I Asked AI to shop for me <insert clickbait>

That title was mine. Not AI. But hey, that’s how it works right. Your AI tool know better than you and knows how to go viral. Right?

Any way this was the challenge I set to see how AI search actually works so I can, hopefully validate what VIA Labs is all about!.

I ran the same prompt through ChatGPT and Gemini.

First: Find me a charcoal BBQ for terrace use, four people, under 100 SGD, available in Singapore.
Then: explain exactly how you selected and ranked those results.

Both models answered the third question honestly. And here is the analysis:

What the models returned

Both gave me a near identical shortlist: best overall, budget option, compact tabletop, foldable. Both followed with a top 10 supplier list. The outputs looked structured and useful.

Bu when I pushed on methodology, ChatGPT said:

“I do not control ranking algorithms. I do not know if any result is sponsored. Results often skew toward Lazada, Shopee, Amazon-style listings because those dominate structured product data.”

Gemini went further and named its biases explicitly: “availability bias toward commonly stocked items, platform bias toward large marketplace listings, practicality bias toward lower-risk options.”

Two different models. Same inputs. And near-identical outputs and disclosures.

What is actually happening

Both ChatGPT and Gemini are pull systems. Before a query arrives, the data they draw from has already been shaped by years of platform economics. The AI applies intelligence on top of distorted input. It cannot actually fix the distortion.

Lazada, Shopee, (the main marketplaces over here) and their equivalents dominate structured product data because they have the engineering resources, SEO investment, and volume to populate that data at scale. A specialist supplier with excellent product but no digital marketing team simply does not appear.

The AI is not choosing the large marketplace over the specialist. It is reflecting a data environment where the large marketplace already won before the query ran.

Neither model can confirm stock. Both returned products described as available without any live verification. You, the buyer still has to click through, check the listing, confirm the price is current, and navigate a checkout independently. So no better than a normal search!

The refinement offer ws a telling detail. Both models offered to switch modes and improve results when the first pass was imprecise. That confirms the initial output is acknowledged as approximate, improvable only through additional human input.

Independent research confirms the pattern

A recent study by AI Plus Automation analysed 10,138 citations across ChatGPT, Perplexity, and Google AI Mode. AI citation is driven by Google ranking breadth, not product quality. Sites ranking for one query had a 34% citation rate. Sites ranking for 16 or more had a 100% rate.

AI discovery rewards accumulated SEO investment, not relevance to the buyer’s actual need. The merchant who spent years building search presence gets cited. The specialist with the better product but no SEO budget does not appear.

How VIA works differently

VIA is not a search engine layered on top of existing platforms. It is a protocol that connects verified buyer intent directly to verified merchant agents, with economic mechanics that change who responds and why.

When a buyer - or a buyer’s Personal Shopper or Concierge - expresses a purchase intent, VIA broadcasts it as a structured signal across the network. The message carries product requirements, constraints, location, and verified buyer identity. No merchant is pre-selected. No ranking is applied. The intent goes out neutrally to all merchant agents on the network.

Merchants respond only if they can actually fulfil. The act of responding costs a micro-fee, deducted from their pre-funded wallet balance. This single mechanic changes the entire economics of the interaction. A merchant paying a micro-fee to respond to a specific buyer intent is motivated by actual match, not by visibility budget. Blanket responses to every query regardless of fit become economically irrational. The micro-fee turns carpet-bombing into a cost that erodes margin.

Every merchant agent on VIA holds an ERC-8004 on-chain identity. Every buyer agent is also verified. When an AI agent is completing a purchase without human oversight, it needs to know it is transacting with a verified counterpart. LLM search pulling from marketplace data cannot provide that.

Payment options include settlement via the x402 protocol - on-chain, no redirect to a checkout page, no third-party processor in the critical path, no manual re-authentication. For autonomous agents, this is what closes the loop. LLM search ends with a suggestion. VIA ends with a completed transaction.

The same query, two architectures

The BBQ test makes the difference concrete. A buyer needs a charcoal BBQ: terrace use, four people, under 100 SGD, in stock.

Via ChatGPT or Gemini: a shortlist of products that appear frequently in indexed marketplace data, with acknowledged caveats about stock, sponsored placement, and platform bias. The buyer has a list. They do not have a confirmed, purchasable offer.

Via VIA: a buyer agent broadcasts the intent. Merchant agents with matching inventory receive the signal. Those who can fulfil respond, paying a micro-fee. Each response is from a verified merchant with live inventory. The buyer agent selects on price, delivery speed, or trust tier. Payment settles on-chain. No caveats. No click-through. The loop closes.

Why this matters for agentic commerce

The shift from human-directed search to agent-directed purchasing is already underway. ChatGPT Operator, Google AI Mode, Amazon Buy for Me, Perplexity shopping - all in market. The question is not whether agents will be doing the buying. It is whether the infrastructure they run on is built for that purpose.

An AI agent completing a purchase does not click ads. It does not browse visually merchandised storefronts. It does not respond to promoted listings. The entire stack of digital marketing that shapes what surfaces in LLM search has no effect on an agent executing a transaction against a protocol. Merchants who have built their discoverability on SEO and paid placement are invisible to agents operating natively through commerce protocols.

For an agent to complete a transaction autonomously, it needs to know the merchant is who they say they are, that the inventory is there, and that payment will settle cleanly. LLM search was not designed to provide any of these. It was designed to help humans find information and make their own decisions.

Physical retail required shelf placement. Digital retail required SEO and paid search. Agentic commerce requires protocol participation. A merchant agent connected to VIA is reachable by any buyer agent that routes intent through the protocol, regardless of marketing budget, platform presence, or search ranking. A specialist BBQ supplier with no digital marketing presence becomes as reachable as a major marketplace merchant. Both pay the same micro-fee to respond to a matched intent. The buyer agent selects on fit.

The ChatGPT and Gemini tests produced the same result because they draw from the same structural reality. LLM search layers intelligence on top of data shaped by advertising economics before the query arrived. Both models said this plainly when asked.

For human buyers tolerating friction, that is workable. For AI agents completing transactions autonomously, it is not.

VIA is built for the agent-to-agent layer. Intent goes out verified. Responses come back from merchants with economic skin in the game. Settlement happens on-chain. The loop closes without a human in the middle.

The brands that get found in agentic commerce will not be the ones who spent the most on visibility. They will be the ones whose infrastructure speaks the same language as the agents doing the buying.

VIA Labs Pte Ltd | getvia.xyz