Lindsey Aliksanyan, founder and CEO of Metricform, explains how AI shopping agents are changing holiday recommendations—and why a brand’s customer service record can determine whether it makes the shortlist.
For years, holiday shopping started with a search box, and brands learned to compete for position on the results page. This season, a fast-growing share of shoppers are handing that first step to an AI assistant and asking it what to buy.
Salesforce predicts that 20% of holiday ecommerce traffic will come from AI chat agents this year, and its latest State of Commerce research found that use of AI assistants as the first stop in the shopping journey grew 200% in a single year. Customer service leaders should pay attention, because these agents build their recommendations from the public record of how companies treat their customers and from the product data those companies publish. The narrative that formed around a brand months ago is now deciding whether it makes a shopper’s shortlist at all, and the teams who answer the complaints and respond to the reviews are shaping sales that will never be attributed to them.
When a shopper asks an assistant to recommend a gift, the agent does not watch a commercial or take a homepage at its word. It reads what has already been written across the open internet and compresses all of it into two or three recommendations, and with retailers as large as Walmart now letting customers complete orders inside ChatGPT, that shortlist sits closer to the checkout than any search result ever did. My team measured how those shortlists get made this fall, putting 32 holiday gift questions to ChatGPT, Claude, Gemini and Grok, three times each, under three different set-ups, for 1,295 coded answers. Answering from memory, with no store to search, the assistants reached for the names everyone already knows: Sephora appeared in 147 of 288 answers and Ulta in 127, while a prestige brand we tracked appeared in five. That familiarity was earned over years of accumulated reputation, and no brand rewrites it before Black Friday.
Much of that record gets written in service moments. A warranty claim that drags on for six weeks becomes a forum thread that never expires, and a delivery failure repeated across a few hundred one-star reviews becomes a pattern a language model spots in seconds. In my experience measuring how narratives form around brands, the theme that ends up defining a company rarely appears overnight; it surfaces in small pockets of conversation one to three days before it spreads widely, then hardens over months into the summary an AI recites to shoppers.
The other half of the recommendation rests on what the assistant can read about a company’s products right now, and that half can still be fixed before the holidays. When we connected the same assistants to a live shopping catalog, the prestige brand went from 5 answers to as many as 31, but being findable proved very different from being chosen: one popular brand was named in more than 80% of the answers whose search results contained it, while another surfaced in 89 searches and was never named once. From memory the assistants quoted $38 for a $40 gift set in 23 separate answers, and given the live listing they copied the price correctly every time; when a question mentioned a delivery deadline, all 72 answers addressed it. The agent is, in effect, a customer reading a company’s product data aloud to other customers, and it does not hedge.
What happens after the recommendation lands on service teams in full. Alongside the assistant study, we read 299 first-hand accounts from shoppers describing their own attempts to buy with an AI agent, and of the 125 that said how far the agent got, 119 stopped before any payment went through. The questions that stopped them, such as whether the price is real and whether the gift arrives by the 24th, arrive at a service desk, and so do the failures, from wrong picks and wrong prices to the seller who found an assistant telling shoppers his food company was halal certified when it is not. Hardly any of those contacts arrive labeled as agent traffic, which is why tagging agent-originated orders now is how a team will understand what January’s returns are telling it.
We also asked every question three times, and the brand an assistant led with changed between sessions in roughly half of cases, so a screenshot of an assistant recommending a brand is a sample of one, and a screenshot of it ignoring one is too. Visibility in these answers has to be measured repeatedly and across many questions, the way service organizations already measure satisfaction and resolution time. Very few bring that discipline to how their company is talked about, or to how quickly word of a damaging claim reaches leadership. Closing that gap is what I call narrative measurement, and this is the first holiday season where it will show up in the sales numbers.
Skeptics will note that agent-referred visits remain a minority of traffic, and that all of this can sound like search optimization wearing a new name, a problem for the marketing department to solve with better-structured product data. Part of it is, and our own catalog numbers say the feeds are worth fixing. But a catalog only determines whether an assistant can see a brand. The choosing rests on what customers have already said and on whether what the machine reads is true, and the early advertiser reports our research collected were mostly weak. Treating this as marketing’s problem alone leaves out the one team that produces most of the material the machines are reading.
Customer service has always shaped loyalty quietly, and the profession has spent decades struggling to prove its connection to revenue. That connection is now being drawn by machines, which read years of accumulated customer voice and recite their conclusions to shoppers at the moment of choice. The leaders who treat each support interaction and each public reply as part of a permanent record, and who measure the narrative forming around their company with the same rigor they bring to satisfaction scores, will run the brands that agents keep recommending. And the service teams who built that record deserve to walk into their next budget conversation knowing that a measurable share of the season’s sales started with them.
About the Author
Lindsey Aliksanyan is the founder and CEO of Metricform, which measures how AI assistants answer shopper questions and what the public conversation says about brands.
Related Articles:

Customer Experience: Moving Away from Pure Self-Service to Welcome the Augmented Advisor

The Power of Reducing Customer Friction

Black Friday Is Won or Lost at the Returns Page, Warns SEKO Logistics

Complaints About Automated Nuisance Calls Nearly Triple – Why a Human Answer Matters

How to Make Automated Customer Experience More Human
