Why Is Your Business Not Showing Up in AI Answers?

Why Businesses Are Not Showing Up in AI Answers

Paul Rowe — Founder, Chief Generative Engine Optimisation Officer & CEO, NeuralAdX Ltd:

Hello, Phil Haoyang Pang, and thank you very much for joining us today on the NeuralAdX GEO & AI Search Strategy Podcast.

You deserve a substantial introduction. You are the founder of Canlah.ai, a company specialising in generative engine optimisation, helping brands get discovered, cited and recommended by AI search platforms such as ChatGPT and Google AI Overviews, with the aim of turning attention into measurable revenue.

Your client experience includes Bedou AI Cloud, Google’s cross-border accelerator, ByteDance’s global expansion team and SGX-listed companies. You are also involved in one of the world’s most active AI agent communities, with more than 1,000 AI skills published and over 300,000 cumulative downloads. Your AI agent research experience spans NTU, UC Berkeley and Stanford.

So we are speaking with someone who has substantial experience across AI, agents and marketing.

The topic today is one that I think will resonate with a lot of business owners and leadership teams: why is your business not showing up in AI answers?

If a client asked you that question today, how would you break the problem down stage by stage?

GEO as a Layered Visibility Strategy

Phil Haoyang Pang — Founder, Canlah.ai:

It is a complicated question because it covers a lot of different layers.

From our perspective, GEO covers much more than a company’s own website. It includes the authority websites that may mention or cite the company, the sources the company itself cites, public relations, social media and the wider information environment surrounding the brand.

We normally begin with monitoring. We look at what the customer is currently experiencing and identify the specific visibility problems. From there, we can understand the different steps that need to be addressed and build a structured plan around them.

Crawlability, Retrieval and Semantic Relevance

Paul Rowe — Founder, Chief Generative Engine Optimisation Officer & CEO, NeuralAdX Ltd:

That aligns closely with my own methodology. The word you used that stands out is "layers". I see generative engine optimisation as a layered, step-by-step process.

The first distinction is between traditional search visibility and AI visibility. In conventional browser search, businesses have historically focused on the SERP, or search engine results page. In an AI platform environment, you are dealing with additional retrieval, synthesis, citation and recommendation behaviour.

So a business cannot simply assume that because it performs well in traditional Google search, it will automatically be selected, cited or recommended inside an AI-generated answer. The optimisation strategy has to evolve for the AI search environment.

The first fundamental layer is crawlability. I think of that as making sure the gates are open. You could spend an enormous amount of time improving a website, but if the relevant crawlers cannot access the content, the work cannot have the intended effect. If access for systems such as Googlebot or relevant AI crawlers is blocked, important information may never become available for retrieval.

After crawlability, I look at content retrievability. Is the content constructed in a way that an AI system can identify and use an individual passage? For example, does a section clearly present the question being addressed and then provide a concise, direct and evidence-supported answer?

That matters because AI systems can retrieve and evaluate passages rather than treating an entire page as one undifferentiated block of text. A useful passage should make it easy for the system to understand what claim is being made and what evidence supports it.

Then there is search intent and semantic relevance. The page needs to align closely with the prompt or question for which the business wants to surface. The content should answer that intent thoroughly without drifting into unrelated topics.

The next layer is query fan-out. An AI system may take one user prompt and expand or decompose it into multiple related searches and semantic variants. That means the content needs what I call generative answer coverage: it should address the important sub-questions, related concepts and likely variations surrounding the main intent.

After that comes citation worthiness. This is where the original Princeton GEO research becomes useful. In my own implementation, I place particular emphasis on citations, statistics and quotations because those evidence-oriented methods were among the strongest-performing approaches in the original study.

The aim is not to insert them mechanically. The point is to make important claims easier to verify and more useful to an AI system constructing an answer.

Then we have authority. That includes independent research, evidence, methodology, credentials and clear author information. From a GEO standpoint, author biographies are important because they help establish who is responsible for the content and why that person’s expertise is relevant.

From there, I move into third-party mentions and external validation. AI systems can encounter a brand across relevant industry websites, authoritative directories, independent publications and other sources. That wider footprint helps establish that the entity exists beyond its own website.

Then there is entity clarity and schema. A website should make it unambiguous who the company is, who its people are, what services it provides and how those entities relate to each other. I think about entity disambiguation and entity salience on the visible page, with appropriate structured data reinforcing those relationships for machines.

Source diversity is another layer. You do not want every important point to depend on one article or one type of source. Drawing appropriately from official organisations, academic research, institutions and credible industry expertise gives the content a stronger evidential foundation.

