Ferda

It is a real shift and it is early, which is exactly the position where it is cheap to prepare and expensive to be told you should have.

What is changing: people increasingly ask an assistant to find and compare products rather than browsing results themselves. Whatever the assistant can read determines whether you are in the comparison. If it can read your competitor's catalogue and not yours, you are not losing a ranking position, you are absent.

What you should do is mostly things you already know how to do, which is the good news:

Structured data on product pages. Proper product markup with name, description, price, currency, availability, condition, identifiers and review data. This is long-standing, well-documented, and it is what a machine reads when it lands on your page.

A clean product feed. You have one for advertising. The same discipline — accurate availability, real identifiers, sensible categories — is what makes a catalogue machine-readable anywhere.

Let the crawlers in. Check your robots rules and your bot protection. A shop that blocks unknown agents is invisible to them by configuration.

On the newer protocol layer, and this is where I would be measured rather than enthusiastic.

There are emerging specifications for exposing a catalogue to agents directly, and for letting an agent complete a purchase, rather than having it scrape a page. Several large platforms and several standards efforts are moving in this area at once, and they do not yet agree.

My read: the structured data and the feed are safe investments — they are useful today for ordinary search, for advertising and for any agent that reads pages, and they will not be wasted whatever wins.

The protocol-specific integrations are a bet. If your platform ships support as a setting, turn it on; it costs nothing. Paying for a bespoke integration with one specification this early is a bet on which one survives, and I would want to see my own traffic justify it first.

One habit worth adopting whichever route you take: when you correct an assistant about something for the second time, write it into the file immediately rather than after the session.

The corrections you make repeatedly are exactly the content that is worth keeping, and they are the ones you forget existed once the tool has been trained out of the behaviour. Capturing them at the moment of irritation is the only reliable time.