Why Your English-Only Site Disappears in Multilingual AI Search
Table of Contents
An English-only website disappears in multilingual AI search because answer engines prefer evidence written in the same language as the question. When a buyer asks ChatGPT, Gemini, Perplexity, or Google AI Overviews in Spanish, French, German, or Japanese, the system looks for Spanish, French, German, or Japanese pages, reviews, and entity names. Your English homepage can still exist. It just stops being useful evidence.
I'm William Spurlock, an AI Solutions Architect and Fractional AI CTO. I spend my days on AI visibility — getting a business named in the answers buyers already ask. I have spent 20,000+ hours on agentic systems and AEO work. The pattern I keep seeing in August 2026 is not a mystery ranking leak. It is a language mismatch.
This spoke sits under how to get ChatGPT and Perplexity to recommend your business. That pillar covers third-party proof and entity consistency. This post owns one query: why does an English-only website disappear in multilingual AI search results? I am talking about language, not country expansion. A US company that never ships overseas still disappears when a Spanish-speaking buyer asks in Spanish.
The cost is quiet. You keep ranking for English Google queries. Your English prompt tests still name you. Then a buyer asks the same question in another language and your competitors who published one real locale page get the citation. The lead never arrives as a "lost ranking." It never arrives at all.
Why Does an English-Only Website Disappear in Multilingual AI Search Results? #
It disappears because the model has almost no on-language proof that your brand answers that question. Retrieval pulls pages, reviews, and mentions that match the query language. Synthesis then quotes the sources it can verify in that language. An English-only site gives the model English proof. That is the wrong pile.
Google states this in plain language for Search. In the multi-regional and multilingual sites guide (last updated 10 December 2025), Google defines a multilingual website as a site that offers content in more than one language, and says Search tries to find pages that match the language of the searcher. A multi-regional site is a different problem — country targeting. I am staying on the language side.
Answer engines are not a carbon copy of Google Search. OpenAI does not publish a formula that says "drop English URLs from Spanish answers." ChatGPT search (updated mid-August 2026) says it ranks results on multiple factors and does not guarantee placement. I still treat language match as a load-bearing factor because the visible behavior is consistent: a Spanish prompt gets Spanish citations when those citations exist.
The disappearance is a retrieval miss, not a penalty #
I do not tell owners "Google punished your English site." Nothing got deleted. Three things fail at once:
- Crawl never sees a second language. If the only URL is English, there is no Spanish document to index.
- Evidence language does not match the question. Reviews, FAQs, and product names stay English, so they are weak support for a non-English answer.
- Third-party mentions stay English. Directories, listicles, and Reddit threads that name you do it in English, so they help English prompts and sit idle for other languages.
That is why an English-only site can "win" English ChatGPT tests and vanish the minute you change the prompt language. The brand did not get weaker. The evidence language stayed fixed.
What "disappear" looks like in a real prompt test #
I run the same buying question in English and in one other language the owner actually receives. The English run often names the business. The second-language run names two competitors and a directory. Sometimes it invents a generic category answer and cites no one. The owner's site is not in the citation list. That is disappearance.
| Prompt language | What the engine can quote | What happens to an English-only site |
|---|---|---|
| English | Your pages, English reviews, English listicles | You can still get named if the rest of your AEO is in place |
| Spanish / French / German / Japanese | Locale pages, on-language reviews, local-language entity strings | You drop out unless a third party already described you in that language |
| Mixed (English brand name + other-language question) | Brand string plus whatever on-language proof exists | Brand mention without a recommendation — or a competitor who has both |
If you only test in English, you are auditing the easy half of the problem.
What this post is not #
This is not a country-expansion brief. I am not writing about "getting cited outside the United States" as a market problem. A business can stay domestic and still lose Spanish, French, or Mandarin prompts from people who already buy from them. Language is the unit. Country targeting, ccTLDs, and shipping zones are a different brief.
I also am not claiming OpenAI, Google Gemini, or Perplexity published a "drop English URLs" ranking clause. They did not. I am claiming the evidence pile is language-shaped, and an English-only site only fills the English pile.