Finally, there is recency. Where the subject changes quickly, recent evidence matters. If a current study, official publication or dataset exists, it can be preferable to relying only on something that is already outdated.

So the layered approach is broadly: crawlability, retrievability, search intent, semantic relevance, generative answer coverage, citation worthiness, authority, third-party validation, entity clarity, structured data, source diversity and recency.

There are more layers beyond those, but that is the core framework I work from.

Phil Haoyang Pang — Founder, Canlah.ai:

That is a very detailed way to look at optimising a single website.

GEO Beyond the Website: PR, Social Media and Third-Party Authority

Phil Haoyang Pang — Founder, Canlah.ai:

We also see GEO as a well-rounded approach across different marketing channels.

Website optimisation is absolutely important, but when we measure different industries, we sometimes find that AI systems do not cite the company’s own website very often. They may instead cite government sources, industry experts, ranking lists, newspapers, publications or other third-party sources.

That is why I see GEO as a holistic marketing discipline rather than just a website activity.

People working across public relations, social media, video, content and other channels can all contribute if they understand what the overall objective is.

For example, the way information is structured in YouTube subtitles or transcripts may matter because search and AI systems can process that information. LinkedIn articles, public relations coverage and social content can also contribute to the wider information environment around a brand.

So GEO can bring together website content, PR, social media, video and third-party authority into one broader strategy for helping AI systems understand and retrieve the brand.

Paul Rowe — Founder, Chief Generative Engine Optimisation Officer & CEO, NeuralAdX Ltd:

That is one of the things I find most compelling about the field.

If you step back and take an orbital view, the whole discipline points toward creating information that is genuinely useful, trustworthy, accurate and well evidenced.

Traditional search has had periods where simplistic tactics could be overused, such as putting excessive emphasis on keywords rather than the actual usefulness of the content. GEO, at its best, should push businesses toward providing information that answers real questions, supports important claims and makes expertise easier to verify.

I like that because the direction of travel is toward better information.

That does not mean that being honest or writing high-quality content automatically guarantees AI visibility. There are still technical, retrieval and authority factors involved. But if your objective is to provide the clearest, most accurate and most useful answer, you are moving in the same direction as the systems you want to be selected by.

For me, that is a positive by-product of AI search. Businesses have more incentive to produce genuinely useful material rather than content created purely to manipulate a ranking signal.

Case Study: Improving AI Visibility Through Authority Signals

Paul Rowe — Founder, Chief Generative Engine Optimisation Officer & CEO, NeuralAdX Ltd:

Can you give us an example or case study of a client who came to you with the problem, "Why is my business not showing up in AI answers?"

What did you test, what did you change and how quickly did you see movement in their AI visibility?

Phil Haoyang Pang — Founder, Canlah.ai:

Yes. One example was a customer in the money-lending industry, which is a highly regulated sector.

The company had been established for many years. It had very good Google reviews, reasonable traditional search performance and the necessary credentials, but it was barely appearing in ChatGPT recommendations.

During our assessment, we looked closely at authority and certification. In a regulated industry, being properly licensed or certified is fundamental because a company cannot legally provide those services without the necessary approvals.

The client had the credentials, but the problem was that the evidence was buried deep in the website. Important badges, certifications and supporting information were difficult for a visitor or retrieval system to encounter quickly.

We moved those authority signals and relevant certification badges much closer to the primary landing experience so that the evidence became much more prominent.

That relatively simple technical and content change made a significant difference in our monitoring. Initially, the brand was being seen in only around 10% of the tracked situations. That moved to roughly 40%, and with further optimisation it continued to improve. Over time, it was appearing among the leading recommendations for a much larger percentage of the prompts we were measuring.

The important point is that the company did not suddenly acquire new authority. The authority already existed. The problem was that the evidence was not being surfaced clearly enough.

Paul Rowe — Founder, Chief Generative Engine Optimisation Officer & CEO, NeuralAdX Ltd:

Exactly. It sounds simple once the problem has been identified, but the diagnosis is the important part.

A website can contain excellent information, but if the most important evidence is buried too deeply, a retrieval system may not encounter or prioritise it in the way you expect.

That makes information hierarchy very important.

I experienced something similar with my own NeuralAdX benchmark and proof pages. Initially, I structured some of those pages in a more academic order. I had the methodology and explanatory material first, with the actual benchmark results further down the page.

The pages contained a lot of knowledge and evidence, but I felt they should have been performing better in AI retrieval.

