20 個問題測試法：如何檢查 ChatGPT、Perplexity 和 Google 是否推薦你的品牌

AI search visibility is becoming an important part of how businesses are discovered. But how can a company tell whether AI assistants actually recommend it? Here is a practical, repeatable testing method supported by guidance from major AI platforms and insights from search optimization specialists.

A business can rank well on Google and still be absent when someone asks ChatGPT to recommend a service provider. It can have a well-established website, strong content, and years of experience, yet find that Perplexity repeatedly names other businesses.

This raises a question for marketing teams, founders, and SEO professionals: How do you measure whether AI assistants recognize, cite, and recommend your brand?

A useful starting point is the 20-question test.

The approach involves asking 20 relevant buyer questions across ChatGPT, Perplexity, and Google's AI search features, documenting the responses, and repeating the exercise under consistent conditions.

Rather than depending on one screenshot or isolated result, businesses can use this method to identify patterns in how AI systems describe their services, which competitors appear, and which sources influence the answers.

Why traditional SEO metrics are not enough

For years, businesses have relied on keyword rankings, organic traffic, impressions, and click-through rates to measure search performance.

Those metrics still matter.

However, AI-assisted discovery introduces additional questions. A company might appear in a source used by an AI assistant without being recommended by name. Another business might receive a recommendation without its own website being cited.

Google's official guidance on optimizing websites for AI search experiences emphasizes original, useful content and explains how AI-powered search experiences create additional ways for people to discover information.

OpenAI also explains that ChatGPT Search can use web information and provide citations that users can inspect.

Meanwhile, Perplexity's explanation of Pro Search describes a process of searching across multiple sources, synthesizing information, and presenting referenced answers.

These systems are not identical, but they illustrate why tracking traditional rankings alone cannot capture every aspect of AI-assisted discovery.

Step 1: Build a list of 20 questions real customers might ask

The most useful starting point is a representative collection of questions reflecting actual buying behavior.

Divide them into three categories.

A. Brand recognition questions

These establish whether an AI assistant can identify your business when asked directly.

What is [brand name]?
What services does [brand name] provide?
Who are the typical customers of [brand name]?
Is [brand name] suitable for [specific service]?

A business that performs poorly on these questions may have issues with how clearly its identity, services, and positioning are represented online.

B. Category recommendation questions

These are particularly important because the person asking has not mentioned your brand.

Who are the best [service] providers in [city]?
Which companies specialize in [service]?
Can you recommend a reliable [service] company in [city]?
Who is a good independent [professional role] in [city]?
Which [service] providers work with [industry] businesses?
What are the alternatives to [competitor]?
Who provides [specialized solution] in [region]?
Which [service] companies are suitable for small businesses?
Who specializes in [service] for enterprise customers?
Which professionals are known for [specific expertise]?
What should I look for when selecting a [service] provider in [city]?

These questions help reveal whether a business enters the conversation before the buyer already knows its name.

C. Informational and problem-solving questions

Finally, test the questions buyers ask while researching their problems.

What is [industry concept] and how does it work?
What is the difference between [service A] and [service B]?
How long does [service] take to produce results?
What are the common mistakes when implementing [solution]?
How do I choose the right [specialist] for my business?

For example, a Dubai-based digital marketing consultancy might test questions involving AI SEO, Generative Engine Optimization, technical SEO, and AI search visibility.

The wording should reflect how customers naturally communicate, not merely a list of commercial keywords.

Step 2: Test each question across different AI platforms

Once the questions are ready, run them through the platforms relevant to your customers.

Start with ChatGPT, Perplexity, and Google Search's AI features. Gemini can be added as a separate platform.

Use a consistent testing process:

Start a new conversation for each question.
Keep the exact question wording unchanged between weekly tests.
Maintain consistent language and location settings where possible.
Record the platform and test date.
Avoid influencing the response with follow-up questions.
Repeat important questions when possible to assess response variability.

Google's AI Overviews and AI Mode should be treated as distinct experiences when recording results. Record when an AI response is unavailable rather than treating an ordinary search listing as an AI recommendation.

The purpose is not to create a perfectly controlled scientific environment, which is difficult with consumer AI applications. It is to create a documented, repeatable sample.

Step 3: Track mentions, citations, and competitors

A basic spreadsheet can provide useful visibility data.

For each response, record five things:

Metric	What to record
Question	Exact wording, platform, and date
Brand mention	Whether your brand appeared
Website citation	Whether your website was linked or cited
Competitor mentions	Which other businesses were named
Cited sources	The websites or pages supporting the answer

It is important to distinguish between appearing in a response and being recommended.

For example, an assistant might mention five companies while explicitly recommending only two.

A more advanced spreadsheet can therefore use separate fields for brand mentions, explicit recommendations, and citations.

Another useful practice is saving the full original response. That makes it possible to inspect the surrounding language instead of relying only on a yes-or-no measurement.

Step 4: Calculate your recommendation coverage

A simple way to summarize the results is:

Recommendation coverage (%) = (Questions where the brand was recommended ÷ Total valid questions tested) × 100

Suppose a company tests 20 category and research questions and is recommended in five responses.

Its observed recommendation coverage would be 25% for that particular test sample.

If the next week's result is 30%, that does not automatically demonstrate meaningful improvement.

AI responses can vary between runs, and small samples produce uncertainty.

