DEV Community: Saulius Steponkus

When Customers Stop Searching Websites: How Businesses Can Be Found in AI Conversations

Saulius Steponkus — Fri, 13 Nov 2026 07:40:00 +0000

AI is changing not only search tools but the decision-making process itself. Businesses need to focus not only on rankings, but also on whether their information is clear, current and useful in conversational AI.

Introduction: search is becoming a conversation

People increasingly want more than a list of links. They want a clear answer: what to choose, how two solutions differ, which option fits their situation and what to do next. This behavior is changing the rules of business visibility.

In an AI environment, a company is not competing only for a position on a search results page. It is competing for the chance that its information will be clear, trustworthy and current enough to become part of an answer.

Several recent examples point in the same direction. JioStar built a conversational content discovery experience that lets viewers use voice to search for movies, shows, sports moments and cricket statistics. At the same time, Businesses in ChatGPT gives companies ways to manage information about themselves, understand how people engage with them and, in some cases, interact with customers through a Business Agent. Together, these examples show that discovery is becoming contextual, conversational and focused on decisions.

From keywords to user intent

Traditional search often starts with a few keywords. In a conversation, users provide more context: their budget, situation, priorities, concerns or even emotional state.

In JioStar’s example, a person can describe a mood instead of entering a film title, or search for a specific moment in a sports match. This means business content should be built not only around phrases, but also around the real questions customers ask.

For example, a website about accounting services needs more than a page titled “Accounting Services.” More useful content answers questions such as:

When should a small company outsource its accounting?
What is the difference between monthly accounting and financial management advice?
What documents should a business prepare before choosing a provider?

Questions like these help AI understand not only what a company offers, but also when that offer becomes relevant in a customer’s decision process.

Accurate information becomes the foundation of visibility

The Businesses in ChatGPT example shows that companies can claim a business profile, update key information and provide product data such as pricing and availability. This is a useful lesson even for companies that do not yet use these specific features.

AI systems need information that is consistent, current and free of contradictions. If a website lists one set of service conditions, social media presents another and directories show an old address, the resulting answer may be unclear or unreliable.

That is why the first step is not necessarily another blog post. It is often better to strengthen the information foundation:

describe services and their limits clearly;
explain who each service is for;
update prices, availability and contact details where relevant;
align company descriptions across channels;
separate verifiable facts from marketing promises.

Strong visibility in AI answers begins with strong information management.

Content discovery is more than recommendation

JioStar’s experience reveals another important point: users do not always need a text-only answer. The company designed content discovery as a combination of visual and conversational interaction. People can speak naturally, while still seeing options that help them decide.

For businesses, this means content should work across several modes of use. A service page should be understandable to people, accessible to search systems and structured clearly enough for specific information to be extracted.

In practice, this can include:

clear service categories;
comparison tables;
frequently asked questions;
specific use cases;
consistent product or service descriptions;
clear next steps for the customer.

The value of this content is not limited to traffic. It helps people understand their options faster and helps AI connect a question with the right offer more accurately.

Trust matters more than a single click

Gilbert + Tobin’s experience suggests that AI adoption works best when an organization creates a safe environment for experimentation, establishes clear rules and keeps people accountable for the final result. The same principle applies to business visibility.

The goal should not be to make AI say as much as possible about a company. The goal should be to make it say what is accurate. Marketing content needs to reflect real capabilities, and automated answers need clear boundaries.

This is especially important in areas requiring professional, medical, legal or financial advice. Responsible content strengthens trust and reduces the risk of incorrect interpretation.

Where should a business start?

A practical process can be straightforward.

First, collect 20 to 30 real customer questions. Include not only product names, but also concerns, comparisons and situations in which decisions are made.

Second, check whether the website answers those questions directly. If an answer is hidden inside abstract marketing language, rewrite it in a clearer structure.

Third, align the core information across channels. Addresses, opening hours, service scope, pricing and contact details should not change depending on where a customer finds them.

Fourth, establish a review process. Information should be checked not only before publication, but also later, because offers, prices and availability change.

Fifth, measure more than visits. Track which questions customers ask most often, which topics lead to contact and where users still struggle to find a clear answer.

Conclusion

AI search is not a separate channel that can be fixed with one technical trick. It is a new way for customers to understand a market, compare options and begin a conversation with a business.

In this environment, the winning company is not necessarily the one that publishes the most content. It is the one whose information is clear, consistent, current and shaped around real human questions.

That is why SEO strategy needs to expand: from keywords and rankings to information quality, user intent and readiness to be understood correctly in an AI conversation.

Originally published at insights.aisys.lt/kai-klientas-nebeiesko-svetaineje-kaip-verslui-tapti-randamam-di-pokalbiuose.

