Why your own website is not the main lever
There are two ways an engine arrives at your name. It retrieves live web results at question time and answers out of them, or it answers from what it absorbed in training. Both are fed by the same thing: how many places across the web mention you.
Your own site controls neither. What it controls is whether you are eligible — crawlable, readable, quotable. That is necessary and it is not sufficient. A site can be technically flawless and still never be named, which is the ordinary case for a brand with no third-party footprint.
This is the single most common mistake in published GEO checklists: they spend most of their length on on-site tweaks because those are the ones you can do alone, and skip the off-site work that actually decides the answer.
What we measured, and what it changed about the advice
In July 2026 we asked eight engines the same twelve B2B buying questions, once in English and once in Chinese — 384 calls, 356 valid answers. The result reordered our own priorities.
The language of the question, not the nationality of the engine, decided who got recommended. Both groups of engines named Chinese products in 1.3% of recommendations when asked in English. Asked in Chinese, the same engines jumped to 26.0% and 31.1%.
The practical reading: visibility is earned per language, not per engine. Adding engines to your monitoring shows you the same brands again; adding a language changes a quarter of the list.
| Chinese-brand share of recommendations | Asked in English | Asked in Chinese |
|---|---|---|
| Western engines (ChatGPT · Claude · Perplexity · Grok) | 1.3% | 26.0% |
| Chinese engines (豆包 · 通义千问 · DeepSeek · Kimi) | 1.3% | 31.1% |
The Split Answer Report, measured July 2026. Single-sample — directional, not precise.
What to do, in order of how much it moves the answer
Ranked by leverage rather than by ease, because the easy items are the ones every checklist over-invests in. Evidence strength is stated so you can argue with the order.
| Lever | Why it ranks here | Evidence |
|---|---|---|
| Get mentioned on third-party sites | Feeds both retrieval and memory. Dominant by a wide margin. | Strong |
| Publish original data and statistics | Stat-rich pages get cited markedly more, and the data doubles as the reason anyone links to you. | Strong |
| Answer-first structure, question-shaped headings | Makes you quotable once retrieved. Engines fan a question into sub-queries and match each separately. | Strong |
| Schema.org structured data | Hygiene. Helps machines parse you; limited direct evidence it lifts citation. | Medium |
| Entity records and consistent profiles | Resolves who you are across sources. | Medium |
| Let the AI crawlers in | Necessary, not sufficient. Verify on the live site — a CDN can override robots.txt. | Baseline |
| llms.txt | Cheap, harmless, no engine has confirmed using it. Ship it; do not believe it. | Weak |
Ordering follows the leverage hierarchy we operate by, published in full at /method.
How do I know whether it worked?
Ask the engines the questions your buyers ask, with your brand name absent from the prompt, and count how often you are named. Any prompt containing your name measures whether the engine can read your site — which a site published an hour ago also passes.
That distinction is the whole measurement. We keep it separate on every report we produce, and it is worth being strict about on your own numbers too.
Common questions
Isn't this just SEO with a new name?
The retrieval half overlaps heavily with SEO — be findable, be credible, be quotable. What differs is the unit of success. SEO wins a position on a list of links; this wins a sentence inside an answer, where there is no second page and usually only three names.
How long does it take?
On-site changes are typically crawled within days. Being recommended accumulates over weeks to months, because it depends on third-party pages appearing and being re-crawled. Anyone quoting a shorter timeline is describing indexing, not recommendation.
Can anyone guarantee that an AI will recommend my brand?
No. AI answers are non-deterministic and personalised — the same question returns different answers to different people on different days. The honest claim is raising probability, and a vendor promising guaranteed placement is telling you something they cannot know.
Does llms.txt work?
No major engine has confirmed using it for ranking or retrieval, and Google has said plainly that no new machine-readable file is needed to appear in Search. It costs nothing and is a clean brand summary, so we ship it — but it is the last item on the list, not the first.
Scope and limits
The figures on this page come from a single-sample run of eight engines in July 2026, on twelve B2B software questions, through official APIs rather than consumer apps. They are directional. We publish the dataset and the method so you can disagree with them specifically: /research/split-answer, CC BY 4.0.
Last updated 2026-09-26.
Is your brand in the answer?
Avowd measures how eleven AI engines describe and recommend a brand, in English and Chinese, then does the work to change it. The free check takes about thirty seconds.