What AI agents do when the product page runs out

When AI shopping agents could consult a brand's own product expert, they chose the right business 93% of the time. Without it, they chose correctly 47% of the time. Those figures come from a controlled test AIVO, Inc. ran on October 26, with test agents on Claude, Gemini and Perplexity that were never told the expert existed (157 of 168 against 79 of 167).

The expert is AIVO Assist, which went into active deployment this week. The more useful lesson for brand owners is in how the agents behaved, because it shows where sales are being lost today on stores that have no expert at all.

Agents ask when the pages run out

Agents read the product pages first. When those pages answered the customer's question, agents took the answer and moved on. When the pages ran out, agents went looking for someone to ask.

In 89 live tests with real consumer agents between October 30 and November 9, agents used the expert 48 times, and 34 of those were unprompted. ChatGPT agent, Perplexity Comet, Claude in Chrome, Instinct, Fo by Wajo and ordinary ChatGPT chat all used it. On a Spanish-language Shopify store, ChatGPT agent consulted the expert unprompted three times out of three.

There is a catch for brand owners. Anything on a page that looks like an answer wins, even when it is wrong. In one test, Instinct met a discount banner that contradicted the store's real terms. It overrode the banner by citing the expert's signed-off answer. On a store with no expert to ask, the banner would have stood.

Shopper phrasing matters too. A question like "Can you deliver a dress to London?" sent agents to the expert far more often than a research-style question did.

Why an honest no matters

AIVO Assist gives one of four judged answers: Fits, Fits, if…, No suitable option, or Need a little more information. Every word comes from facts the business has signed off, with sources attached.

The "no" is the part brand owners may hesitate over, and it is what makes the yes credible. An expert that can say "this serum is not a hair-loss treatment" or "we only ship within the EU" earns trust for its yes. In the controlled test, every agent that consulted the expert chose the right business, 149 of 149.

The visitors your analytics can't see

On October 26, alongside the selection test, AIVO, Inc. ran a controlled attribution test. Every session that asked the expert and then booked was credited, 14 of 14, with no false credits. A standard analytics script running alongside recorded 0 of 314 AI agent visits.

If standard analytics behave the same way on other stores, AI agents can visit, compare and leave without any of it reaching a dashboard. Each AIVO Assist answer carries a reference code and a tracked link, so the business can see what agents asked and which answers led to a sale.

What we have not proven yet

All of the live tests so far were run by AIVO, Inc., on our own sites and one merchant store. As of our November 10 log audit, there was no agent use yet from merchants' own customers. Meta's Muse, OpenAI's Dots and Grok are untested because we have no access to them in the UK or EU.

The next result we owe readers is one from a real merchant's real customers. We will publish it with sample sizes, baselines and variance, in the same way as everything else in the AIVO Standard record.

What this means for brand owners

The facts that decide a sale are usually already on your store. Agents struggle to find them, read them and match them to the customer in front of them. Product pages are written to be browsed by people, and agents arrive with one customer's specific need.

Two things follow from our tests. Contradictions on your own pages cost you sales, because agents believe whatever looks like an answer. And when an agent can ask, it often does (48 of 89 live tests), so the brand that answers at that moment is in the decision.

Sources

AIVO Assist deployment note, AIVO Standard on Zenodo, November 13, 2026: 10.5281/zenodo.24189003

AIVO Assist working paper, Sheals and de Rosen, October 27, 2026: 10.5281/zenodo.23694598

Both are DOI-anchored, not peer-reviewed.