What Language Signals Do AI Answer Engines Actually Read? #
They read the language of the visible page, the language of the reviews attached to the entity, the strings people use for your name, and — for Google — explicit hreflang pairs that point to real locale URLs. They do not read a translate widget, a cookie, or an html lang tag as proof that a second language exists.
Google is explicit about the last part. The localized versions documentation (last updated 22 December 2025) says Google does not use hreflang or the HTML lang attribute to detect the language of a page. It uses algorithms on the visible content. hreflang tells Google which other URL is the match for a language or region. If there is no other URL, the annotation has nothing to point at.
I group the signals owners can actually ship into four buckets. This is the table I walk in audits.
| Signal | What it is | What it tells a crawler or answer engine | What fails on an English-only site |
|---|---|---|---|
| hreflang | Reciprocal annotations (HTML, HTTP header, or sitemap) that pair language versions | "This English URL has a Spanish sibling at this other URL" | There is no sibling. Tags that point at the same English page, or at a /es/ URL that 404s, get ignored |
| Locale pages | Separate, crawlable URLs whose main content is in the target language | "Here is a Spanish document I can quote" | One English URL plus a widget. Googlebot usually crawls from the US and does not send Accept-Language (Google, 10 Dec 2025) |
| Entity names | Legal name, trade name, product names, and the strings people actually search | "This brand is the same thing in both languages" | English-only name with no alternateName, no transliteration, no on-language product title. The model cannot join "Acme Roofing" to the local-language query |
| Reviews language | The language of Google, Yelp, industry, and on-site reviews | "People who speak this language already describe this business" | All reviews in English. A Spanish answer has no Spanish testimony to lift |
How I treat each signal in practice #
- hreflang is a pointer, not a translation. I never ship
hreflang="es"until the Spanish URL returns 200 and the Spanish URL points back. - Locale pages must carry the offer, the FAQ, and the proof in that language. A translated nav bar with English body copy is still an English page. Google says localized versions are only treated as duplicates if the main content stays untranslated (localized versions, 22 Dec 2025).
- Entity names need a written map: English name, local-language name if you use one, and the misspellings customers already type.
- Reviews language is the cheapest third-party proof you can add without buying ads. One honest Spanish review is more citable in a Spanish answer than a perfect English testimonial.
schema.org inLanguage and availableLanguage sit under these four. They label the language of a CreativeWork or the languages a service can be used in (schema.org/inLanguage, Schema.org V30.0, 19 March 2026). They do not create a locale page. I add them after the URL exists.
If you only "fix hreflang" on an English-only site, you annotated a ghost. The other three signals are what the model quotes.
How Does hreflang Tell Google Which Language Version Exists? #
hreflang tells Google, "this page has a language sibling at that other URL." It does not translate anything. It does not detect language. It only pairs URLs. If the sibling is missing, broken, or is the same English page, the annotation does no work.
Google accepts three equivalent delivery methods: HTML link tags, HTTP Link headers, or sitemap annotations (localized versions, 22 Dec 2025). Pick one and keep it correct. Running all three at once does not help Search and makes drift more likely.
The rules I refuse to skip #
Google's own guidelines, compressed into owner language:
| Rule | Why it matters | English-only failure mode |
|---|---|---|
| Every language version lists itself and every other version | Missing self-reference or missing return link can cause Google to ignore the set | You add hreflang="es" on the English page and never put the return tag on /es/ |
Alternate URLs are fully qualified (https://…) |
Relative paths and protocol-relative URLs are invalid | href="/es/" looks fine to a human and is wrong for Google |
| Return links are reciprocal | "If page X links to page Y, page Y must link back to page X" | A vendor injects tags on the homepage only |
| Language codes are ISO 639-1, optional region is ISO 3166-1 Alpha 2 | Unsupported codes get ignored. Google calls out es-419 and UK as examples that fail |
hreflang="uk" or hreflang="sp" sits in the head and does nothing |
x-default is a fallback for unmatched language settings |
It is for the selector or catch-all, not a fake second language | You set x-default and call the site multilingual |
Google also says you can omit some languages on some pages if a full bidirectional mesh gets unmanageable, but new language pages should still pair back to the originating language. For most owners I work with, that means every Spanish service page points at the English service page and back. Not a homepage-only cluster.
