"Introduced AI" is a phrase behind which everything is hidden: from a chatbot on five prepared scenarios to a restructuring of the way the store stores product data. Between these two things, months of work and completely different consequences. Let's analyze in parts what the models really change in the online store, under what conditions they are generally turned on and how to check the readiness of your own site without hiring anyone for this.
"AI in the store" is three different projects
Confusion begins with the fact that incompatible things are called one word.
Assistant in dialogue
A chatbot or operator assistant that answers questions about availability, delivery, compatibility. Here, the language model does exactly what it does: it understands the person's phrasing and retrieves the answer from your database. If there is no base, the model will start to invent - and the worst thing is that the fiction will sound confident, and the client will discover it, not you.
Ranking and recommendations
"Similar goods" block, issue sorting, personal selection on the main page. This is not a language model, but statistics on your data: browsing history, shared purchases, card attributes. The quality is determined not by the algorithm, but by how neatly the names, categories and characteristics are filled.
Visibility in other people's AI responses
The newest part. A person asks the assistant where to buy a specific thing, and the assistant names the shops. You were either mentioned there or not. It's no longer about the AI inside your site, but whether an external agent can read your storefront.
Three different projects with different budgets and deadlines. When a contractor sells "turnkey AI" as a single package, it makes sense to ask which of the three.
Data first, models second
In late July - early August 2026, we captured technical parameters from more than thirty live sites in our portfolio, most of them online stores. One of the most frequent defects lies neither in design nor in speed.
A store of professional cosmetics: a catalog of 5,561 products, works, sells — and no counter at the main office. No GTM, no gtag, no pixel. Conversions are not measured in any way. The second case is a B2B catalog of spare parts for sewing machines, where Universal Analytics is still connected: Google turned it off on July 1, 2023, that is, analytics has been dead for three years, but looks alive in the code. The third, the most subtle: the clinic's inline GTM snippet is self-blocking, because the exit condition checks the container identifier for equality with the GA4 identifier and always works. There is a script on the page, the container does not load.
Now imagine a recommendation system on such a site. It receives an empty behavior stream and emits a random set of items. Checking whether it has gotten better is also nothing.
The mechanism here is simple and unpleasant. Analytics doesn't make money by itself — it makes everything else verifiable. Without correct conversion events, the advertising office has nothing to optimize, and the recommendation model has nothing to learn from.
