CapstonAI is an AI visibility platform that helps brands track, audit, and optimize their presence in AI-generated answers. It monitors how your brand is mentioned across ChatGPT, Claude, Gemini, Perplexity, and Google AI Overviews, then provides actionable fixes to improve your share of voice in AI search results.

How does CapstonAI track AI visibility?

CapstonAI sends structured brand queries to multiple large language models (LLMs) on a recurring basis and records whether your brand is mentioned, cited, or recommended. It aggregates these results into a GEO score, a share-of-voice metric, and a citation report.

Which AI models does CapstonAI monitor?

CapstonAI currently monitors ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google), Perplexity, and Google AI Overviews. Coverage is continuously expanded as new AI search surfaces emerge.

CapstonAI offers a free WordPress plugin that provides basic AI visibility scanning for your site. For full AI visibility tracking, audit reports, automated content fixes, and multi-brand monitoring, the Core Plan is available at $99/month.

How is CapstonAI different from Semrush or Ahrefs?

Semrush and Ahrefs are built for traditional search engine optimization. CapstonAI is purpose-built for the AI search era: it tracks your brand presence inside AI-generated answers from ChatGPT, Claude, Gemini, and Perplexity.

Cross-Language GEO Stability: Why Claude Reuses Domains And GPT Swaps Ecosystems

Cross-Language GEO Stability: Why Claude Reuses Domains and GPT Swaps Ecosystems (Toronto Study)

The University of Toronto study by Chen et al. (arXiv:2509.08919, 2025) tested cross-language stability across English, Chinese, Japanese, German, French, and Spanish for the same intents. The result is one of the most engine-dependent findings in the entire dataset: Claude maintains relatively stable evidence sets across languages, GPT effectively swaps in different site ecosystems by language (cross-language overlap near zero), while Perplexity and Gemini sit between these poles. Below: the exact heatmap patterns, what they mean for multilingual brands, and the 7-step playbook to build cross-language GEO visibility.

TL;DR: Cross-language domain overlap is engine-dependent and generally low. Claude is the most stable (high reuse of authority domains across languages). GPT is the most unstable (near-zero cross-language overlap; entirely different site ecosystems by language). Perplexity and Gemini are intermediate. Practical implication: multilingual GEO requires authoritative local-language earned-media coverage plus English baseline; you cannot rely on English coverage spilling into other languages on most engines.

Free CapstonAI scan → GEO Research hub

What the heatmaps reveal

Chen et al. computed Jaccard overlap on cited-domain sets for the same query in English vs. each target language, across 10 verticals. The patterns by engine:

Google baseline. Cross-language domain overlap is generally low (0 to 0.1), with the maximum cell only slightly above 0.1 (EN-ES Electric Vehicles).
Claude. Much higher cross-language stability than Google in all verticals. Frequent high overlaps indicate strong reuse of the same authority domains across languages.
Perplexity. Mostly comparable to Google with a few pockets slightly above it (e.g., EN-DE laptops around 0.22).
Gemini. Modestly higher overlaps than Google in several language pairs (EN-CN exceeding 0.2 in pockets; EN-DE peaking at 0.32).
GPT. Shows lower overlap than Google. Near zero across the board. GPT effectively switches to different site ecosystems by language.

Website-language localization patterns

Under non-English prompts, citations tilt toward the target language, but to different degrees by engine. The Toronto pooled view shows:

Engine	Localization tendency	Implication
GPT	Most local-language heavy	Requires local-language earned coverage per target market
Perplexity	Local-language heavy	Requires local-language earned coverage per target market
Claude	English-heavy across all languages	English coverage compounds; local-language is secondary
Gemini	Balanced, varies by language	Mix of English + local; engine-specific tuning required

Across languages, the localization gradient is: Japanese (strongest localization, target-language often above 75% of citations), French and German (highly localized but keep a larger English slice), Spanish (mixed, English slightly exceeds target), Chinese (the exception — English dominates, over 75% of citations even with Chinese prompts).