What hreflang cannot do #
I get asked to "add hreflang so ChatGPT finds us in French." That request mixes two systems.
- Google Search documents
hreflangas a hint for which URL to show a searcher. It is a hint, not a command. Google still determines page language from visible content. - ChatGPT, Perplexity, Claude Opus 4.8, and Gemini 3.1 Pro do not publish a public
hreflangconsumer. I have not seen an official OpenAI or Google Gemini doc that says those products readhreflangthe way Search does. Treathreflangas a Google Search signal that keeps locale URLs discoverable. The answer engines then have a real URL to crawl.
If you need ChatGPT to cite you in French, you still need a French page, French-language third-party mentions, and French-language reviews. hreflang is the map. The locale page is the territory.
A clean pairing vs a decorative one #
| Setup | What Google can do | What an answer engine can quote |
|---|---|---|
English URL + Spanish URL, both 200, reciprocal hreflang, main content translated |
Serve the Spanish URL to a Spanish-language searcher | The Spanish page, if it answers the question |
English URL + hreflang="es" pointing at the same English URL |
Nothing useful | Still English |
English URL + /es/ that auto-redirects by IP or browser language |
Risk of hiding a version from crawl | The crawler may only ever see English |
English URL + sitemap hreflang to a staging domain |
Tags ignored or wasted | Nothing public to cite |
I would rather ship two URLs and no tags than tags and one URL. You can add the annotations in a day once the pages are live. You cannot annotate your way into a language that does not exist.
Why a Translate Widget Is Not a Locale Page #
A translate widget changes pixels in one browser. A locale page is a separate, crawlable URL whose main content is written in that language. Answer engines cite documents. A widget does not create a document.
Google's multilingual guidance is blunt. Use different URLs for different language versions rather than cookies or browser settings to swap the language on one page. If you change content in place or reroute by language settings, Google might not find every variation. Googlebot usually originates from the USA and does not send Accept-Language (managing multi-regional and multilingual sites, 10 Dec 2025).
That paragraph is why "we have Google Translate on the site" fails as an AI-visibility plan. The crawler fetches the English HTML. The widget runs later, in a user's browser, if it runs at all. There is no Spanish URL for Search, for OAI-Searchbot, or for Perplexity's fetch to collect.
Locale page vs the lookalikes #
| Approach | Separate URL? | Crawlable main content in the target language? | Useful for multilingual AI search? |
|---|---|---|---|
Locale page (/es/servicios/, es.example.com, or example.es) |
Yes | Yes, if you actually wrote or professionally localized the body | Yes — this is the unit I ship |
| Translate widget (Google Translate, Weglot overlay, JS swap) | No | No. Source HTML stays English | No. Humans may read it. Crawlers usually do not |
Cookie / Accept-Language swap on one URL |
No | Only the variant the crawler happens to receive | No. Google says this hides variants |
| Auto-redirect by guessed language | Sometimes the user never sees the other URL | The crawler may never reach the other version | No. Google tells you not to auto-redirect |
| PDF or image of a translated brochure | Maybe | Weak. Hard to extract, easy to ignore | Rarely. I want HTML |
Google also tells you to let the user switch language with visible links, and to keep one language per page instead of side-by-side translations. That last point matters for extraction. A page that mixes English body copy with a Spanish sidebar gives a language detector a messy primary language. Common Crawl's own stats run CLD2 on HTML and assign a primary language. Mixed pages become noisy training and retrieval objects.
URL shape I actually recommend #
Google's table is still the right starting point. I default to subdirectories on one host (example.com/es/) for most owner sites: one sitemap, one CMS, one certificate, one Search Console property to reason about. Subdomains work when a team already treats es. as a separate app. A ccTLD is a country signal, which is the multi-regional problem I am not solving in this post.
What I will not ship:
- Query-parameter locales (
?lang=es). Google lists URL parameters as not recommended. - A single English URL that "detects" language.