So I tested a different structure. I moved the key benchmark results and the table of metrics much higher up the page, with only a short introductory section before them. The methodology was still there, but the strongest evidence was encountered immediately.

After making that change, I observed stronger performance for relevant queries across multiple AI platforms.

I have documented this through recurring live retrieval testing and screen recordings over time. That is important because I do not want to rely on one isolated answer. I want to see whether the behaviour repeats across platforms and over a longer period.

So your example reinforces something I have found as well: do not make an AI system, or a human reader, work unnecessarily hard to find the most important evidence, result or authority signal on a page.

Sometimes the difference between weak and strong AI visibility is not that the information is missing. It is that the information is poorly positioned, poorly structured or insufficiently explicit.

Measuring GEO ROI and Commercial Impact

Paul Rowe — Founder, Chief Generative Engine Optimisation Officer & CEO, NeuralAdX Ltd:

I think we have covered that question well. Do you have anything you would like to ask me?

Phil Haoyang Pang — Founder, Canlah.ai:

Yes. I have one serious question that I have discussed with partners and other experts in Southeast Asia and China.

A lot of clients care about commercial numbers. They care about cost, return on investment and what measurable value they receive from a marketing activity.

Many of our customers are SMEs or smaller companies, so every amount they spend matters. They may say, "If I spend $500, $1,000 or $5,000 on GEO, what exactly will you give back to me?"

How do you answer that question?

Paul Rowe — Founder, Chief Generative Engine Optimisation Officer & CEO, NeuralAdX Ltd:

I completely understand that because I face the same question.

Part of the answer is education. There is no honest way to say that a specific GEO investment will always produce an exact amount of revenue.

Generative engine optimisation can improve the probability of a brand being discovered, mentioned, cited or recommended in AI answers, but there are still several stages between visibility and a completed sale.

I can help put a business in front of the customer through a brand mention, domain citation or recommendation. I can influence how clearly the company is represented and potentially improve the sentiment and context around the recommendation. But I cannot control the final human decision.

A person may see the brand in an AI answer and not click. They may compare several companies. They may choose a competitor for reasons that have nothing to do with the quality of the optimisation. Human behaviour is variable.

So I would never tell a client, "Give me £1,000 and I guarantee you a particular amount of revenue in return." That would not reflect the reality of a volatile and probabilistic environment.

What we can measure are the indicators that sit much closer to GEO itself: AI answer visibility, brand mentions, domain citations, recommendation frequency, relative position, coverage across tracked prompts, citation counts, share of voice and changes in those metrics over time.

Those measurements do not replace revenue attribution, but they show whether the brand is becoming more visible inside AI-mediated discovery.

There is also an important difference between traditional search and AI answers. With the traditional ten-blue-links model, a user may have had many businesses available across a results page and subsequent pages. In some AI answers, the system may recommend only a small number of companies.

That makes inclusion itself strategically valuable.

If AI-assisted discovery continues growing, a company that is consistently absent from those answers risks being absent from the customer’s consideration set altogether.

That is why I think the commercial case for GEO is broader than simply saying, "We moved you from number five to number three." The first question is whether the business is present, accurately represented and competitively positioned when customers ask AI systems for recommendations, comparisons or solutions.

Then you can measure how that visibility develops over time and connect it, where possible, to website traffic, enquiries, pipeline and revenue.

As the market matures, I think businesses will become more familiar with those measurements. They will understand that appearing in AI answers is itself a meaningful stage in the customer acquisition process.

I also think the value of proven GEO capability will increase because large companies with substantial resources will invest heavily in this area. Smaller companies will still be able to compete, but they will need efficient, evidence-led strategies rather than relying on guesswork.

That is why our job is to keep testing, documenting and improving the process.

Phil Haoyang Pang — Founder, Canlah.ai:

I agree.

There are customers who care about branding and whether they are being shown by AI systems, and there are customers who care almost entirely about direct numbers.

Sometimes those clients need education. In other cases, a particular industry may decide that another marketing channel deserves more budget in the short term.

Over time, I think the market will reach an equilibrium and companies will develop a better understanding of where GEO contributes commercially.

Paul Rowe — Founder, Chief Generative Engine Optimisation Officer & CEO, NeuralAdX Ltd:

I agree, and I think this specialist field is going to become extremely important.

The hard work now is building genuine capability: studying the systems, testing them, documenting results and learning what actually changes AI visibility.

If someone becomes very good at that and can demonstrate the results, the market will place significant value on that skill set.

