AI Does Not Fix Confusion. It Scales It.
Thirty questions about AI adoption, business performance, and AI visibility—and the hard lessons behind a more human approach.
By Miklos Roth | AI Strategy, SEO, and Complexity Research
I have spent more than twenty years working with businesses on digital performance. I have also spent enough time with artificial intelligence to make mistakes that I would rather discuss than conceal. One of them was simple: for a while, I used AI for almost everything. I was fascinated by what it could produce. But producing more is not the same as achieving more. Eventually I had to ask whether all those prompts, reports, and experiments were improving revenue, profitability, or the quality of our work.
That question changed how I work. Thirty interview questions led me to one uncomfortable conclusion: a powerful model cannot compensate for a business that has not clarified what it wants or how it operates. It may accelerate the confusion instead.
Start with four questions, not four software subscriptions.
The four-question method: clarity before software.
Before recommending a model, a platform, or an automation, I ask a leadership team four questions: Who are we? What do we want? Why do we want it? How do we want to achieve it?
These are business questions, but they are also psychological questions. People often ask an AI to improve performance by 20 percent without explaining which performance measure matters, why that measure matters, whom the business serves, or which trade-offs it will accept. The answer may be articulate and entirely beside the point. If a company needs better customer retention but asks only for more acquisition, even a technically competent response may reinforce the wrong priority.
The quality of the prompt matters, but a clever prompt is not enough. Relevant historical data, previous analyses, operating constraints, customer knowledge, and a definition of success have to enter the conversation. The model does not know the unstated history of the company. When it delivers something irrelevant, the problem may begin with the question we gave it. We must accept our share of responsibility.
I sometimes describe AI as an extension of our own thinking. It can help us examine far more information than we could alone. But it is not a human colleague with our lived experience or accountability. Neural-network analogies do not mean machines think exactly as people do.
A fast task inside a broken system is still a broken system.
I have seen organizations confuse AI activity with transformation. They purchase licenses, celebrate pilots, and count generated documents while the actual decision process remains unchanged. A report that takes minutes to prepare can still wait in a two-week approval queue. One department may save hours creating proposals while the next spends those hours repairing them.
The unit of analysis is the complete workflow. I begin with a baseline: how was the work performed before, what information supported it, and what did a successful result look like? Then we can redesign rather than simply accelerate.
Sometimes five approval stages could become three, provided the two removed stages did not add independent protection. Human review must remain. A meaningful expert check is more valuable than several ceremonial signatures. The time recovered can go into better analysis, stronger source verification, and decisions that actually improve the outcome.
I favour a standard AI operating system: documented inputs, ownership, steps, quality criteria, and a feedback loop. It should survive managers and technology while preserving evidence of what worked. If the underlying process is chaotic, I would initially use AI only for contained, supervised tasks. Automating the whole disorder makes the disorder faster.
My most useful lesson came from a rejected document.
At our agency, AI has helped us produce more content across different areas of work. Most clients have been satisfied. Yet one episode taught me more about reliability than a successful demonstration ever could.
We sent a client a document of about sixty pages. We had carefully reviewed the first thirty, and they looked good. We failed to give the second half equal attention. That half was weak enough for the client to reject the work. We corrected the document, revised our instructions, and reviewed it in full. The client was satisfied with the revised result.
The failure was ours, not merely the model's. We had mistaken a strong sample for a verified deliverable. An AI system can perform well nine times and fail on the tenth; that tenth output may be the one that causes the damage. Human oversight must cover the work that will actually be used.
I compare working with AI to learning to walk or learning a language. We make attempts, fall, correct, and become more fluent. The useful part of the analogy is repeated feedback, not anthropomorphizing machines. People need permission to make mistakes in supervised learning, while consequential mistakes must be caught before publication or deployment. For many operating tasks, feedback belongs within hours or a day, not weeks later.
Treat the company as a complex system.
S-I-C-T, a proposed conceptual lens for business systems.
My research at Roth Complexity explores a framework I call S-I-C-T: Structure, Information, Cohesion, and Transformation. Its practical ambition is to give different disciplines a shared way to describe how systems change.
In a business, structure includes processes and boundaries. Information concerns the data and understanding available to people. Cohesion concerns whether teams, incentives, and interpretations support one another. Transformation is the change we can observe. I am exploring whether carefully selected, company-specific proxies can make these dimensions measurable and reveal overlooked constraints or opportunities. This is research to operationalize and validate—not a commercially proven equation that automatically discovers growth.
Use-case lists describe visible departmental tasks. A systems diagnosis also examines the spaces between them. Employees and executives often disagree because each sees incomplete information and constructs a different picture. Shared definitions, education, and regular feedback can reduce that misunderstanding. Fear and hype are not opposites here; they can be parallel responses to uncertainty.
AI visibility exposes the limits of generic content.
Search marketing has made this problem especially visible. AI can produce thousands of competent-sounding pages. Much of that content repeats what every competing page already says. In our own agency work, I have watched a business appear in AI-generated answers and later disappear. When I compared competing pages, I sometimes noticed that more distinctive source material was being quoted. I do not know the hidden ranking formula.