From AI Experiment to Operating Model: Lessons from the New Generation of Agents

Saulius Steponkus — Thu, 12 Nov 2026 20:03:35 +0000

From AI Experiment to Operating Model: Lessons from the New Generation of Agents

AI agents are moving from isolated experiments into real operating processes. Recent research and business examples show that the greatest value comes not only from model capability, but from clear boundaries, validation and cross-functional adoption.

AI agents are changing how work is organized

For years, businesses introduced artificial intelligence as a standalone tool: a chatbot, text generator or coding assistant. Recent examples point to a broader shift. AI agents are becoming part of operating processes. They analyze information, help build solutions, test security and hand results back to people for review.

This changes the starting question for companies considering AI adoption. Instead of asking only which model to choose, leaders should ask which process can be redesigned reliably once the agent’s responsibilities and human approval points are clearly defined.

Three lessons from recent AI agent deployments

1. Value appears when the agent is connected to the full workflow

Indian fitness and healthcare company cult.fit began by standardizing software development. Codex was used not only to generate code, but also to create shared conventions for development, testing, review and feedback. The company later extended agents into incident analysis, security work, data analysis and internal application development.

The lesson is straightforward: an agent creates more value when it is connected to a defined operating cycle that includes a requirement, relevant data, validation, an accountable person and a measurable outcome. A one-time generated answer may save minutes. A repeatable workflow can change how an entire team works.

Cult.fit reports that non-engineering teams built more than 20 internal applications in recent months. In another case, the company rebuilt an internal dashboard capability and avoided more than $100,000 in annual software costs. These figures should not be treated as a universal promise. They do show, however, that AI can reduce bottlenecks around centralized teams when employees receive the right tools, data and support.

2. Reliability begins with recognizing when evidence is insufficient

BioBasicsBench evaluated how AI agents interpret everyday laboratory evidence, including gel images, microscopy, instrument readouts, protocols and other imperfect inputs. The 100 tasks tested more than answer generation. Agents also had to identify errors, check controls and recognize when the evidence did not justify a confident conclusion.

At the stated reasoning setting, GPT-6 Astra earned an average of 66.7% of available credit, compared with 43.8% for GPT-5.6 Sol. The results show progress, but they also underline an important limitation: even a capable agent is not an automatic source of truth. It needs an environment that defines when an answer is accepted, when additional review is required and when the process must stop.

The same principle applies outside science. Finance, customer service, legal operations and business analytics all involve incomplete fields, conflicting signals and ambiguous inputs. A useful AI system should not only produce an answer; it should also indicate what is missing before a decision can be trusted.

3. Security must be part of the workflow

OpenAI’s partnership with Hacktron illustrates how AI agents can be used not only to build systems, but also to discover weaknesses in them. The partnership focuses on high-priority areas, including how systems interact with third-party services, where complex vulnerabilities can emerge.

GPT-6 Cyber is described as a model designed to find and validate vulnerabilities while remaining within a researcher’s authorized scope. In reported evaluations, it avoided exploiting out-of-scope services in 99.93% of tests without an additional deployment prompt and in 100% of tests with that prompt.

For businesses, the practical principle is clear: the more actions an agent can take, the more important it is to define permissions, data boundaries, approval points and audit trails in advance. An agent that can prepare a code change or propose an action should not automatically be allowed to deploy it without human review.

How should a company choose its first AI agent use case?

These examples suggest starting with a process that meets four conditions:

the desired outcome is clearly defined;
the agent has access to the necessary but limited data;
errors can be detected through review;
a named person or team remains accountable for the final decision.

Suitable starting points may include recurring data analysis, incident summaries, internal application prototypes, code review or customer-service classification. The first goal should not be complete automation. A safer and more useful approach is to have the agent prepare a result while a person evaluates it against predefined criteria.

AI literacy is becoming an organizational capability

The experience of Australian universities shows that preparation for an AI-enabled economy begins before graduates enter the workforce. In the cited survey, 81% of students said they already used AI several times a week, 28% had built an AI agent or assistant, and 79% wanted more opportunities to use advanced AI tools on real-world projects.

Access to a tool does not create competence by itself. Universities are also developing frameworks for appropriate AI use in learning and assessment, combining practical projects with subject knowledge and human judgment. Businesses need the same approach. Employees must be able to assess source quality, identify uncertainty, verify outputs and understand when an agent’s actions should be restricted.

Conclusion: the advantage belongs to the better operating system

The latest examples lead to one common conclusion: competitive advantage comes not from the model alone, but from the organizational system built around it. That system requires clear processes, reliable data, security rules, employee capability and measurable outcomes.

Companies should begin with one concrete process where AI can analyze, build or test while responsibility and control remain visible at every stage. That is how an AI experiment becomes a managed change in the operating model rather than a short-lived technology trend.

Originally published at insights.aisys.lt.