AI & Agentic commerce guide: the new customer journey

As a Partner at Converteo and a specialist in AI, Agentic, & Data strategies, Guilhem Bodin helps Marketing and Digital departments transform their operations. He works with Thibauld Vian, a Principal who translates these strategies into data and e-commerce performance roadmaps, and Camille Daudet, a Senior Manager who oversees their deployment through an omnichannel, consumer-centric approach.

Key Takeaways:

• The Three Faces of Agentic Commerce: This isn’t a monolithic block but three distinct trends: onsite agents (deployed by brands on their own ecosystems), operator agents (which “take control” of an interface for the user), and protocol agents (technical infrastructures like OpenAI’s ACP or Google’s UCP, allowing agents to trade among themselves).
• The End of SEO Hegemony, the Dawn of GEO: Web traffic likely peaked in 2025/2026. Visibility is no longer just about the 10 blue links (SEO) but about the synthesized answers from generative engines (Generative Engine Optimization – GEO). In this paradigm, the citation (being mentioned by an authoritative source) supplants the hyperlink, and e-reputation becomes a major battleground.
• The Paradox of the Onsite Agent: Current deployments (Zalando, Boulanger, etc.) show functional advances (cross-category comparison, understanding constraints) but suffer from breakdowns in the user experience (no “add to cart” function, silos between the assistant and the site), often due to internal governance where AI and e-commerce teams are not sufficiently integrated.

We are witnessing a tectonic shift in digital commerce. The emergence of generative AI and agentic systems is not just a technological evolution; it is a paradigm shift that is altering consumer thinking and fragmenting the traditional purchasing journey. How can brands position themselves to avoid massive disintermediation?

The era of answer engines and the rise of GEO (Generative Engine Optimization)

The initial premise is clear: website traffic as we know it has likely reached an all-time high. A Gartner study predicts a structural decline, already visible on news sites across the Atlantic, as users get synthesized answers directly from answer engines (ChatGPT, Gemini, etc.) without needing to click further.

This shift in user behavior from “search” to “conversation” gives rise to a new discipline: GEO. While the fundamentals remain (technical readability, content relevance, reputation), the rules of the game are profoundly altered:

• From Link to Citation: The value of a hyperlink (backlink), a pillar of SEO, is collapsing in favor of the citation. What matters to AI is not that a site points to you, but that it talks about you, and in what terms. The work of digital reputation (PR, influence, review management) becomes more crucial than ever. The example of a telecom operator judged a “poor performer” because of unmoderated Trustpilot reviews, while its “Verified Reviews” are excellent, perfectly illustrates this new challenge.
• From Structuring for the Bot to Richness for Synthesis: AIs do not read; they synthesize. They look for unique data, comparisons, lists, tables, and a rich context of use. The product feed, once a technical tool for Google Shopping, is becoming the backbone of agentic commerce, requiring the integration of previously ignored information: compatibility, testimonials, usage context, and more.

The three dimensions of agentic commerce

The term “agentic commerce” is often overused. In reality, it covers three distinct models that present neither the same opportunities nor the same threats:

1. Onsite Agents: This is the agent the brand deploys on its own site (web or app) to advise and act. Examples from Zalando, which can suggest a complete outfit considering the season and budget, or Boulanger, which compares its entire category, show the potential. Their current limitation is an often “siloed” experience, disconnected from the rest of the navigation and unable to finalize the purchase.
2. Operator Agents: These agents, like Atlas (OpenAI’s browser) or ChatGPT Agent, act as an overlay on the existing web. They “take control of the mouse” for the user, filling out forms or navigating third-party sites. This is a direct threat to the advertising revenue of the visited sites, which explains why a player like Amazon actively blocks these browsers.
3. Protocol Agents: This is the most advanced and structuring form. Technical infrastructures like the Agentic Commerce Protocol (ACP) from OpenAI or the Universal Commerce Protocol (UCP) from Google aim to create a standard so that agents can dialogue and trade directly with each other via APIs. The purchase is made within the agent’s ecosystem (OpenAI, Google) without ever redirecting traffic to the brand’s site. This is the scenario of maximum disintermediation, but also that of a new distribution channel in its own right.

