---
title: "Why your brand is missing from AI answers: start from the customer question"
lang: "en"
canonical: "https://winin.ai/en/learn/ai-brand-visibility/"
alternate: "https://winin.ai/zh/learn/ai-brand-visibility/"
datePublished: "2026-09-12"
dateModified: "2026-09-12"
section: "learn"
---

# Why your brand is missing from AI answers: start from the customer question

Buyers shopping for GEO usually do not start by asking "should we do GEO?". They start with a concrete question: when a customer asks AI, am I in the answer? And if not, at which layer does it break? Answer that first, because it decides what you are actually buying: more monitoring, or fixing facts and content first.

## 1. Write down the questions customers actually ask

Do not start with "how many models should we monitor". Start with "what did the customer actually ask". Collect 10-20 questions that sit closest to the moment before purchase, covering category recommendation, head-to-head comparison, usage scenarios, alternatives, and budget constraints.

Example: suppose you sell home espresso machines. A user might ask "What home espresso machines are actually worth buying?", or "I want a home espresso machine that lasts five years, budget around 5000, which one?" These are not the same prompt. The first asks for a category recommendation; the second carries budget, lifespan and usage context, so the candidate list, ranking and reasons may all differ. Record them as two separate questions, otherwise you cannot see which one you are missing from.

Once a prompt is in use for recording, do not casually rewrite it. Freeze wording, language and punctuation. If you need to change it, add a new prompt and keep the old one in the record. Numbers produced under a changed prompt are not comparable to the earlier numbers, whatever direction they move.

## 2. Four outcomes, four different actions

The same question can put a brand into four different outcomes. Their surface symptoms look similar ("AI does not show me" or "AI describes me wrongly"), but the work required is very different.

| Outcome | How you appear in the answer | Question to ask first | Priority action |
| --- | --- | --- | --- |
| Absent | You are not in the competitor list at all | Does the prompt contain a category term or a brand term? | Separate "not found" from "found but not recommended" |
| Wrong | Outdated price, discontinued status, incorrect specs | Which public page is this wrong claim coming from? | Turn editable pages into checkable, dated factual statements |
| Mentioned but not recommended | You are listed, but the reasons go to competitors | Which of your points is treated as a selling point, and which competitor point is treated as a buying reason? | Fill missing factual entries instead of repeating brand adjectives |
| Blind spot | You do not know what the answers look like | Who asked, when, and what was answered? | Fix prompts and model set, then start archiving |

A few notes.

**Absent** is the most commonly misdiagnosed. Brands often rank in search and assume AI will surface them too. But crawl and index status in a search engine does not directly represent whether a given model will recommend you; these are two separate measurements. In its AI features documentation, Google states that its AI search features build on SEO fundamentals, value helpful reliable content and discoverable internal links, do not require special AI text files, and do not guarantee crawling, indexing or display (https://developers.google.com/search/docs/appearance/ai-features). That statement describes Google's own mechanics only. It cannot be extrapolated to other models, and it certainly does not mean "indexed implies recommended".

**Wrong** is handled facts-first, wording second. Example: suppose an answer says one of your products is "discontinued", when in reality only the model name changed. The first step is not a long explanatory article. It is making the page for that model state the situation clearly, checkably and with a date, so that human readers and retrieval systems get the same version.

**Mentioned but not recommended** is the outcome most easily hidden by metrics. Rising mention counts feel like progress, but customers remember the reasons. If you are mentioned only because you are cheap, or because you are "the other option", that does little for being chosen.

**Blind spot** is not "no problem"; it is "no data". Without archives you cannot separate model drift, competitor moves and your own publishing.

## 3. Seeing the problem is only a necessary condition

Monitoring answers "was I mentioned". To turn risk into work a team can approve and execute, three things go with it: collapsing the claims that get cited into versions that have a source, an owner and a validity window; executing changes only with authorization; and retesting under the same conditions to confirm whether the change is actually connected to your action. These map to fact governance, authorized execution and retest discipline. Monitoring without them usually produces one of two results: reports get thicker while nothing that matters changes, or many things change and nobody can say which one worked.

## 4. Measure five numbers separately

The most common mistake is blending different layers into one "AI performance score". At minimum, record these five things separately:

1. **Mention**: does your brand name appear in the answer.
2. **Recommendation**: are you inside the candidate or recommendation list the answer gives.
3. **Citation**: does the answer cite your page or content as a source.
4. **Visit**: is there site traffic attributable to these answers.
5. **Opportunity**: did enquiries, trials or deals result.

Example: suppose a model's new answer mentions you (mention +1) but does not put you in the recommendation list (recommendation 0) and does not cite your page (citation 0). In that case, traffic and enquiries staying flat is normal, and you cannot use "mentions went up" to explain business results. The reverse also holds: a rise in traffic may come from a campaign or a media mention, not from AI answers. Correlation over time is not causation.

## 5. Model-layer differences must be written into the record

The same question can produce different results on a consumer product versus a direct API call: the consumer surface may include web retrieval, memory and personalisation, while an API call is often a fixed single-turn exchange; models with native retrieval return citations, those without it do not. So every record should state at least: which entry point was used, whether retrieval was on, and whether a logged-in or memory state applied. Mixing these together destroys comparability.

## 6. When this method breaks down

- If your category appears very rarely in AI answers (a very niche industrial product, for example), the sample will be small and any short-term conclusion is unreliable.
- If prompts keep being rewritten, the numbers are not comparable and discussing movement is meaningless.
- If the model version, retrieval switch or system prompt changed, the change you observe may be model drift rather than your action.
- If your brand has almost no checkable factual pages in public, the first step is filling those gaps, not increasing monitoring frequency.

## 7. A diagnostic template you can use directly

One row per record, with fields such as:

customer question | language | model | entry point and retrieval on/off | mentioned | in recommendation list | reason owner (self / competitor / none) | cited sources | record date | uncertainty notes

Do not skip the uncertainty column. It captures the part of the change you cannot yet explain, and it becomes the most useful material later when you do attribution.

## FAQ

**Is checking a few models manually each week enough?**

It can work as a starting point, with two caveats: manual records rarely keep prompts and conditions identical, and a single result is only a directional signal, not a formal baseline. Operational decisions need a fixed baseline and consistent recording.

**What should we do first?**

Freeze a prompt list and a model list, run one round of archiving, and see how you split across the four outcomes. This needs no tooling; a spreadsheet is enough to start.

Winin's direction here is to connect monitoring, fact governance, authorised execution and same-conditions retest into one checkable chain, so a brand knows not only whether it was mentioned, but what to change next and how to confirm it afterwards. Scope depends on your actual question set and model list.