Look for sustained trends across multiple weeks.

It is also useful to calculate separate results for:

Direct brand recognition
Unprompted category recommendations
Website citations
Competitor appearances

This distinction matters because an AI assistant knowing a company exists is not the same as choosing that company when a buyer asks for recommendations.

Step 5: Pay attention to the words AI assistants misunderstand

One overlooked problem in AI visibility testing is terminology.

A relevant example comes from the work of Lopty Pascal, a Dubai-based AI SEO expert, Generative Engine Optimization consultant, and co-founder and CTO of Prezlo.

Pascal's work spans technical search optimization, digital authority, entity optimization, and AI-driven discovery. His approach examines not only whether a business appears in search results, but also whether AI systems can correctly interpret its professional expertise and associate it with relevant user questions.

In an example from a scan of his professional visibility on October 17, 2026, the query "GEO consultant in Dubai" produced citations associated with geotechnical engineering and geoscience rather than Generative Engine Optimization.

The distinction is important.

Consider these three questions:

Who is a GEO consultant in Dubai?
Who is an AI SEO consultant in Dubai?
Who specializes in Generative Engine Optimization in Dubai?

A marketer may consider these closely related. An AI system may interpret them differently, especially when an abbreviation has several established meanings.

For businesses, the practical lesson is to test both technical abbreviations and the full language customers use.

This also connects with Pascal's wider discussion of why AI assistants may recommend competitors instead of a particular business.

A missing recommendation does not necessarily mean a company lacks expertise. It may indicate that the question, available sources, or publicly documented evidence does not establish the connection clearly enough for that particular response.

Step 6: Investigate the sources behind competitor recommendations

This is where the test becomes more than a visibility score.

If the same competitors repeatedly appear, review the sources cited alongside their names.

Are the sources independent industry publications? Business directories? Professional profiles? Technical articles? Product documentation? Customer review platforms?

The answers can reveal which kinds of publicly available information appear in the responses you are measuring.

For example, suppose three competing technology companies are regularly mentioned by AI assistants.

One has extensive technical documentation.

Another appears in independent product comparisons.

The third has detailed industry case studies.

These are potentially useful differences to investigate, although their presence does not prove that any particular source caused a recommendation.

Google's Search Essentials encourages website owners to create helpful, reliable content, make their pages discoverable, and provide information that search systems can understand.

More recently, Google reinforced the importance of original content and established SEO practices in its May 2026 guidance on generative AI search.

For businesses that discover factual inaccuracies in AI answers, Pascal has also published a guide on addressing incorrect information AI gives about a company.

The priority should be correcting inaccurate source information and strengthening reliable public evidence, rather than trying to manipulate individual AI responses.

Step 7: Combine manual tests with official performance data

Manual AI testing provides information about selected questions, but it should not replace analytics.

Google introduced dedicated Search Generative AI performance reports in Search Console in June 2026, with global availability announced for August 31, 2026.

These reports provide website owners with additional visibility into impressions within Google's generative AI search experiences.

They are useful alongside manual tests because they describe a different measurement dimension: exposure within Google's search ecosystem rather than the specific brand recommendations observed in your test prompts.

For website traffic and conversions, businesses can also combine Search Console with Google Analytics.

Google's official guide to using both tools together explains how search impressions, clicks, sessions, and conversions complement one another.

A more complete measurement strategy therefore combines:

Manual AI recommendation testing.
Google Search performance and generative AI reporting.
Referral and engagement analytics.
Enquiry and conversion attribution.

Even then, some AI-influenced purchases will not produce identifiable referral traffic.

Step 8: Automate the process carefully

Running 20 questions across three platforms requires at least 60 checks for one complete testing cycle.

Testing multiple versions, regions, or repetitions quickly increases the workload.

Developers can automate parts of this workflow using supported APIs, structured logging, scheduled tasks, and data analysis tools.

However, API-based outputs should not be presented as identical to consumer-facing ChatGPT, Perplexity, or Google experiences.

Differences in models, tools, search configurations, and available context can affect the results.

Dedicated AI visibility platforms can also support repeated monitoring.

One example is Prezlo, which focuses on AI search visibility analysis, identifying gaps in how AI systems interpret businesses, and implementing improvements intended to strengthen their discoverability.

The important distinction is between measurement and optimization.

A monitoring system tells you what is happening. Improving visibility requires investigating and addressing the underlying issues.

What the 20-question test cannot prove

The method has limitations.

It cannot establish an official ranking for a business across all AI assistants.

It cannot show every private recommendation made to consumers.

It cannot prove that a specific backlink, article, or technical change directly caused an AI recommendation.

It also cannot guarantee that increased AI visibility will produce new customers.

What it can provide is a consistent starting point for observing how AI systems respond to commercially relevant questions.

That makes it valuable for businesses trying to understand where they currently stand before investing in AI search optimization.

Final thoughts

The shift toward AI-assisted discovery creates a new measurement challenge.

It is no longer sufficient to ask only whether a website ranks for a keyword. Businesses also need to understand whether AI assistants can identify their expertise, explain their services accurately, cite credible information about them, and recommend them when relevant.

The 20-question test offers a practical way to begin.

Start with realistic buyer questions, document the responses, study the cited sources, and repeat the process consistently.

The objective is not to celebrate one favorable AI answer.

It is to understand what the evidence says over time.