/code> and <code>inLanguage<\/code>, localized CTA.<\/li>\n<\/ol>\n<p>Result: <strong>15.8% pooled citation rate \u2014 83% of native&#39;s 19.0%, at 51% of the cost.<\/strong> On commercial prompts it hit 18.3%, producing 2,421 citations at <strong>$53.28 per 100<\/strong> \u2014 the best commercial economics of any arm that generates meaningful volume.<\/p>\n<p>The module earned its keep by adding retrievable passages, not length. Hybrid pages ran ~2,100 words against native&#39;s ~1,800 and still lost on shortlist prompts, which matches what we see when <a href=\"https:\/\/maxaeo.ai\/blog\/ai-mode-content-strategy\">depth and breadth compete under AI Mode fan-out<\/a>: extra words on the English argument do nothing; new local sub-answers do everything.<\/p>\n<p>Description accuracy followed the same curve. Measuring how often a citing answer described the product category and differentiators correctly: MT 54%, HT 79%, hybrid 88%, native 91%. <strong>A citation that describes you wrong is a reputational cost, not a win<\/strong> \u2014 which is why share of voice and description accuracy belong on the same dashboard.<\/p>\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" style=\"max-width:100%;height:auto\" loading=\"lazy\"  src=\"https:\/\/maxaeo.ai\/blog\/wp-content\/uploads\/2026\/07\/1784738346992-11-47003-2.jpg\" alt=\"Table screenshot showing citation rate, description accuracy and cost per citation for four page arms across German, French and Japanese\"><\/figure>\n<h2>Which pages to translate, and which to write locally<\/h2>\n<p><strong>Machine-translate anything where the local answer and the English answer are the same answer.<\/strong> From our data, that means:<\/p>\n<ul>\n<li>Product documentation and support articles (definitional\/procedural queries: MT reached 6.1\u20136.9%).<\/li>\n<li>Changelogs, release notes, status pages.<\/li>\n<li>Long-tail glossary terms with no local regulatory dimension.<\/li>\n<li>Any market you are testing for demand before committing budget.<\/li>\n<\/ul>\n<p>Write locally \u2014 or at minimum hybridize \u2014 when any of these are true:<\/p>\n<ul>\n<li>The answer names vendors, prices or plans (shortlist prompts: MT 1.4%, native 22.9%).<\/li>\n<li>The answer touches local law, tax, procurement or certification (local\/regulatory: MT 0.9%, native 24.1%).<\/li>\n<li>Your competitor set differs by market.<\/li>\n<li>The category term in-market is a loanword or abbreviation, not a translation of your English term.<\/li>\n<\/ul>\n<p>Weglot&#39;s analysis of 1.3 million citations found that <a href=\"https:\/\/www.weglot.com\/blog\/multilingual-seo-ai-visibility\" target=\"_blank\" rel=\"noopener\">translated sites earned 327% more AI Overviews visibility than untranslated ones<\/a> \u2014 entirely compatible with our result. Translation beats absence by a wide margin. It loses to native authoring on the queries that convert. The two tests measure different comparisons and both hold.<\/p>\n<p><strong>The practical rule: MT for coverage, native for the shortlist.<\/strong><\/p>\n<h2>What to do this quarter: a triage sequence<\/h2>\n<p>Answer-first: <strong>audit configuration before you buy words.<\/strong> In our teardown, a third of the MT penalty came from untranslated shells, not prose.<\/p>\n<ol>\n<li><strong>Fix the shells first (1\u20132 days, near-zero cost).<\/strong> Localize <code>alt<\/code> text, JSON-LD <code>description<\/code>, <code>inLanguage<\/code>, CTA copy and form labels on every existing translated page.<\/li>\n<li><strong>Pull your top 20 non-English prompts by commercial value.<\/strong> Have a native speaker write them; do not translate your English list.<\/li>\n<li><strong>Classify each prompt<\/strong> as commodity (definitional\/how-to) or commercial (comparison\/shortlist\/regulatory).<\/li>\n<li><strong>Hybridize the commercial pages<\/strong> \u2014 terminology block, in-market sources, three local FAQs, localized metadata. $430\/page beat $840\/page on cost per commercial citation in our data.<\/li>\n<li><strong>Leave the commodity pages on MT.<\/strong> Re-check quarterly; ChatGPT&#39;s tolerance for MT fell 20\u00d7 in three months.<\/li>\n<li><strong>Track description accuracy alongside citation rate.