Introduction

I’m Mia Sato, an AI researcher.

On February 24, 2026, Google announced a new agent capability for the “generate” step in Opal. This update makes it easier to choose tools and models based on your goal, and it also improves the system’s ability to remember context, ask follow-up questions when needed, and change its workflow depending on the content. In this article, I’ll break down those changes as simply as possible and look at where they may be especially useful in e-commerce operations.

At GDX, we often hear questions from teams running e-commerce sites, such as:

“Ad performance is down. Should we also look at discount promotions at the same time?”
“Can you draft product descriptions that match our brand tone?”
“Can you separate FAQ drafts by inquiry type?”

These kinds of questions come up frequently in day-to-day e-commerce operations. With this latest Opal update, it has become easier to carry context through a task, ask for missing information along the way, and shift the flow depending on the situation.

That means tasks like reviewing ad reports, creating first drafts of product descriptions, and organizing FAQ drafts may now be completed more quickly than before. However, to make these use cases work in real operations, teams still need to define at least some basic assumptions in advance, such as prohibited expressions, brand tone, and customer support policies. One of the key strengths of this Opal update is that it makes follow-up questions and branching workflows easier to apply in real business settings where those kinds of assumptions already exist.

First, what does Opal do?

Opal is Google’s tool for building AI mini-apps using natural language.

Google officially describes Opal as a way to “build, edit and share mini-AI apps using natural language.” In other words, rather than stopping at conversation, Opal is designed to make it easier to build small apps that take input, process it, and return results. The Overview and Quickstart materials also explain that users can build multi-step workflows using both a natural-language editor and a visual editor.

A simple way to think about it is this: earlier versions of Opal were tools for “building a workflow and running it,” while the new Opal is better understood as a tool for “building a workflow while also reasoning through the process as it goes.”

How is it different from ChatGPT custom GPTs?

The reason it makes sense to compare Opal with ChatGPT custom GPTs is that both can appear to be tools for making AI easier to use in your own work. That can make it hard to tell what the difference is and which one to choose. In particular, tasks like writing product descriptions, answering internal questions, organizing FAQs, or drafting reports can all seem possible with either tool.

Because of that, people encountering these tools for the first time often struggle to decide whether they want to create “an AI specialized in conversation” or “a system that follows a process step by step.” Sorting out that difference first makes it easier to think about which tool fits which type of work.

Broadly speaking, a custom GPT is closer to a dedicated conversational assistant, while Opal is closer to a small app that is built around a workflow.

To put it even more simply:

A custom GPT is like “a knowledgeable store associate.”
Opal is like “a front-desk tool that walks you through a process step by step.”

If your goal is something like having AI write product descriptions or respond in line with your brand tone, the custom GPT model is a closer fit.

On the other hand, if you want to build the flow itself—for example, “review ad metrics → ask for missing information → organize the next items to check”—then Opal is generally a better fit.

Of course, individual tasks can often be done with other generative AI tools as well. What makes Opal distinctive is that it lets you turn those tasks into a shareable mini-app built around the workflow of “ask, sort, and organize.”

From GDX’s perspective: where can Google Opal be used in e-commerce operations?

For e-commerce operations, Opal tends to work especially well for tasks that involve the same checks over and over again. Examples include reviewing ad reports, drafting product descriptions, and classifying customer inquiries.

That said, to use it reliably for these purposes, it is important to have at least some baseline assumptions prepared in advance, such as brand tone, prohibited expressions, and inquiry classification rules.

Below are concrete examples of how the capabilities Google describes for Opal can be applied to e-commerce operations. In each case, the point is not just to generate text, but to confirm information partway through, branch depending on conditions, and make the whole workflow easy to reuse. That is exactly the kind of scenario where Opal is a strong fit.

1. Ad report review app

Where it can be used: Before weekly or monthly ad meetings
What the app does: You enter a KPI summary, and the app returns “key changes → possible causes → items to check next.”

For example, if CVR has dropped, the app could ask follow-up questions such as: “Were there any discount promotions?” “Are out-of-stock products included?” or “Was the landing page updated?” It would then organize the discussion points based on those answers as well. This is a use case where the newly added interactive chat and dynamic routing are especially effective. With memory, it also becomes easier to maintain assumptions such as “for this brand, gross profit matters more than ROAS.”

How to use it:

Open a similar demo from Opal’s gallery and edit it as a starting point.

In the initial instruction, write something like:
“Take a summary of ad KPIs and organize the key changes, possible causes, and the next items to review. If information is missing, ask whether there were promotions, inventory issues, price revisions, or landing page updates.”

Use Preview and test it with the previous week’s summary.

If the output is not sufficient, add instructions in the natural-language editor such as:
“If inventory issues are likely involved, do not draw conclusions based on ad performance alone.”
According to Google’s official materials, editing can be done in either the natural-language editor or the visual editor.

Once you have tested around three patterns and confirmed that it works well, you can share it internally.

Why this update matters:
Previously, users had to define the “checkpoints” in much more detail beforehand. Now that the app can ask for missing information more naturally, it is much easier to build a report-review app for use before meetings.

2. Product description and promotional copy drafting app

Where it can be used: Before product registration, before a sale, or when preparing a featured page
What the app does: When SKU information is entered, it returns draft product descriptions, key selling points, and short promotional copy.

