I asked ChatGPT to shop like a human – Here’s what broke (and what worked)

When OpenAI finally launched ChatGPT Agent, we read how AI agents were about to change everything and that the new age of agentic mode is coming to change our lives — already here.

So I had to see it for myself. I decided to test out Agent mode to get a firsthand look at what “agentic interactions” actually mean in practice.

In this post, I’m sharing a real-world test: how ChatGPT Agent handled a product search and made an actual inquiry on a furniture website. You’ll see what worked, what didn’t, which websites handled the agentic interaction, and which ones broke under it.

Let’s dive in!

How to use ChatGPT Agent mode

1. You’ll need a ChatGPT Plus subscription (or higher) to access Agent mode. Just click the ➕ icon and select Agent mode from the menu.
2. Activate Agent mode and give it a task. Something like, “Find me a luxury sofa and request a quote.”
3. The Agent gets the assignment. It recognizes this as a multi-step task that involves both searching and interacting with websites.
4. It starts browsing the web. The Agent runs multiple search queries, checking out different sites to gather options.
5. It sifts through the search results. It scans snippets, opens top-ranking links, and starts evaluating based on things like product details, price, reviews, and availability.
6. It selects the best options. From there, it shortlists the most promising products or suppliers.
7. At the final stage, it gets ready to reach out. The Agent drafts a message to execute on your product inquiry or quote request.
8. Before completing the task, it checks with you first. It asks for your approval before sending an inquiry or making the purchase.

If you’re a visual person like me, here’s how that agentic interaction sequence could look like:

My prompt

To start, I gave ChatGPT Agent this prompt:

“I want to buy luxury furniture. Get a quotation for a luxury outdoor daybed in Sydney, Australia.”

The luxury outdoor daybed, Sydney was a main keyword target.

Also, I wanted to test whether our client’s website (Osier Belle) was “agentic-AI friendly”. Basically, could an AI agent complete a real conversion task on their site?

Oh, ChatGPT responded with a smart follow-up: asking when I needed the product delivered and if I had any material preferences (like teak, wicker, or aluminum). I told it I wanted it by Christmas and preferred wicker.

Now the fun part—watching how Agent goes about completing the task just like a human would in a browser! This entire interaction took 11 minutes, ChaGPT executed 13 searches and discovered 108 sources.

1. Agent starts with live keyword searches.

Based on my original prompt, here are the keyword search phrases ChatGPT Agent used:

• wicker outdoor daybed Sydney luxury in stock
• wicker daybed Sydney available
• Osier Belle daybed Sydney
• Tressé daybed Osier Belle
• Wicker Outdoor Day Beds Lavita
• lavita wicker daybed
• wicker daybed site:amazon.com.au
• luxury outdoor wicker daybed Sydney Australia

This is the multi-query retrieval technique ChatGPT uses. Instead of searching with a single keyword query, Agent fans out across multiple queries to gather the most relevant information.

2. Agent retrieves real-time search results

Next, ChatGPT Agent pulled websites directly from Google and Bing. These are live search results, not static databases. That’s part of why OpenAI recently added Google as a source: it provides fresher, more relevant pages.

When I asked ChatGPT how it chose those keywords, it explained that it tries to balance relevance, specificity, and local context based on the task.

3. Agent selects the #1 search result

From there, Agent showed a list of top-ranking websites and selected the #1 result on Google—in this case, Luxo Living. I tested this manually and confirmed it was indeed ranking first at the time.

Interestingly, Bing showed different top contenders, Bay Gallery and Osmen. But Agent leaned toward Google’s results.

Is that ChatGPT preferring Google’s search results? Or just a case of better relevance for the query? Something to keep an eye on.

4. Agent visits the #1 ranking result but encounters dynamic content

Here’s where things got tricky.

Although Luxo Living’s product category page was SEO-friendly (great keyword targeting, detailed product content, clear FAQs), ChatGPT Agent couldn’t interact with it properly, and eventually exited their website.

These were the dynamic content issues ChatGPT explained it encountered:

• Client-side rendering: The HTML was mostly empty until JavaScript (JS) loaded the actual content, which the Agent’s browser environment doesn’t execute.
• Redirect issues: The site returned a redirect.html placeholder, which JS normally replaces with real content. Agent can’t run that JS.
• API-only product data: Product names, prices, and specs were loaded via background requests (XHR), not available in the raw HTML.
• Bot protection: Likely Cloudflare or CDN checks were blocking bots unless they passed a JavaScript challenge.
• Infinite scroll/lazy loading: Products loaded only as the page was scrolled. Again, this was a JavaScript-driven challenge.