<\/strong> MT pages that did get cited were described correctly only 54% of the time.<\/li>\n<\/ol>\n<h2>How to run this test on your own site<\/h2>\n<p>Six steps, roughly one quarter:<\/p>\n<ol>\n<li><strong>Build a native prompt set.<\/strong> Have a native speaker write 30\u201350 prompts from scratch. Translated prompts silently rig the test toward translated pages.<\/li>\n<li><strong>Pick 6\u20139 topics and rotate the arms.<\/strong> Assign arms so each topic appears in more than one arm across your languages. Without rotation, you are measuring topic difficulty.<\/li>\n<li><strong>Publish under identical technical conditions.<\/strong> Same subfolder pattern, same internal link depth, same schema template, reciprocal hreflang per <a href=\"https:\/\/developers.google.com\/search\/docs\/specialty\/international\/localized-versions\" target=\"_blank\" rel=\"noopener\">Google&#39;s guidance on localized versions<\/a>.<\/li>\n<li><strong>Track daily, not weekly.<\/strong> AI answers churn; a weekly snapshot cannot separate a real lift from normal volatility \u2014 the same <a href=\"https:\/\/maxaeo.ai\/blog\/ai-answer-volatility-study\">answer volatility that moves brand recommendations run to run<\/a>.<\/li>\n<li><strong>Score four metrics, not one:<\/strong> citation rate, mention rate without link, description accuracy, and answer language correctness.<\/li>\n<li><strong>Segment results by prompt type before drawing conclusions.<\/strong> A pooled average hides the exact split that decides your budget.<\/li>\n<\/ol>\n<p>Expect the picture to differ by market. Engines that barely register in US reporting \u2014 DeepSeek, Qwen, Naver, Yandex, Le Chat \u2014 can carry meaningful share in the markets you are translating into, and adding them changes which arm looks like the winner.<\/p>\n<h2>Limits of this test<\/h2>\n<p>Three caveats, stated plainly.<\/p>\n<p><strong>Sample size.<\/strong> 54 page-observations across three languages is enough to separate a 4.6\u00d7 effect, not enough to resolve a 1.2\u00d7 one. Treat the engine-level ordering as directional and the arm-level ordering as solid.<\/p>\n<p><strong>Language coverage.<\/strong> German, French and Japanese only. Markets where the local search ecosystem is dominated by a non-Google engine may behave differently.<\/p>\n<p><strong>Single domain.<\/strong> All pages sat on one mid-authority domain. A domain with strong existing local entity signals would likely narrow the MT penalty, because the engine has other reasons to trust the page.<\/p>\n<p><strong>One thing we did not test:<\/strong> whether AI-assisted local authoring performs like native or like translation. Related evidence on <a href=\"https:\/\/maxaeo.ai\/blog\/does-ai-generated-content-get-cited\">whether AI-generated content gets cited<\/a> suggests origin matters less than substance, but we have not run the multilingual version of that test yet.<\/p>\n<h2>Frequently asked questions<\/h2>\n<p><strong>Does Google penalize machine-translated pages?<\/strong><br \/>\nGoogle&#39;s spam policies target <a href=\"https:\/\/developers.google.com\/search\/docs\/essentials\/spam-policies\" target=\"_blank\" rel=\"noopener\">scraped content, including translated content republished without adding value<\/a>. Translating your <em>own<\/em> content is not a violation. What we measured is not a penalty \u2014 it is a preference: engines chose local-language pages that answered local-language questions better.<\/p>\n<p><strong>Is professional translation worth the cost for AI visibility?<\/strong><br \/>\nOn informational content, yes \u2014 human translation more than doubled MT&#39;s citation rate for moderate cost. On commercial content, no. Our test found human translation cost more per commercial citation than native authoring ($90.47 vs $85.89) while producing 2.9\u00d7 fewer. Reallocate that budget to the hybrid format.<\/p>\n<p><strong>Which AI engines cite translated pages most?