My working hypothesis is that original expertise, specific evidence, and an identifiable brand offer an advantage over interchangeable prose. Sometimes a relatively poor page is mentioned while a more polished one is not; an AI answer can even describe a recommended company inconsistently. I investigate, rather than claim to know how sources are selected.
For me, AI visibility extends SEO into a more integrated examination of technical accessibility, content, entities, links, PR, and brand recognition. The writing can be faster, but examining the additional signals and reviewing drafts often makes the total work harder, not easier. Clients do not always recognize that trade-off. If AI search vanished tomorrow, technical quality, useful content, credible references, and clear visual communication would still matter. Too often, leaders choose PPC because its daily numbers are easier to explain than foundational work. To me, technical SEO resembles the keeper in football: when it protects the team, nobody applauds; when it fails, everybody notices. Paid advertising has its place, but it cannot substitute for those foundations.
Work backward from the result, but do not pretend to know the algorithm.
I usually begin with Semrush and available analytics: organic traffic estimates, brand searches, keyword positions, referring sites, and public relations coverage. I examine what Google, Bing, and relevant AI assistants actually show for specific questions. Then I compare competing businesses and work backward through their visible technical, editorial, and external signals. I have called this a regression-like approach because it is comparative and iterative; it is not a fitted model of a search company's secret ranking weights.
AI visibility tracking is still imperfect. Third-party tools sample queries and estimate exposure; they cannot provide a complete census of every personalized AI answer. Google Search Console, Google Analytics, first-party lead records, and internal measurements each answer different questions. I triangulate them instead of treating any dashboard as the truth. Nor would I blindly average unlike metrics: traffic, sampled mentions, and attributed sales need their own definitions before they can be interpreted together.
My years in competitive athletics shaped this habit of continual testing. You examine the current result, identify what the strongest performers do, change one thing you can defend, and measure again. The objective is not a prettier visibility chart. It is more qualified interest, more revenue, and, ultimately, more profit.
Small brands can compete through specificity, not imitation.
When large companies dominate broad search terms, I often move toward precise, four- or five-word customer questions—sometimes longer, sometimes absent from commercial keyword databases. A car buyer may search for a particular model, year, and specification rather than simply 'cars for sale.' A large brand cannot cover every possible product detail with equal depth. A smaller specialist sometimes can.
The same logic applies internationally, but translation alone is insufficient. I would first research local demand, purchasing power, competitors, likely ROI and the questions appearing in local search and AI answers. I often favor a country-specific domain and a genuinely localized website, possibly maintained through one underlying system. It is a preference, not an automatic search advantage.
Brand visibility does not require disclosing confidential information. Nike, Red Bull and other enduring brands illustrate the distinction between communicating a human story and publishing operational secrets. A company's public story should connect a real customer problem to its own distinctive way of helping. Confidential information stays private.
A successful campaign must survive the financial question
Visibility only matters when it supports real business value.
One of my strongest disagreements with superficial marketing is its dependence on traffic as a proxy for success. If AI mentions rise but inquiries do not, I look beyond the content. Is the product competitive? Is the price appropriate? Are we reaching the buyers who can afford and value it? Is the offer convincing once they reach the website? Sometimes the problem is segmentation or the business model, not the ranking.
I have described a one-hour intervention for a business with limited resources: find an important existing landing page, review the strongest relevant search questions, supplement rather than erase its expert content, add useful FAQs and prepare external or social distribution through our automated workflows. I can deliver the work quickly; AI recommendations are not guaranteed. The next stage is to observe indexing, visibility, qualified visits, inquiries and return on investment over a suitable period.
Not every company should buy this work. SEO-based AI visibility is generally a medium- to long-term investment. If a business cannot cover immediate obligations, I would not push an investment whose return may take time.
Psychology determines whether capability becomes performance
Small supervised tasks and frequent feedback build useful AI capability.
My interest in psychology and self-knowledge developed over roughly a decade. It eventually became inseparable from my work with AI. I have a teacher's instinct. As I explored AI, I realized that many of the obstacles were not software defects. They were misunderstandings, unrealistic expectations, fear of failure and an unwillingness to examine our own habits.
I know the temptation personally. On well-structured days, twenty-, thirty- or sixty-minute meetings with clients can lead to focused, high-impact decisions. On quieter days, I can still lose time testing interesting new AI tools without a concrete business reason. My safeguard is goal-oriented work: let the objective choose the tool, method and workflow—not the other way around. Intuition can be useful, but mine is intuition tested against years of projects and revised when results disagree.
I also think leaders should understand when AI is being overused. If the team spends increasing time interacting with models while sales, quality and profitability stagnate, more AI is not automatically the remedy. The experienced human still has to connect context, trade-offs and responsibility. AI is excellent for research, searching, drafting and exploring alternatives; deciding what is right for a particular business remains our job.
Education is part of my professional mission. The first step into AI is often the hardest. Start with a small, low-risk task. Practice, inspect the result, correct the error and move gradually to more demanding work. Build a culture where learning is permitted and carelessness is not.
I do not believe the future belongs to the company with the most subscriptions or the longest collection of prompts. It belongs to organizations that know who they are, know what they want, know why it matters, and can test whether their chosen way forward actually works. AI can extend what those organizations achieve. It cannot supply their purpose for them.