Because of these barriers, Agent couldn’t extract product info from Luxo Living. So it pivoted to other websites that used server-side rendering (like Osier Belle, United House Furniture, and Lavita), which served clean, accessible HTML with all the needed details: prices, specs, delivery options.

I find this to be a major website accessibility factor we can all improve and optimize for! When agentic interactions become more widespread, this will be an optimization opportunity for website platforms like Shopify and BigCommerce.

These were the dynamic content ChatGPT encountered on their website when I followed up and asked what specific issues prevented further action:

This is probably an understatement. But I think one of the easiest and most effective ways to optimize for ChatGPT or LLMs (Large Language Model) is to use the AI platform and ask it, how can I do better? What problems did you encounter?

Now, this is a major problem. Despite great SEO and ranking #1 on Google, Luxo Living’s site had Javascript issues and Agent exited their site, and they lost the conversion.

So if you’re ranking at the top of search results, but your website does not support agentic interactions, you’re losing traffic and revenue.

5. Agent goes to the next website with accessible static pages

Agent moved on to sites with clean, server-rendered HTML. Like Osier Belle and Lavita where it could easily extract product info.

For each URL fetched, Agent performs a quick quality check:

• Confirms the page has loaded (status 200, not a 404 or redirect).
• Inspects the HTML snapshot available at load time (not the JavaScript-rendered version a human browser would see).
• Assesses content relevance. The page must directly match the search intent (e.g., a product page for a wicker daybed, not just a category or blog post).

6. Agent looks for key product details in the structured data

Once a page passes validation and loads cleanly, Agent looks for key product details:

• Product name/model
• Price (including “from” indicators)
• Material (e.g., PE wicker, teak, Sunbrella fabric)
• Dimensions/seating capacity
• Warranty info
• Shipping or delivery estimates (to Sydney)
• Stock status (“in stock”, “made to order”, etc.)

To verify consistency, Agent cross-references within brand sites, opening multiple related pages:

• On United House Furniture, it compared daybeds like Newport, Coco 3-seat, and Coco 2-seat.
• On Osier Belle, it reviewed the Tressé Double and Revé to cross-check materials and pricing.

If essential details aren’t pre-rendered in the HTML, the site is marked as “dynamic/unreadable,” and Agent moves on.

7. Error handling & Fallback logic

If a site is detected as JavaScript-heavy (like Luxo Living or Bay Gallery), Agent recognizes it by:

• Seeing empty page bodies or placeholder redirects.
• Failing to find usable product data.

In those cases, Agent won’t attempt blind scraping. Instead, it:

• Re-queries search engines for alternate suppliers or similar models.
• Logs the exclusion reason (e.g., “client-side rendering issues”).

Fallbacks included Osier Belle, Lavita, and Remarkable Furniture, where content was cleanly server-rendered. This was why those websites were chosen at the end.

Conclusion

Testing ChatGPT Agent showed how AI can interact with the web like a real user. But with some important caveats. While the Agent is smart and can get things done, it depends heavily on how accessible and readable your website is. If your content is hidden behind JavaScript or dynamic loading, the Agent simply can’t “see” it. And if it can’t see it, it can’t buy from you, request a quote, or even consider your business in a comparison.

It moves on to the next website, and you lose the conversion.

So what can you do? Start by testing your website’s accessibility in Agent mode. Assign it a real task a user might do, like requesting a quote or checking product availability—and see how far it gets.

Then, opt for server-rendered HTML wherever possible, especially for product names, prices, stock status, and shipping info. Avoid hiding key data behind JavaScript or modal pop-ups.

Also, don’t ignore traditional SEO. Your search ranking still plays a major role in whether Agent finds your site at all.

The impact of Agentic AI on websites and SEO comes down to two things:

1. Whether agents can find your site in search results.
2. And whether they can actually use your website and complete the task once they land there.

Your website’s accessibility can make or break the interaction. If it’s clunky or confusing, the Agent may abandon ship and move on to a competitor’s website.

And just like traditional SEO, your ranking in Bing and Google still matters. Because it helps determine whether the Agent even finds your site in the first place.

Agentic AI is available now and can be widely adopted. If your site can’t support it, you risk being skipped by the bots, and soon, by the people relying on them.