That is what makes this period exciting. We are still in a stage where there is a lot to learn, which means the practitioners doing the detailed research now are building knowledge that could become very valuable later.

I find that genuinely motivating. I can spend late nights researching this because I am interested in what the systems are doing and how the environment is changing.

AI Agents, Marketing and the Future of AI Discovery

Phil Haoyang Pang — Founder, Canlah.ai:

My background is in AI and machine learning.

I studied AI, carried out research in AI and worked with machine learning. I also started working with AI agents around 2023, so I became very interested in how agents could move closer to everyday life and how people could use them in practical workflows.

At the same time, I had experience in marketing and advertising environments. Eventually I realised something important: no matter how technically strong you are, if you do not market yourself, people may never discover or adopt what you have built.

That brought the technical and marketing sides together for me.

Generative engine optimisation is interesting because it is one of the more technical and systematic areas of marketing. It sits naturally with my background.

I also use AI agents within GEO work for customers. My thinking is that AI systems can be used to analyse other AI systems at scale.

If you understand the way these systems process information, retrieve sources and form answers, you can structure a client’s information so that it is clearer and more useful to those systems.

Paul Rowe — Founder, Chief Generative Engine Optimisation Officer & CEO, NeuralAdX Ltd:

Yes, essentially understanding the language and behaviour of the systems themselves.

Phil Haoyang Pang — Founder, Canlah.ai:

Exactly. It raises an interesting question about how information influences an AI system’s perception and how a brand can become more prominent within that information environment.

Paul Rowe — Founder, Chief Generative Engine Optimisation Officer & CEO, NeuralAdX Ltd:

My view is that GEO is fundamentally about understanding the directives and signals already present in AI retrieval and recommendation systems, and then making sure the information we publish aligns with those requirements.

I do not see it as controlling an AI system. I see it as making the right signals available so that the system can understand, retrieve and trust the information.

That perspective comes from studying the Princeton University GEO research and the subsequent research that has followed it, then testing those principles in live AI environments.

Different AI platforms can behave differently. ChatGPT may not respond to the same signals in exactly the same way as Gemini, Google AI Overviews or another platform.

That is where platform-specific testing becomes important.

You can run repeated tests and observe whether a particular platform appears more responsive to citations, quotations, recency, technical terminology, source diversity, fluency or other characteristics.

The important thing is to move away from assumption and toward measurement.

Research-Backed GEO and the Princeton Study

Phil Haoyang Pang — Founder, Canlah.ai:

That is something I have been studying over the last few weeks as well.

I have been going through the academic papers you mentioned, including the Princeton GEO paper.

You are one of the first people I have spoken to in this field who consistently refers back to that research, even though it should be an obvious starting point.

A lot of people now offer GEO services, but many do not seem to discuss or study the academic foundations in much depth.

Paul Rowe — Founder, Chief Generative Engine Optimisation Officer & CEO, NeuralAdX Ltd:

That has been one of the biggest lessons for me.

If you want to understand what generative engine optimisation is, one logical starting point is the research that introduced and evaluated the concept.

The original GEO study tested multiple optimisation methods and measured their effects. That gives you something concrete to investigate rather than relying entirely on marketing terminology.

From there, you keep learning. Read the research, test the ideas, document what happens and compare those results with live AI retrieval behaviour.

That is how I have approached NeuralAdX.

I have conducted recurring benchmarking and live AI retrieval testing over an extended period. I record assessments, document the implementations and then look at whether the changes appear to improve visibility, citations or recommendations.

The important thing is the evidence trail. I want to know what changed, what happened afterwards and whether the result repeats over time.

I do not think one academic paper permanently defines the field. AI systems continue to change. But the research gives you a disciplined foundation for experimentation.

That is very different from simply using GEO terminology because it is fashionable.

Phil Haoyang Pang — Founder, Canlah.ai:

There is another company doing similar research and testing in Singapore: Canlah.ai.

Paul Rowe — Founder, Chief Generative Engine Optimisation Officer & CEO, NeuralAdX Ltd:

Yes, of course. As soon as you said it, the penny dropped.

Multi-Platform AI Testing: Browser, API and Source Overlap

Paul Rowe — Founder, Chief Generative Engine Optimisation Officer & CEO, NeuralAdX Ltd:

What are you doing in terms of benchmarking and testing?

What infrastructure are you using to keep improving your knowledge so that what you learn can influence the next optimisation cycle?

Phil Haoyang Pang — Founder, Canlah.ai:

We have been doing a lot of research across different AI systems and different versions of those systems.