The 7-step cross-language GEO playbook

Step 1: Map your target-market priority. Rank your top 3-5 non-English markets by revenue contribution. The Toronto findings imply that GEO investment per market needs separate budget, not a blanket “international” line.
Step 2: Build local-language earned-media coverage in each priority market. The Toronto data shows that GPT and Perplexity prompts in French, German, Japanese, or Spanish heavily cite target-language sources. Get coverage in tier-1 local publications (Le Monde for French, FAZ/Spiegel for German, Nikkei/Toyo Keizai for Japanese).
Step 3: For Chinese markets, prioritize English-language coverage too. The counter-intuitive finding: even Chinese-prompt outputs are dominated by English sources (over 75%). Investing only in Chinese-language PR for ChatGPT visibility in China underestimates the English-source weight.
Step 4: Maintain a stable English baseline for Claude. Claude reuses English authority domains across all language prompts. A strong English Wikipedia article and English authority-publication coverage will drive Claude citations across all language markets.
Step 5: For GPT, accept that each language is a separate market. GPT’s near-zero cross-language overlap means English authority does not transfer. Each target market needs its own earned-media campaign, its own local domain coverage, and its own tracking dashboard.
Step 6: Run cross-language prompt panels quarterly. Build a 30-prompt panel per priority language. Track citation rate per engine per language. The patterns shift quarterly; don’t extrapolate from English-only data.
Step 7: Localize structured data and schema, not just copy. Pages targeting French markets need French Schema.org content (name, description, FAQs) in addition to French body copy. AI engines parse the structured data layer first; bilingual schema markup compounds visibility.

What this means by language

Language	Localization strength	English share retained	Priority engines for local coverage
Japanese	Strongest (75%+ target-language)	Small	GPT, Perplexity
French	Strong (high target-language)	Moderate English slice	GPT, Perplexity
German	Strong (high target-language)	Moderate English slice	GPT, Perplexity
Spanish	Mixed (English slightly higher)	Significant English	GPT, Perplexity + English content
Chinese	Exceptional (English dominates 75%+)	Dominant English	English Wikipedia, Reuters, English authority

Common errors with cross-language GEO

One-size-fits-all international strategy. Treating French, German, Japanese, Chinese identically misses the per-engine, per-language patterns. Each combination needs separate planning.
English-only PR for global markets. Works for Claude (English-heavy across languages) and partly for Chinese markets. Fails for GPT in French, German, Japanese, Spanish where local-language sources dominate.
Local-language only for Chinese markets. Misses the dominant English-source bias even in Chinese prompts on most engines.
Translating English content without local Schema. Structured data layer matters. Bilingual Schema.org markup is a higher-leverage investment than copy translation alone.
Skipping engine-specific tracking per language. ChatGPT citation behavior in French is different from ChatGPT citation behavior in English. Run separate prompt panels per language.

FAQ — Cross-language GEO stability

Why does Claude maintain English dominance across languages?

Chen et al. interpret this as Claude prioritizing high-authority English sources (Wikipedia, Reuters, AP, Consumer Reports) over local-language alternatives even when the prompt is in another language. The mechanism may relate to training data composition or retrieval scoring that favors high-DR English domains. Practically, this means English Wikipedia notability and English authority coverage are high-leverage for global Claude visibility.

Why is Chinese the language where English dominates the most?

The Toronto study notes this as the exception across the language gradient. Possible explanations include the strength of English-language Chinese sources (Reuters Chinese desk, English-language Chinese tech publications), the underweighting of mainland Chinese-only sources, and engine training data biases. The practical implication: for Chinese-market GEO, English authority placements remain critical alongside any Chinese-language coverage.

How do we measure cross-language citation share?

Build a translated version of your top 30-prompt panel in each target language. Run it through ChatGPT, Perplexity, Claude, Gemini weekly. Track: (a) citation rate per language per engine, (b) target-language vs. English source share, (c) overlap between your English and local-language citation sets. The CapstonAI platform supports multilingual prompt panels with per-language reporting.

Tools and related reading

GEO Scientific Research 2026 (silo hub)
Earned Media Bias in AI Search
AI Engines Domain Diversity Compared
How to rank in Le Chat (Mistral) (French AI engine)
How to rank in DeepSeek (Asia-heavy engine)
Wikidata setup for brands
sameAs in Organization schema
CapstonAI AI Citation Tracking

Ready to build cross-language GEO?

Free CapstonAI scan →

Last updated: May 2026. Primary source: Chen, M., Wang, X., Chen, K., & Koudas, N. (2025). Generative Engine Optimization: How to Dominate AI Search. University of Toronto. arXiv:2509.08919. https://arxiv.org/abs/2509.08919