AI Visibility: Why Assistants Don't Cite Your Brand, and How to Measure It
- Jul 2
- 3 min read
A growing share of your future customers no longer type their question into Google. They ask an assistant — ChatGPT, Perplexity, Gemini, or, for Chinese consumers, Doubao and DeepSeek.
The answer they get doesn't contain ten blue links. It contains a recommendation: two or three names, sometimes a cited source. Either your brand is in that answer, or it doesn't exist for that customer.
We audit European brands' visibility in AI assistants — on the Western side (ChatGPT, Perplexity, Google's AI Overviews) and on the Chinese side (Doubao, Kimi, DeepSeek). The finding fits in one sentence: almost everyone still optimizes for an engine that ranks, while discovery is shifting to engines that cite. That is what GEO — Generative Engine Optimization — is about.
Key takeaways
SEO optimizes for an engine that ranks; GEO optimizes for engines that cite. They stack — neither replaces the other.
A brand appears in an AI answer only if it climbs three steps: being readable (AI crawlers can access the site), being citable (content breaks into answers), being cited (the sources the model reads talk about you).
AI visibility is measurable: share of voice on a real-prompt matrix, assistant by assistant, against named competitors. No baseline, no provable progress.
The engine that ranks and the engines that cite
Google ranks ten links; being fourth is still a business. An assistant synthesizes two or three sources into a single answer; not being in it means being invisible — and the user often doesn't click at all anymore. This shift doesn't kill SEO: AI Overviews sit on top of Google results, and models readily read pages that already rank. That's why we apply a hard rule on every content project: each page must cover both tracks at once, SEO (ranking) and GEO (citation). Doing only one means half of discovery passes you by.
Three steps: readable, citable, cited
Step 1 — be readable. Before any strategy, a trivial question: can AI crawlers read your site? A robots.txt blocking GPTBot, ClaudeBot or PerplexityBot, or content rendered only in JavaScript, and the conversation ends there. Add an llms.txt file — the emerging standard that hands models a map of your site. It's the cheapest audit and the fastest fix.
Step 2 — be citable. Models don't cite slogans; they cite passages that answer a question. Citable content makes direct claims with numbers and dates, breaks into question-answer pairs (FAQ), and carries structured data (schema.org: Organization, Article, FAQPage). A beautiful but vague page will never be cut into an answer.
Step 3 — be cited. The step everyone forgets: the model doesn't recommend what your site says, it recommends what its corpus says. On the Western side: Wikipedia and Wikidata, press, comparison sites, forums. On the Chinese side, eight sources dominate — Zhihu, Xiaohongshu, Baidu Baike, Ctrip, Dianping, Weibo, Mafengwo, WeChat Official Accounts. Absent from the corpus means never recommended, however good your site is.
Measuring: share of voice inside the answers
AI visibility is not an impression, it's a metric. Our method: a matrix of 30 to 50 real prompts in your category ("best cashback app in France," "where to buy luxury in Paris"), run at a regular interval on each assistant. We log: your presence in the answer, your rank, the sentiment, the cited sources. Baseline against named competitors, produce content on the sources that matter, then re-measure — the same quarterly audit-produce-measure loop our dedicated practice runs on Chinese assistants (Doubao, Kimi, Qwen, Wenxiaoyan, DeepSeek).
Urgency varies by market. On the China side, Doubao counts 226 million monthly active users (+126% year over year) and over 60% of Chinese outbound travelers now plan with an AI assistant — against close to zero European retailers optimized for it. Positions are being claimed now; whoever enters the corpus first gets cited for quarters.
Where a European brand should start
Week 1: crawler-access audit + share-of-voice baseline. Quick fixes: robots.txt, llms.txt, schema. Then produce where the models read — not only on your own site. Re-measure monthly. The entry budget is modest; the effect compounds, because a corpus is built, not rented.
FAQ
Is SEO dead? No. AI Overviews live on top of Google and models read well-ranked pages. GEO adds to SEO; it doesn't replace it.
How long before a measurable effect? Technical fixes (crawlers, schema) show within weeks. Entity presence in the corpus takes one to two quarters of steady production. Hence monthly measurement.
What about Chinese assistants? Different corpus, different sources, same three-step logic. It's a practice of its own — ours covers five Chinese assistants and the eight sources their models read.
ECTIME AI Lab is the applied-AI research and deployment unit of ECTIME Group. We build, ship and stress-test agentic systems in production, from GEO/SEO automation to multi-step autonomous agents. We maintain open-source Claude Skills for GEO/SEO and advise European brands on deploying AI that is not just autonomous, but verifiable and authorized.
Comments