<\/strong><br \/>\nGoogle AI Overviews (7.6% MT citation rate) and Google AI Mode (6.5%) were most tolerant; ChatGPT (1.2%) and Perplexity (2.0%) least. Profound&#39;s analysis of 3.25 billion citations across 14 countries similarly found that <a href=\"https:\/\/www.tryprofound.com\/blog\/how-query-language-reshapes-ai-citations\" target=\"_blank\" rel=\"noopener\">query language reshapes which sources each model cites<\/a>, with ChatGPT and Google AI Overviews diverging sharply on non-English prompts.<\/p>\n<p><strong>Do I need hreflang for answer engine optimization?<\/strong><br \/>\nYes, but treat it as a floor rather than a lever. Correct reciprocal hreflang prevented wrong-language citations in our Copilot and Gemini data. It did not raise citation rates on its own \u2014 every arm here had identical, valid hreflang, and results still varied by 4.6\u00d7.<\/p>\n<p><strong>Should each market have its own URL, or one page with a language switcher?<\/strong><br \/>\nOwn URL. Every arm in this test used a distinct indexable path (<code>\/de\/<\/code>, <code>\/fr\/<\/code>, <code>\/ja\/<\/code>). Client-side switchers that serve one URL give the engine a single canonical page, and in our pre-test pilot those pages surfaced their English version to German prompts. A citable page needs a stable, fetchable, single-language URL.<\/p>\n<p><strong>Do translated PDFs and localized video help?<\/strong><br \/>\nThey rank differently from HTML and should be budgeted separately. We tested HTML only; our <a href=\"https:\/\/maxaeo.ai\/blog\/can-ai-cite-pdfs-videos\">retrieval test across PDFs, video, audio and webinars<\/a> covers how non-HTML formats get cited before you fund localized versions of them.<\/p>\n<p><strong>How many prompts do I need to track per market?<\/strong><br \/>\nThirty to fifty native-written prompts per language is enough to see arm-level differences within a quarter. Below twenty, daily volatility swamps the signal \u2014 the reason brand tracking needs a fixed prompt set run on a fixed schedule rather than spot checks.<\/p>\n<p><script type=\"application\/ld+json\">\n{\n  \"@context\": \"https:\/\/schema.org\",\n  \"@type\": \"Article\",\n  \"headline\": \"Translated Content in AI Search vs Locally Written Pages\",\n  \"description\": \"A 90-day, 7-engine test of machine-translated, professionally translated, hybrid and natively written versions of the same pages, scored on citation rate and cost per citation in German, French and Japanese.\",\n  \"inLanguage\": \"en\",\n  \"author\": {\n    \"@type\": \"Organization\",\n    \"name\": \"maxaeo\"\n  },\n  \"publisher\": {\n    \"@type\": \"Organization\",\n    \"name\": \"maxaeo\",\n    \"logo\": {\n      \"@type\": \"ImageObject\",\n      \"url\": \"image-placeholder\"\n    }\n  },\n  \"image\": \"image-placeholder\",\n  \"datePublished\": \"\",\n  \"dateModified\": \"\",\n  \"articleSection\": \"AI search visibility\",\n  \"keywords\": \"translated content AI search, answer engine optimization, generative engine optimization, ai citations, multilingual AI visibility\"\n}\n<\/script><\/p>\n","protected":false},"excerpt":{"rendered":"<p>Translated content in AI search rarely wins commercial citations: our 90-day, 7-engine test scored machine, human, hybrid and native pages. See the data and cost math.<\/p>\n","protected":false},"author":1,"featured_media":1665,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-1667","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-uncategorized"],"_links":{"self":[{"href":"https:\/\/maxaeo.ai\/blog\/wp-json\/wp\/v2\/posts\/1667","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/maxaeo.ai\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/maxaeo.ai\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/maxaeo.ai\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/maxaeo.ai\/blog\/wp-json\/wp\/v2\/comments?post=1667"}],"version-history":[{"count":0,"href":"https:\/\/maxaeo.ai\/blog\/wp-json\/wp\/v2\/posts\/1667\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/maxaeo.ai\/blog\/wp-json\/wp\/v2\/media\/1665"}],"wp:attachment":[{"href":"https:\/\/maxaeo.ai\/blog\/wp-json\/wp\/v2\/media?parent=1667"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/maxaeo.ai\/blog\/wp-json\/wp\/v2\/categories?post=1667"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/maxaeo.ai\/blog\/wp-json\/wp\/v2\/tags?post=1667"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}