An AI Product Recommendation Agent That Knows Your Catalogue

A shopper types “trouble sleeping, already on blood pressure tablets, what should I try?” into your site search and gets zero results, or forty. An AI product recommendation agent answers that question the way a good shop assistant would: one product from your live catalogue, a reason tied to what the product actually contains, and the caution that applies to someone on medication. It picks only from what you sell, and it says nothing your product data does not back up.

We built one for a consumer health-products brand in our own group. The model was the easy part. Most of the work went into rules: what it asks before it recommends, what it will not say, and what happens when the catalogue changes underneath it. That is also where the risk sits, so that is where this post spends its time.

Why site search fails the questions shoppers actually ask

Most store search matches words. Shoppers describe situations. The Baymard Institute’s e-commerce search benchmark, updated 29 April 2026 from more than 10,000 performance ratings across 170+ leading sites, found that 56% of sites fail to adequately support users’ search needs.

The breakdown by question type is the useful part. Searches by use case failed on 43% of sites. Compatibility searches (“does this work with…”) failed on 44%. Symptom searches (“something for dry skin”) failed on 37%. Those are the exact questions a product advisor exists to answer.

Baymard reviews leading US and European retailers. Our inference, not theirs: a small store with a default search box and thin product pages does worse.

The channel keeps growing regardless. The US Census Bureau put US online sales at $340.2 billion in the second quarter of 2026, 17.1% of all retail, up 12.2% on a year earlier (released 18 August 2026). More of the buying happens on pages where nobody is standing behind the counter.

Shoppers already ask AI which product to buy

Adobe Analytics measured traffic to US retail sites from generative AI tools rising 693.4% year on year over the 2025 holiday season, 1 November to 31 December, across more than a trillion site visits. That growth comes off a small base, and Adobe did not publish AI’s share of total traffic. Read it as direction, not size.

Salesforce, which sells shopping agents, estimates that AI and agents drove 20% of global holiday retail sales in 2025, $262 billion, modelled from data on 1.5 billion shoppers (published 8 January 2026). The release never defines “drove”, so treat it as a vendor’s upper bound.

The direction is still plain. Shoppers are getting used to asking a machine what to buy. If that question gets answered on your site, the answer comes from your catalogue and your cautions. If it gets answered somewhere else, you have no say in it.

What the agent does, step by step

Ours is a tomte, our word for one production agent with one defined job. This one’s job is matching a shopper’s need to one product in the catalogue. Here is a single conversation, described without the plumbing.

1. Sort the message. Every message lands in one of five buckets: a concern that needs working out, a direct question about a product, a store FAQ, someone ready to buy, or off-topic chat. Only the first gets the full treatment. The others get a short answer or a link.
2. Ask as little as possible, but ask the question that matters. If the concern is clear and low-risk, the agent asks at most one safety question (medication, pregnancy, a diagnosed condition) and then recommends. If something sounds serious, it asks more before going anywhere near a product. A shopper who wants to skip the questions can say so.
3. Search the catalogue by meaning. Candidates come from the live catalogue, matched on what each product does, so “can’t switch off at night” can find a sleep product whose page never uses those words.
4. Check, then ground. The shopper’s stated constraints are checked against each product’s documented cautions before anything is shown. The pitch has to come from the product’s listed ingredients and benefits, never from its name or its marketing copy, and any caution on file gets stated.
5. Answer in a fixed shape. One product, linked, with the price, why it fits, how and when to use it, and who should not take it.
6. Close with the safety line. Every recommendation ends by telling anyone with a diagnosed condition, a pregnancy, or a prescription to check with a doctor first.

Step 4 carries the most weight. A recommendation grounded in a product’s name is a guess wearing a lab coat.

The rules it runs under

The advisor is set up as a wellness guide, not a doctor, and it is told never to diagnose. A fixed list of red flags changes its behaviour: symptoms that have lasted weeks, severe or chest pain, trouble breathing, pregnancy, children under 12, prescription medication, a diagnosed condition. Any of those means more questions and a plain referral to a doctor before any sale.

It cannot recommend anything outside the catalogue, and it may not invent a benefit or a caution that is missing from the product data. When the data is thin, the answer is thin. We think that is the correct trade.

One honest gap: ours has no live handoff to a support person. Its only escalation is the doctor referral, which suits a health catalogue, since a support rep is the wrong person to ask about medication anyway. For most other catalogues the handoff should go to your support inbox with the conversation attached, the way our internal knowledge assistant passes what it cannot answer to a named person.

Regulators care about the gap between what an AI says and what anyone checked. In February 2025 the US Federal Trade Commission finalised its order against DoNotPay, with $193,000 in monetary relief, after charging that the company never tested whether its “AI lawyer” performed at the level of a human lawyer. That case was about marketing claims, not a chatbot’s answers. The principle transfers anyway: if your AI tells customers something, expect to be asked how you know it holds.

When the catalogue changes underneath it

A product advisor goes stale the week something is discontinued. Ours re-syncs from the live store every week: new products get indexed, changed ones updated, removed ones switched off.

The sync has one guard we would put in any build. If the product count coming back is abnormally low, the job treats the pull as broken and switches nothing off. A bad sync cannot empty the catalogue overnight.

Failure behaviour follows the same logic. If the search-by-meaning layer is down, the agent falls back to a plain catalogue lookup instead of an error. A shopper sending messages too fast gets a polite “slow down” message and the chat stays open. If saving the conversation history fails, the failure is logged and the reply still goes out.

Memory is opt-in. The agent can remember what a shopper told it across visits (never what it inferred), but only after asking once, and a shopper can withdraw that consent mid-conversation with immediate effect.

Site search, a generic chatbot, or a catalogue advisor

	Site search	Generic chatbot	Catalogue advisor
”Something for…” questions	Often fails (37% to 44% of leading sites, Baymard 2026)	Answers fluently	Answers from your products
Where answers come from	Keyword match on product text	The model’s general knowledge	Your product data only
Can suggest what you do not sell	No	Yes	No
Safety questions first	None	Only if prompted	Built in, with red-flag rules
Stays current	Automatically	Not without work	Weekly sync with a guard
Who owns it	Your platform’s feature	A vendor subscription	You: code, data and prompts in your accounts

What changed

We do not publish usage numbers for this one, because none have been signed off. What changed is easier to describe than to count. The same question gets the same careful answer at 3 a.m. as at 3 p.m. The safety question gets asked every time, including the times a tired person would skip it. And because conversations are stored, the questions shoppers actually ask became readable, which shows exactly where the product pages are thin.

Anything customer-facing we install gets the discipline we apply to our own agent fleet: registered before it runs, every action logged, one fence per agent. That fleet counted 41 registered agents, 9 of them live in production, as of 30 June 2026.

A caveat against ourselves. In our ranking of what to automate first, a customer-facing bot sits in the “wait” column, and we still stand by that. Do the boring back-office jobs before this one, because the guardrails above take longer to build than the features.

What it costs to install

A catalogue advisor is a standard AI Workforce Sprint: from $9,500 fixed, four to six weeks, built on your store, your accounts and your catalogue, then handed over with documentation.

Two things move the price. The first is product data. The advisor can only ground a recommendation in what your product records actually say, so if ingredients, specs or compatibility live in a PDF or in someone’s head, getting them into structured data is part of the scope. The second is integrations: your store platform, stock levels, and whether returning customers sign in.

The written scope fixes the price before you commit. Code, prompts and conversation data stay in your accounts, and you can revoke our access whenever you like.

If your inbox keeps filling with “which one should I get?” messages, send the store link and the five questions customers ask most. Start async and a written answer comes back within one business day, with no meeting at any stage.