For example, we compare ChatGPT in the browser with API-based behaviour. We look at Gemini in the browser and through APIs, Google AI Overviews and other AI search environments.

We want to understand whether the browser and API versions source and rank information differently, what their preferences appear to be and which signals overlap.

We also investigate the underlying search and retrieval systems that different AI products may use. We are interested in the similarities and differences between those retrieval results and Google’s results.

Another important area is source overlap. We want to know how many sources are being used across different AI systems, which sources appear repeatedly and which types of sources are cited most often by a particular platform.

For client assessments, we can run more than 40 prompts and analyse roughly 200 to 500 different sources. Within those sources, we look at which ones are cited most frequently and by which types of AI system.

From that, we can build a strategy for the specific client, product and stage of the market.

We also test browser and API environments because future AI agents may increasingly rely on API-based models rather than only the consumer interfaces people use today.

If a customer’s brand is visible in one environment but absent in another, that difference may become commercially relevant as agent-based discovery grows.

We have substantial infrastructure in Singapore that allows us to run this kind of testing at scale. The objective is to understand not just one answer, but patterns across systems, prompts, interfaces and sources.

Paul Rowe — Founder, Chief Generative Engine Optimisation Officer & CEO, NeuralAdX Ltd:

That is the challenge: deciding the scope of what you want to measure, because the possibilities are almost limitless.

There are so many experiments you could run.

I have my own research interests, but I try to bring myself back to the commercial question. I ask: what information would a client genuinely find valuable enough to pay for?

That helps me separate intellectual curiosity from commercial usefulness.

I might personally want to investigate dozens of different behaviours inside an AI system, but for a client I need to identify which measurements actually affect their business decisions.

So I try to work within that framework: rigorous research, but directed toward commercially meaningful outcomes.

That is also why it is useful speaking with someone else who works deeply in this field. There are still relatively few people who spend their day thinking about retrieval behaviour, source selection, AI visibility and how these systems may evolve.

AI Answer Variability and Repeated Testing

Phil Haoyang Pang — Founder, Canlah.ai:

There is another major problem with testing: you can ask the same question repeatedly and still get different answers.

In our experiments, the broad structure of an answer may remain similar, but the selected sources, details and reasoning can continue to vary across repeated runs.

We have tested prompts three times, eight times and even more. The answers can remain broadly similar while certain elements still change after many repetitions.

That creates an important commercial question.

How many times do you need to repeat a prompt before you have enough evidence to make a useful conclusion for a customer?

At some point, extra testing may increase confidence, but it also increases cost and complexity. So you have to decide what level of repetition is statistically useful and commercially practical.

Paul Rowe — Founder, Chief Generative Engine Optimisation Officer & CEO, NeuralAdX Ltd:

Exactly. Variability is inherent in these systems, which is why I prefer to think in terms of repeated testing and patterns rather than treating one answer as definitive.

Source diversity also plays into this.

One approach that has performed well in my own testing is to draw important evidence from several different classes of source rather than relying on a single type.

The three categories I have found particularly useful are government or official organisations, academic or institutional sources, and credible industry experts.

I do not claim that those categories are universally preferred by every AI system in every situation.

In some of my testing, AI platforms have cited community sources such as Reddit or other pages that would not necessarily be considered the strongest traditional authority sources. Retrieval behaviour varies by query, platform and context.

But across my own citation testing, using strong official, academic and industry sources has generally improved the evidential quality of the content and has coincided with stronger citation performance.

That is why I keep monitoring it.

AI systems are continually adjusted. Their retrieval and ranking behaviour can change, so a strategy that performs well today may need to evolve.

For me, GEO is therefore not a one-off optimisation exercise. It requires observation, testing, measurement and adaptation.

Source Diversity, Monitoring and GEO Adaptation

Phil Haoyang Pang — Founder, Canlah.ai:

Always.

Paul Rowe — Founder, Chief Generative Engine Optimisation Officer & CEO, NeuralAdX Ltd:

Exactly. There is no universally applicable formula for every AI platform.

You have to understand the client, the query environment, the sources being retrieved, the evidence available and the behaviour of the specific AI systems you are targeting.

Then you test, measure, learn and adapt.

Phil, it has genuinely been a pleasure speaking with you. We have covered why businesses may fail to appear in AI answers, how crawlability and passage-level retrieval affect visibility, why semantic relevance and generative answer coverage matter, how authority and third-party evidence influence retrieval, how to think about GEO measurement and ROI, and why repeated multi-platform testing is increasingly important.

Thank you for joining me on the NeuralAdX GEO & AI Search Strategy Podcast.