- A locale folder that contains machine-translated thin copies of every blog post and no money pages.
What "main content in the target language" means #
I want the H1, the offer, the FAQ, the pricing or package language, and the proof on that URL to be in the target language. Nav and footer are not enough. Google warns that translating only boilerplate while leaving user-generated or body content in one language creates a bad result — the same content repeating with different chrome (managing multi-regional and multilingual sites, 10 Dec 2025).
If budget is tight, I would rather have three Spanish pages that answer buying questions than fifty auto-translated blog posts. Answer engines extract the first clear answer. A thin translation of a 2019 English rant is not that.
Entity Names That Only Exist in English #
If the only string that identifies you is an English brand or product name, a non-English query has nothing to lock onto. The model can understand the English name. Understanding is not the same as retrieving you for a local-language question that never uses that string.
This is the part owners skip because the logo feels universal. The logo is not a retrieval key. The keys are the words in the question and the words on pages and reviews.
The name map I collect in week one #
| String type | English-only example | What I add for a second language | Why answer engines care |
|---|---|---|---|
| Legal / trade name | "North Shore Dental Group" | Keep the legal name. Add the local-language descriptor people actually say ("clínica dental North Shore") | Queries often use the category word in the buyer's language plus a fragment of the brand |
| Product or service name | "Same-Day Crown" | Publish the local-language service title on the locale page and in schema name / alternateName |
Buying questions use the service noun, not your trademark |
| Transliteration | A Latin-script brand asked about in Japanese or Arabic | A documented kana, katakana, or Arabic spelling if customers already use one | Search partners rewrite queries. If the rewritten query uses the local script, your English-only string misses |
schema.org names |
name in English only |
name in the page language, alternateName for the English brand and common misspellings |
inLanguage labels the page. alternateName keeps the entity joinable (schema.org/inLanguage) |
| Third-party listings | GBP, directories, "best of" posts all English | At least one trusted listing that uses the local-language category + your name | ChatGPT and Perplexity synthesize third-party proof. English-only listings help English prompts |
I do not invent a cute translated brand. I write down the strings that already appear in inbox subject lines, invoices, and voicemail. Then I put those strings on the locale page in the first 100 words, in the H1 or H2, and in JSON-LD.
The join problem, in one sentence #
An English-only entity is easy to recommend when the prompt contains the English brand. It is hard to recommend when the prompt is a category question in another language and never mentions you. That is the default buying prompt. Nobody starts with your name. They start with the job to be done.
The recommendation pillar is the longer treatment of third-party proof. The language-specific cut is this: proof that only exists in English does not travel into a Spanish or Japanese synthesis. You need the name to appear next to an on-language category word on a page the engine can fetch.
Why Review Language Quietly Filters You Out #
Review language is third-party proof in a specific tongue. If every review attached to your entity is English, a non-English answer has no customer voice to lift. The model can still know you exist. It will prefer a competitor who has on-language reviews it can quote.
I have not seen an OpenAI, Google, or Perplexity document that says "we filter businesses by review language." Treat that as undocumented. The mechanism I trust is simpler: synthesis quotes evidence. A Spanish answer that cites "clientes dicen que…" needs Spanish sentences. English five-star text does not supply those sentences.
Where review language actually lives #
| Surface | What language the engine sees | English-only failure | What I change first |
|---|---|---|---|
| Google Business Profile reviews | The review body as written | 40 English reviews, zero in the language of the prompt | Ask real customers who already write in that language to review in that language. Do not ghostwrite |
| Industry directories / associations | Listing blurbs and reviews | English-only profile copied from the homepage | A short on-language description + one on-language review if the directory allows it |
| On-site testimonials | HTML on your domain | English quotes on every locale URL | Put on-language quotes on the locale page. Do not machine-translate a testimonial and keep the English name as if they wrote it |
| YouTube / podcast comments | Weak, noisy | Irrelevant unless the transcript is in the target language | Skip until locale pages and GBP language exist |
| Reddit / forums | High citation value when they name you | Threads about you are English | You cannot fake this. You can earn one honest thread by being present in the language community you already serve |
What I will not do #
I will not run a "get 20 Spanish reviews this week" campaign with templates. That is fake proof, and it is the first thing a careful answer engine or a human reporter will smell. I also will not auto-translate English reviews onto a locale page and mark them as local customers.
