Your brand is invisible in AI answers: what to do first

The first step is a baseline test, not new content. Every piece you write before you know why you’re invisible may land in the wrong place. Invisibility has three entirely different causes with three entirely different fixes — and telling them apart takes about half a day.

Three kinds of “invisible”

One: you were never retrieved. When the AI searches the web, your pages don’t make the candidate list, or you simply have no retrievable content. The symptom: ask about your brand by name and the AI says it doesn’t know, or can only repeat the few lines on your homepage.

Two: you were retrieved but not used. Your content is in the results, but the model doesn’t draw on it. The symptom: the answer quotes a competitor’s specific numbers and claims, while your content, though present, isn’t cited at all. This usually happens because there are no extractable facts in it — just words like “leading”, “professional” and “one-stop”, which give the model nothing it can repeat.

Three: you were displaced by a same-name entity. The AI mentions your name but describes another company. That’s a brand entity problem; more content won’t fix it and only makes it messier. The metric for it is name confusion rate.

The three fixes don’t transfer. The first needs crawlability and content that exists; the second needs a different way of writing; the third needs entity disambiguation. Skip the diagnosis and start writing, and whether you hit the target is down to luck.

Who this is for

Brands that find they don’t appear in AI answers, or appear incorrectly. Teams that know they need generative engine optimization but don’t know where to start. If you’re still unclear on how GEO differs from SEO, read GEO in practice first.

How to run the baseline

Step one: write the question set. Twenty to 30 questions, phrased the way customers actually ask — not your internal messaging. Include all three types:

• Branded: “What does Company X do?”, “How is X different from Y?”
• Category: “Which companies in Shanghai do A?”, “What’s a good solution for scenario B?”
• Scenario: the specific questions customers ask before deciding, usually with no brand name at all

Category and scenario questions matter most — being mentioned unprompted is worth far more than showing up only when asked by name.

Step two: run the platforms. Test at least three, each once with web search on and once with it off. Retrieval behavior differs a lot between platforms; see testing Doubao, Yuanbao, Kimi and DeepSeek.

Step three: record four things. For each answer, note whether you’re mentioned (mention rate), where you rank (first-recommendation rate), whether the description is right, and who the citation sources are. The fourth is the most valuable: it tells you directly whom the model trusts — and those sites are where you need to show up.

Symptom-to-fix table

Symptom	Likely cause	First move	Don’t
Asked by name, the AI says it doesn’t know	No retrievable content	Build a facts page on your site; get listed by third parties	Rush into ads
The AI describes a different company with the same name	Entity confusion	Standardize full and short names; add structured data	Add more content — it gets messier
Mentioned, but the description is outdated	Old sources outweigh new content	Update, and make the new facts appear consistently in several places	Fix only your own site
Cited sources are all competitors’ sites	You’re not in the pool of sources the model trusts	Produce content on those sites	Publish only on your own channels
Never appears in category questions	Your content only talks about you, not the category problem	Write content that solves the customer’s problem	Keep writing company profiles
Appears with web search, not without	The model’s long-term memory doesn’t include you	Accept it; work on the web-search side first	Expect a short-term change

Evidence

The projects for Kechuang Yihao and the Shanghai Industrial Transformation and Development Institute both started at this step: fix a test question set, run the baseline, then change content, then retest with the same questions. Kechuang Yihao’s 15-question snapshot on Tencent Yuanbao on 2026-06-15 showed a 100% mention rate and 80% first-recommendation rate; the institute’s 10-question Yuanbao snapshot on 2026-06-14 showed 90% / 70%. Both sets of raw answers and screenshots are archived internally.

Those numbers let you verify “what was answered that day”, but a single snapshot can’t independently prove incremental attribution — it has to be read alongside the baseline, the work in between and later retests. Both projects’ later two-platform numbers fluctuated, which is exactly why regular archiving matters more than keeping only the peak.

Yichi Garden started from a different place: it cared not about branded queries but about its position in category recommendations, and eventually reached third place in category recommendations in AI web-search results. Different goal, different question set — which is also why the question set has to be locked during diagnosis, not assembled after the fact.

The order after the baseline

Ranked by return on effort, the usual order is:

1. Fix the entity first. Name ambiguity, inconsistent short names, inconsistent business descriptions — leave these and all later content works for someone else.
2. Add structured data. Cheap, and it stops machines from having to guess who you are from body text.
3. Rewrite core pages. Use answer-first writing: give the full conclusion in the first sentence, and replace “industry-leading” with specific numbers and specific scope. Only content a model can lift whole has a chance of appearing in the answer.
4. Show up where the cited sources are. The cited sites in your baseline record are your list.
5. Add llms.txt. Very cheap; the effect depends on the platform. Don’t expect it to work on its own.
6. Retest every two weeks.

Limits and risks

• A half-day baseline is a small sample. Twenty to 30 questions set direction; they aren’t a precise metric. AI answers are inherently random — the same question asked twice can give different results — so ask important questions several times.
• Answers without web access won’t move in the short term. They depend on training data; you have to wait for model updates.
• Question sets can be self-fulfilling. Write questions around your own positioning and you’ll measure your own strengths. A version written by someone unfamiliar with the business is often closer to reality.
• Some industries can’t make it work. Extremely long-tail demand, customers who don’t use AI to decide, or a brand name that collides with a huge same-name entity all give poor returns. See how long GEO takes, and when it won’t work.
• Don’t treat an AI error as a PR crisis. Most wrong descriptions come from outdated public information; adding correct facts works much better than demanding removal. Hallucinations can’t be fixed by complaining either.

What to do today

Open three AI apps, ask 20 questions the way your customers would, and log the results in a sheet: mentioned or not, what position, accurate or not, who was cited. That sheet is your baseline and the basis for every later decision. If you’d rather have it done for you, start with a free GEO audit.

For a ready-made diagnostic process, see AI marketing consulting, which is accepted against mention rate and first-recommendation rate. If you want your team to run a round themselves, the AI marketing bootcamp has participants produce exactly this diagnostic report.