An Onsite AI Assistant Does Four Jobs. Most Teams Buy It for One

An onsite AI shopping assistant is usually bought as a conversion tool and then quietly does three other jobs that nobody budgeted for. The gap between those two facts is where most of the disappointment in this category comes from, and most of the upside.

We have been in a run of evaluation conversations with mid-size retailers this year. The pattern is consistent: the brief arrives as "improve website conversion", and within twenty minutes the same buyer is asking about order tracking, about whether ChatGPT can see their catalog, and about what the assistant is hearing that their analytics is not. Those are four jobs, not one.

Job one: conversion, the one on the invoice

This is the job every vendor sells and the only one they publish numbers for. Rep AI publishes "10-30% Lift in CVR". Envive publishes a "4X AVg Conversion Lift" for visitors who engage with its storefront. Alhena publishes a 20% AOV increase at Victoria Beckham. All captured from their own sites on 23 September 2026.

Read those carefully before you budget against them. Only one denominator in that set is stated, and we took the whole category's numbers apart in the conversion-claims audit. The short version: engaged-visitor conversion and sitewide conversion are different claims, and a lift quoted without the denominator tells you almost nothing about what will happen to your business.

Conversion is a real job. It is just the one where the published evidence is weakest.

Job two: support, before and after the sale

Buyers raise this unprompted, and it is usually the point where the conversation gets specific. Pre-sale is sizing, shipping windows, compatibility, returns policy. Post-sale is order tracking and refund status, which means the assistant needs read access to order state, not just the catalog.

The category is converging here from both directions. Gorgias came from support and now publishes that "1 in 7 (14%) Shopping Assistant conversations ends in an attributed order". The sales-first vendors all ship deflection metrics. The practical consequence for a buyer: ask which direction the vendor came from, because the roadmap and the pricing model still point that way.

Post-sale scope is the question to press on. An assistant that answers "where is my order" needs a live order lookup and an identity check, and that is a different integration from a catalog feed. Several products in this category stop at pre-sale and do not say so on the pricing page.

Job three: the catalog work is shared with off-site agents

This is the one almost nobody sells, and it is the one with the longest shelf life.

To answer "something warm and waterproof under £150 that does not look like hiking gear", an assistant needs attributes: material, waterproof rating, fit, colour family, price, availability at the variant level. Most catalogs do not carry them. The enrichment work to make an onsite assistant useful is substantial and it is one-time, with ongoing sync as products are added.

Here is the part that changes the business case. That is the same data an external agent reads. When a shopper asks ChatGPT, Gemini, Perplexity or Alexa for Shopping what to buy, products whose attributes cannot satisfy a constraint drop out of the answer before price is considered. The enrichment you do for the widget in the corner is the enrichment that decides whether you appear in an answer you will never see.

The surfaces are separate. The work is not. Any evaluation that scores an onsite assistant purely on its own conversion is undercounting it, because a second surface is being paid for at the same time. We set out what the external side actually reads in which product data agents use.

Job four: the demand you cannot currently hear

Site search records queries a shopper already knew how to phrase in your vocabulary. Everything else is invisible.

An assistant records the rest: occasion, recipient, relationship, budget framing, compatibility, constraint. A gifting retailer hears "something for my sister who just moved house, under 3,000". Nothing in a search log looks like that, and nothing in a category tree answers it.

Three things fall out of that log that no other system produces:

Gaps in the catalog. Repeated requests with no good match are a buying-plan input.

Gaps in the attributes. Questions the assistant could not answer are exactly the fields missing from your product data, which loops straight back to job three.

The vocabulary shoppers actually use, which is the input to category naming, filter design and the copy on your product pages.

We would go further: for a catalog under a few thousand SKUs, this log is often worth more in the first quarter than the conversion delta, because it changes decisions rather than moving a rate. That is a claim about where attention goes, not a measured number, and we would rather say it plainly than dress it up as a statistic.

What this means for how you evaluate

Score all four. A vendor that is excellent at conversion and has no post-sale scope, no attribute export and no query log is one product being sold as four.

Concretely, in the demo:

• Ask for the denominator behind every conversion figure, and the attribution window.
• Ask whether post-sale order and refund lookups are in scope, or a later tier.
• Ask whether the enriched attributes are exportable and who holds the semantic index. If you cannot get the enrichment back out, you are renting the work that also decides your visibility in external agents.
• Ask what the query log exposes, and whether unanswered questions are reported as a first-class metric rather than buried.

The four-job frame also settles the sequencing question. If qualified traffic is arriving and not converting, start at job one. If shoppers are never arriving because assistants are naming somebody else, no onsite tool reaches them - that gap is measured by Citation Rank, and House of Zelena is what closing it looked like over six months.

The agentic readiness scan checks whether your catalog can answer a constraint question at all, which is the precondition for jobs one, three and four.

Sources

• Rep AI, Envive and Alhena homepages, captured 23 September 2026.
• Gorgias, "AI shopping assistants are now officially influencing sales", gorgias.com/research, captured 23 September 2026.
• Tru Commerce citation corpus, 26,629 citations, 250 prompts, 5 engines, US/UK/India, captured 16 September 2026.
• Buyer-question patterns are drawn from Tru Commerce evaluation calls in September 2026, reported in aggregate and unattributed.