What I will do: identify customers who already communicate in the target language — they email you in it, they call in it, they pay invoices in it — and ask them for a review on the surface that already ranks for your entity. The ask is in their language. The review stays in their language.
Why this filters you even when the locale page is decent #
A locale page is first-party. Reviews are third-party. Getting recommended depends more on the second than the first. I have watched English-only review profiles lose a Spanish prompt to a smaller competitor with six Spanish reviews and a worse website. The engine had quotable Spanish proof. We did not.
If you serve bilingual customers today and all of your public reviews are English, you already have a language-visibility hole. You do not need a new country. You need reviews in the language those customers already use.
How ChatGPT, Gemini, and Perplexity Handle Language in August 2026 #
Each product documents a UI language setting. None of them publish a ranking spec that maps query language to a crawl language. I treat official docs as facts and the ranking step as an observed pattern. If a vendor has not written it down, I will not pretend they did.
What is documented #
| Product | What the vendor says as of August 2026 | What that does not say |
|---|---|---|
| ChatGPT (OpenAI) | Browser or device language is detected and the UI language updates for supported languages. Users can override it in Settings (OpenAI Help, Aug 2026). Search ranks on multiple factors, may rewrite queries for partners, and may use approximate IP location. Placement is not guaranteed (Searching the web with ChatGPT). Sites must allow OAI-Searchbot. | It does not say English pages are excluded from Spanish answers. It does not say hreflang is consumed. It does not publish a language-match weight |
| Gemini (Google) | The Gemini web app follows browser or device language. Gemini Apps can understand and reply in any supported language when you type, even if the UI language is different. Some features are not available in every language (Change Gemini's language). The web app is available in more than 70 languages (Gemini availability). | App availability in 70+ languages is not a promise that your English URL will be cited in those languages |
| Google AI Overviews / Search | Search tries to match the searcher's language on multilingual sites. hreflang points to the right URL. Language detection uses visible content, not lang or hreflang (Google Search Central, Dec 2025). |
It does not publish how AI Overviews pick citations inside a language |
| Perplexity | Perplexity answers in the language of the thread and cites live web sources. I have not found a first-party public spec that defines language routing the way Google defines hreflang. |
I will not invent a Perplexity "hreflang booster." I test it |
Claude Opus 4.8, GPT-5.5, and Gemini 3.1 Pro are the models I am using in production work this month. They all handle multilingual prompts. That capability is not the same as "your English site will be the citation." Capability is the model's. Citability is the evidence.
The behavior I actually measure #
I keep the prompt meaning fixed and change only the language. Then I record:
- Whether the brand is named.
- Whether the site is cited.
- The language of the cited URLs.
- Whether the answer uses the local-language category word or falls back to the English brand.
Across ChatGPT search, Gemini, and Perplexity, the pattern that repeats is: when on-language sources exist, they dominate the citation list. When they do not, the engine either answers from English sources, answers generically, or names a competitor who has on-language pages. That is an observation from prompt panels, not a vendor SLA.
ChatGPT may also rewrite a prompt into partner queries (OpenAI, Aug 2026). If the rewritten query stays in the user's language — and I expect it often does — your English-only title never matches. Hedge: OpenAI has not published the rewrite language rule.
Location is not language #
ChatGPT can infer a country, state, or city from IP and use it for local results. Device location sharing is optional and off by default. That is a geography signal. This post is about language. A Spanish-language prompt from an English-language site visitor is still a Spanish retrieval problem. Do not hide a missing locale page behind "but we are a US business." Language is the unit.
English Ranking Is Not the Same as Being Citable in Another Language #
Ranking in English Google is a statement about English documents. Citability in a Spanish or Japanese answer is a statement about Spanish or Japanese documents. Those are different indexes, different citation lists, and different evidence piles. Winning one does not transfer.