For example, when you enter a product name, features, target users, and a promotional theme, the app might ask follow-up questions such as, “Should the brand tone be softer?” or “Are there any expressions we should avoid?” It would then generate copy based on those conditions. In this use case, memory is especially useful because it can retain assumptions like brand tone and prohibited expressions. If the writing style should change by category, dynamic routing is also a good fit.

How to use it:

Find a template in the gallery that is close to copy generation and edit it, or create a new app from scratch.

In the initial instruction, write something like:
“Take product information and draft a product description and promotional copy in line with the brand tone. Ask follow-up questions if information is missing. Do not use prohibited expressions.”

Start by testing just one product.

As you review the output, add conditions such as:
“For food products, avoid exaggerated claims.”
“For cosmetics, describe the feel of use first.”

Once it is approved internally, product managers can use it by entering SKU-level information.

Why this update matters:
One of the biggest improvements is that the app can now ask for missing product information on its own. That makes it easier to begin drafting even when all the input fields have not been perfectly prepared from the start.

3. Inquiry classification and FAQ drafting app

Where it can be used: Initial CS triage, internal review, or before updating FAQs
What the app does: When you enter the content of a customer inquiry, the app classifies it into categories such as “shipping,” “returns,” “inventory,” or “payment,” and then produces an internal response policy or a draft FAQ answer.

For example, for an inquiry like “My order still hasn’t arrived,” the app might ask, “Has a shipping notification been sent?” or “What was the order date?” It could then determine whether the issue is a shipping delay or a pending inventory allocation and draft a response accordingly. This is probably the clearest example of where dynamic routing is effective. Because interactive chat allows the app to ask for the information it needs, the process can still move forward even when the initial input is rough.

How to use it:

Create a new app with an instruction like:
“Take the text of a customer inquiry, determine the category, and create an internal response policy. Ask for missing information when necessary.”

Start with about four major branches.
Shipping / Returns / Inventory / Payment is enough at first.

Use Preview and test it with three to five examples similar to real inquiries. Google’s official materials note that you can refine the behavior while checking it in Preview or in the visual editor.

If the classification is off, add rules such as:
“Prioritize shipping for phrases like ‘hasn’t arrived yet.’”
“Route phrases like ‘I want to cancel’ to returns/cancellations.”

At first, it is safer to use this only for internal initial sorting rather than sending responses directly to customers. Since FAQ content can still contain mistakes, it is better not to use it for automatic customer-facing replies right away.

Why this update matters:
Previously, you needed to define inquiry branches in much finer detail from the very beginning. Now that it is easier to classify, ask, and organize, Opal is much better suited for use at the front end of customer support workflows.

Points to check before introducing it

1. Start with drafting use cases

Opal’s FAQ clearly states that Opal can make mistakes and that testing is important. Because of that, it is safer to begin with use cases like pre-meeting organization, first drafts of copy, or FAQ drafts rather than immediately using it for auto-sending or automated decision-making.

In practice, it is also easier to start by defining input items, prohibited expressions, and decision rules on a small scale rather than aiming for perfect automation from the beginning.

2. The way you share it changes what others can see

According to the Overview, Opal can be shared or published, and the people you share it with may be able to see the contents and prompts. It also notes that Opal apps are saved as files in Google Drive. If you plan to use it internally, it is better to decide in advance how much of the underlying content you want others to be able to see.

3. Think carefully before entering raw data as-is

The FAQ explains that Opal prompts and outputs are not used to train generative AI models. At the same time, it also notes that some prompts may be reviewed by humans for troubleshooting or for understanding use cases. If you are testing it in e-commerce operations, it is generally better to start with summarized values or classification labels rather than entering customer names, order IDs, or cost data directly.

Conclusion

This Opal update brings the experience closer to “handing over what you want done and letting the system guide you through it,” rather than “thinking first about which model to use.” In e-commerce operations, it is especially well suited not to full automation, but to use cases that shorten the time needed for shared context and first-draft creation.

More specifically, Opal is a strong fit not so much for one-off AI consultations, but for situations where the same confirmation steps are repeated over and over and where teams want to package those workflows into small shareable apps.

Honestly, that is probably where it will be most helpful in practice. Tasks like reports, FAQs, and promotional drafts are especially worth trying because teams often have to explain the same things from scratch every time.

The best approach is probably to start small and begin with drafting use cases.

References

Official: Build dynamic agentic workflows in Opal / Google Blog / https://blog.google/innovation-and-ai/models-and-research/google-labs/opal-agent/

Official: Frequently asked questions and best practices | Opal / Google for Developers / https://developers.google.com/opal/faq

Official: Introducing Opal: Describe, build, and share AI mini-apps / Google Developers Blog / https://developers.googleblog.com/ja/introducing-opal/

Commentary / expert analysis: Google adds AI agent to Opal mini-app builder / Paul Krill / InfoWorld / https://www.infoworld.com/article/4136919/google-adds-ai-agent-to-opal-mini-app-builder.html

Commentary / expert analysis: Google's Opal just quietly showed enterprise teams the new blueprint for building AI agents / VentureBeat / https://venturebeat.com/technology/googles-opal-just-quietly-showed-enterprise-teams-the-new-blueprint-for

For more information about GDX Inc., please visit:
Company website: https://gdx.inc/

Parts of this article were created with the support of ChatGPT and then revised and expanded by the author. The content reflects the author’s personal views and does not represent the official opinion or statement of GDX Inc. The information is provided for reference purposes only. Please refer to official announcements and primary sources.