AI in eCommerce: Start with the friction

A recurring pattern we’re seeing in AI deployments is that eCommerce teams start with the capability before defining the friction. A tool becomes available, a use case gets attached to it, and the team works backwards to identify the customer problem it was meant to solve. The more durable commercial wins we’ve seen tend to come from teams that reverse that sequence.

This article draws on a live discussion at Commerce Live. Watch the full session.

Start with the funnel, not the feature

The underlying motivations behind customer decisions haven’t changed, even if the interfaces and discovery patterns around them are shifting quickly. The question worth anchoring to is the same as it’s always been – where do customers slow down, and what’s making them do it?

Chatbots make the risk of skipping that question visible. In practice, chatbot interactions tend to cluster around verification behavior like delivery windows, compatibility questions, return policies, and suitability concerns. When answers resolve that uncertainty clearly, they reduce evaluation friction. When they fail, they create a confidence problem the customer now has to navigate elsewhere, usually by abandoning the experience or contacting support.

That’s where AI can earn its place, particularly in environments where the volume of behavioral data exceeds what teams can realistically analyze manually. Pattern detection, synthesis across large datasets, and reducing the lag between signal and insight are valuable capabilities, especially across search activity, support themes, and product evaluation patterns. But they need a defined friction point to work towards. Without one, AI can add a layer to the experience instead of improving it.

AI can also introduce operational overhead that teams sometimes underestimate – model costs, maintenance, QA requirements, escalation handling, governance. If the underlying friction isn’t clearly defined, those costs can scale faster than the customer value they create.

Measurement is still catching up, and that matters

Standard web analytics were built for a world where customers arrive, navigate, and convert within owned channels. That model doesn’t fully capture what’s happening as AI mediates more of the journey. Shopify has released new metrics to give teams better visibility into AI-influenced behavior, which is a meaningful step, though the industry hasn’t aligned on standards yet.

AI output is only as reliable as the data feeding it. If events aren’t firing consistently, conversion definitions aren’t aligned across tools, or your data isn’t capturing the full journey, your AI is working from an inaccurate picture – and the decisions made from it will be too. It’s something we work through with clients as a starting point.

Even where that foundation is solid, human review still needs to remain in the loop. This was a point Ali from BigCommerce grounded well at Commerce Live – recommendation engines have always required human override to stay aligned with brand strategy. The algorithm might favor a low-cost accessory based on transaction volume alone, while the brand’s commercial priorities point somewhere else completely. The technology scales the volume of decisions. The team still has to own the criteria behind them. 

It’s also worth being clear about what scaling an AI feature actually proves. Engagement with something new isn’t evidence it’s solving the friction it was meant to address. That requires its own measurement pass, typically a controlled test, not a before-and-after read.

When the journey moves outside your site

The uncomfortable truth worth getting used to now is that some customers will never visit your site, and that includes B2B. They’ll form an impression, compare options, and decide what to do next entirely within an AI experience, whether through an OpenAI-powered search or via an agent shopping on their behalf. For large enterprise brands, this is already an active concern.

At Commerce Live, I was asked what teams can do now to prepare for that shift. At the time, the honest answer was that it was still shaping up. Conversion data from LLMs wasn’t publicly available. Within a month of that conversation, platform-level visibility has started to emerge. But the broader picture is still forming, and what looks like best practice today is likely to evolve quickly.

That said, two things consistently hold up:

1. Getting SEO and structured data as solid as possible, things like content signals, entity clarity, schema markup. These are part of how AI systems represent brands in generated answers, and they’re the same foundations that support organic search.

2. Catalog data enrichment, ensuring product data is complete, accurate, and structured well enough to be useful to AI-driven answer engines, not just traditional search.

The more active step for any team is visibility testing. Querying the AI tools your customers are likely to use, with the questions they’re likely to ask, and comparing the output against what your brand actually offers will surface gaps worth addressing. Where answers are incomplete, outdated, or missing at high-intent moments, that’s where to start. The methodology for doing this systematically is still emerging, but the starting position is clear enough to act on now.

Talk to our growth team. If this resonates with where your team is, book a 30-minute session to explore whether a deeper CRO engagement is the right next step.

Book a session

Lori Cantwell

Senior CRO Strategist

Read Lori’s full bio