Owners show me an English rank-tracking screenshot and ask why ChatGPT in French never names them. The screenshot answers a different question. English SEO can be healthy while multilingual AI visibility is empty.
Why English still dominates the crawl — and why that does not save you #
Common Crawl CC-MAIN-2026-30 (July 2026 crawl; CLD2 primary language on HTML) put English at 40.5782% of identified pages. The next large slices in that crawl were Russian 6.8217%, German 5.9862%, Japanese 5.3160%, French 4.8032%, Spanish 4.6418%, and Chinese 4.4292%. English is the biggest single language on the open web. It is not the only language buyers use in prompts.
Two mistakes follow from that table:
- "Most of the web is English, so English is enough." Forty percent is a plurality, not a monopoly. The other ~59% is not English. Training data and retrieval corpora still contain huge non-English mass. Your competitors who published one French page sit in that mass. You do not.
- "The model is English-native, so it will translate my page on the fly and cite me." Models can translate. Citation prefers a source that already matches the answer language. I have not seen a vendor doc that says they will cite your English URL and silently present it as a native-language source. When they do use English sources in a non-English answer, they often keep the citation English — or they skip you for a native-language publisher.
The transfer that does not happen #
| You already have | What it proves | What it does not prove |
|---|---|---|
| Page-one English Google for "best [service]" | English Search likes that URL | A French AI Overview will cite you |
| ChatGPT names you on an English prompt | English evidence is thick enough | The same prompt in German will name you |
| Strong English backlink profile | English entity authority | On-language third-party mentions |
| Perfect English FAQ schema | English extractability | A locale page exists |
GEO vs AEO vs AIO is the vocabulary post. The language cut is narrower: each of those disciplines still needs a document in the query language. You cannot GEO your way around a missing French page.
"But the model can read English" #
Yes. GPT-5.5, Claude Opus 4.8, and Gemini 3.1 Pro can read English and answer in Spanish. That is generation. Retrieval still has to pick sources. If the search partner query is Spanish, Spanish titles and snippets win the first cut. Your English title is in the wrong language for that cut. Hedge: the exact partner-query language is not documented by OpenAI. The safe build is still a locale page.
If your entire content operation is English, you are optimized for English AI search. That is a choice. It is not a default that covers every language your customers already speak.
A 30-Day Language-Visibility Audit I Run With Owners #
I pick one language you already receive, run the same buying prompts in English and in that language, then fix the four signals in order: locale URL, entity strings, reviews, hreflang. I do not open with a ten-language CMS project. I open with proof that you disappear, then one language that would change revenue if citations started landing.
This is the owner sequence. No invented client names. No fabricated before/after percentages. If you want the measurement habit after the pages exist, use how to track when AI tools cite or recommend your business.
Days 1–7 — Prove the disappearance #
- List the languages that already hit the business. Inbox, phone, chat, invoices, support tickets. Not languages you wish you had. Languages that already cost you a reply.
- Pick one. The one with the most buying questions, not the one that looks impressive on a slide.
- Write 10 prompts you already use in English. Translate them carefully. Do not run them through a cheap pass that changes the meaning.
- Run the panel in ChatGPT (search on), Gemini, Perplexity, and Google (AI Overview / AI Mode when it appears). Same week, same wording.
- Score each row: named / not named, cited / not cited, language of cited URLs, language of reviews mentioned.
A blank panel looks like this. Fill it. Do not invent the outcomes in advance.
| Prompt (keep meaning fixed) | EN named | EN cited URL lang | ES/FR/DE/JA named | ES/FR/DE/JA cited URL lang |
|---|---|---|---|---|
| "Who should I hire for [job]?" | ||||
| "Best [category] for [use case]" | ||||
| "[Brand] vs [competitor]" | ||||
| "Does [brand] offer [specific service]?" | ||||
| "What do customers say about [brand]?" |
If the English column names you and the second-language column does not, you have the bug this post is about. If neither column names you, stop and fix English AEO first. Language is not your first hole.
Allow OAI-Searchbot on the locale URLs. OpenAI says a site must allow that crawler to be eligible for ChatGPT search results (Searching the web with ChatGPT). A locale page blocked in robots.txt is an English-only site in practice.