The Amazon Rufus use case: When data trust becomes the weakest link

Amazon’s shopping assistant, Rufus, perfectly illustrates both the immense potential and the current limitations of on-site AI agents. On paper, the tool is a powerhouse. It can guide a consumer through complex queries—such as finding “a good vacuum cleaner”—by maintaining a multi-turn conversation. Rufus refines its selection as new criteria are added: first the highest-rated, then the lightest, then the quietest, all while keeping track of previous constraints. It even goes as far as making a final recommendation, declaring what it considers the “perfect” vacuum that checks every box. It is the promise of ultra-personalized advice, baked directly into the shopping experience.

However, this technical feat hits a brick wall: trust. An agent’s effectiveness is directly tied to the perceived quality of its underlying data. In an ecosystem as vast and open as a global marketplace, the product catalog is often seen as “polluted” by dropshipping offers and customer reviews of questionable reliability. The AI, no matter how high-performing, appears to simply analyze surface-level data (ratings, summarized specs) without the ability to discern substance, nuance, or fraudulent feedback.

Faced with this skepticism, savvy users reflexively turn to external tools like general-purpose LLMs (ChatGPT, Perplexity, etc.). These act as meta-analyzers, capable of cross-referencing a platform’s native info (the Amazon product page) with a wide array of third-party sources: specialized blog posts, comparative tests, and forum discussions. This creates a major paradox: the agent built by the brand for its own ecosystem is deemed less reliable than an outside agent.

Technology alone is not enough. Without impeccable data governance and established trust in the catalog’s quality, the on-site agent risks being little more than a sophisticated gadget. If a user feels compelled to leave the platform to verify information, the journey is broken, and the primary objective—retaining and converting the customer within the ecosystem—is lost. For brands, the real battle will be fought as much over the quality of their data as the performance of their algorithms.

Google vs. OpenAI: The strategic duel

The market is structuring around two opposing strategies:

• OpenAI, the “disruptor” in search of a business model: Starting from scratch, OpenAI is multiplying initiatives to diversify its revenue beyond subscriptions and APIs. The arrival of advertising (a CPM model, for now limited) and sales commission (4% via ACP in partnership with Shopify) shows a desire to become a marketplace in its own right. OpenAI needs to capture behavioral data (hence the interest in a browser like Atlas) to close the gap with Google.
• Google, the giant facing the innovator’s dilemma: Sitting on 20 years of data and a powerful advertising ecosystem (Merchant Center, Google Ads, etc.), Google is proceeding with caution. Its major challenge is not to cannibalize its main advertising revenue while responding to new user behaviors. Innovations like the integration of shopping in AI Mode, the Price Tracker, or Branded AI Agents (allowing users to “chat” with a brand directly from the search results) are gradual responses aimed at keeping the user and the transaction within its ecosystem.

Acting on two fronts

Waiting is not an option. The strategy to adopt must be played out on two fields simultaneously:

1. Externally, master your influence: It is imperative to understand how answer engines perceive your brand. This involves an audit of your presence (which reviews, articles, or forums talk about you?), content optimization for synthesis (GEO), and the semantic enrichment of your product feeds. You must see these platforms as a new distribution channel in their own right, with their own rules.
2. Internally, build the experience of tomorrow: The deployment of an onsite agent should not be a technological gadget but a strategic cornerstone of the customer experience. This requires breaking down organizational silos, defining a clear vision (what role for the agent in 3 years?), and building an evolvable technical architecture (orchestration of specialist agents, connection to existing IS) to avoid creating just another proof-of-concept with no future.

Tomorrow’s commerce will be conversational, personalized, and largely delegated to artificial intelligence. The brands that will succeed are those that start today to treat their reputation as their main asset and their own ecosystem as a laboratory for experiential innovation.