Days 8–21 — Ship one locale, not a widget #
| Build item | Done means | Skip if |
|---|---|---|
| Locale URL for Home, Offer, FAQ | 200, indexable, no language auto-redirect | You only have a widget |
| Main content in the target language | H1, offer, FAQ, proof are in that language | Only the nav is translated |
| Entity name map on the page | Brand + local-language category in the first 100 words | English trademark only |
inLanguage on the CreativeWork / WebPage |
BCP 47 code matches the visible language | Schema on an English page claiming es |
Reciprocal hreflang to the English sibling |
Self-reference + return link, absolute URLs | Tags pointing at themselves |
| Review ask | Real customers who already use that language | Templated fake reviews |
I write the locale pages as answers, not as translations of English marketing. A translated slogan is not an FAQ. The questions on that page should be the questions from the prompt panel.
Days 22–30 — Re-run the panel and keep the log #
Run the same 10 prompts. Do not add new ones yet. You are measuring whether the new URL and the new strings entered the citation set. Citation tracking is a monthly habit after this. For a language launch, the first re-run is at day 30, then monthly.
What I record in the log:
- Date
- Engine
- Prompt language
- Named? Cited?
- Cited URL language
- Notes (invented competitor, generic answer, English-only citations)
If day 30 still shows English-only citations in the second language, I look at third-party mentions next — not another round of homepage polish. The locale page can be perfect and still lose to a directory that already describes you in that language, or to a competitor with reviews in that language.
What to Ship First If You Cannot Afford Ten Locales #
Ship one language you already serve, three money pages, reciprocal hreflang, and a review ask. Do not ship a machine-translated blog archive. Ten locales with thin copy will not get cited. One locale that answers buying questions can.
I am opinionated here. A "translate the whole site" project is how owners spend a quarter and still fail the prompt panel. Answer engines need extractable answers, not coverage.
The first-language cut #
Use this order. It is not a country list. It is a language list.
- Language of inbound buying questions. If 20% of qualified emails arrive in Spanish, Spanish is first. If nobody writes you in Korean, Korean is not first.
- Language of existing reviews and staff. You can support the page. You can ask for reviews without inventing a persona.
- Language where a competitor already gets cited in your prompt panel. That is the gap that is costing you names right now.
I ignore "we should be global" as a sequencing input. Global is not a language.
The three pages that pay for the rest #
| Page | Why it is first | What must be on-language |
|---|---|---|
| Offer / service | This is the buying question | What you do, who it is for, what it costs or how pricing works, FAQ |
| Proof | This is what synthesis quotes | Reviews, credentials, process. No English-only testimonials labeled as local |
| Home or hub | This is the entity landing pad | Name map, language switcher links (not an auto-redirect), hreflang + x-default if you use a selector |
Blog posts wait. Comparison pages wait unless the prompt panel is already "X vs Y" in that language. When you do write comparisons, the comparison-content rules still apply — in the target language.
Machine translation: when I allow it #
I allow a machine pass as a draft for an owner who will edit every sentence. I do not publish raw machine output on a money page. Language detectors and readers both notice it. Google wants the page language to be obvious from visible content. A muddy machine translation makes the primary language harder to assign and gives the model worse sentences to quote.
If you cannot edit the language, hire one editor for three pages. That is cheaper than a multilingual plugin that paints 200 URLs and creates 200 weak documents.
What I tell owners who only have English #
Keep the English site. Make it citation-grade. Then add one locale when a real language shows up in the pipeline. An English-only site is not a moral failure. It is a scope decision. It becomes a visibility failure the moment buyers ask in another language and you still have no document to cite.
The site I want to build for that moment is an AI-visibility-ready site: locale URLs, name maps, review language, and annotations that point at real pages — not a widget and a hope.
FAQ #
Does Google use hreflang to detect the language of my page? #
No. Google does not use hreflang or the HTML lang attribute to detect page language. It uses algorithms on the visible content. hreflang only points Search to the matching language or region URL. That distinction is in Google's localized versions documentation (updated 22 December 2025). If your Spanish URL is still English on the page, the annotation does not make it Spanish.
Will a Google Translate widget make my English site show up in Spanish AI answers? #
No. A widget does not create a crawlable Spanish document. Googlebot usually crawls from the US and does not send Accept-Language, so in-place swaps and widgets are easy to miss (Google, 10 Dec 2025). Answer engines cite URLs. Ship a locale page if you want a Spanish citation.
Do I need a separate URL for every language I care about? #
Yes if you want that language indexed and citable. Google recommends different URLs per language version instead of cookies or browser settings. Subdirectories (/es/) are the default I ship. Query parameters are not recommended. One English URL cannot honestly represent two languages to a crawler.
Does ChatGPT only search English pages when someone asks in Spanish? #
OpenAI has not published that rule. ChatGPT search ranks on multiple factors, may rewrite queries for partners, and does not guarantee placement (OpenAI Help, August 2026). In prompt tests I run, Spanish questions pull Spanish citations when those URLs exist. Treat language match as a working hypothesis, not a documented weight.
How many languages should I launch first? #
One — the language that already shows up in buying conversations. Three money pages beat ten thin locales. Add the next language only after the prompt panel shows citations in the first. "Global" is not a sequencing input.
Do machine-translated pages help or hurt AI citations? #
Raw machine output on a money page usually hurts. Google wants the page language to be obvious from visible content. Muddy translation makes language detection noisier and gives the model worse sentences to quote. I allow a machine draft only when a human edits every line before publish.
Should customer reviews stay in the language they were written in? #
Yes. Do not translate a review and present it as if the customer wrote the new language. On-language reviews are the quotable third-party voice for that prompt language. Ask customers who already write you in that language to review in that language. Leave English reviews on the English surfaces.
Does schema inLanguage replace hreflang? #
No. inLanguage labels the language of a CreativeWork. hreflang pairs URLs. Use IETF BCP 47 codes on inLanguage (schema.org/inLanguage, Schema.org V30.0, 19 March 2026). Add alternateName for English brand strings. Schema does not create a locale page and does not replace reciprocal hreflang.
Can I keep one English site and still get cited in other languages? #
Sometimes, if a third party already described you in that language. That is luck, not a system. First-party English pages are weak evidence for a non-English answer. If you want repeatable citations, you need a locale URL, a name map, and on-language reviews. English-only is a scope decision until buyers ask in another language.
How do I test whether my English-only site disappears in multilingual AI search? #
Run the same 10 buying prompts in English and in one other language you already receive, across ChatGPT search, Gemini, Perplexity, and Google AI Overviews. Score named / cited / language of cited URLs. If English names you and the other language does not, you have the disappearance this post describes. Keep the panel and re-run it 30 days after the first locale ships. Track citations monthly after that.
Get an AI-visibility-ready site built #
If your English site already ranks and still vanishes the moment the prompt language changes, the fix is not another English blog post. It is locale URLs, entity name maps, review language, and annotations that point at real pages.
I build AI-visibility-ready websites and run AI visibility audits for operators who need to get cited in the answers buyers already ask — including the languages those buyers already use. The work sits on the ai-visibility track: AIO / AEO structure, question-shaped pages, and a prompt panel you can keep running.
Book an AI visibility audit and bring two things: the English prompts that already name you, and the same questions in the other language you hear in the inbox. I will tell you whether you are missing a document, a name string, a review, or the site underneath.
Related Posts
How SaaS Brands Get Cited in the AI Answers Their Buyers Already Trust
SaaS brands get cited in buyer AI answers when review sites, comparison pages, implementation docs, and security pages agree on the same extractable facts.
Can Your Business Show Up When Someone Asks ChatGPT Out Loud?
Yes. ChatGPT Voice can name you out loud if the open web already treats you as a shortlist option. Spoken answers are shorter, so vague brands stay unnamed.
Making Your Brand Findable in Image, Video, and Voice AI Search
Your brand is findable in image, video, and voice AI search when pixels, captions, and spoken answers sit on pages a camera or microphone can retrieve.
