Strategic Consequences for Commerce Leaders in the Age of Agentic Commerce

The preceding six articles in this series have described the structural shift in the digital commerce decision journey. The classical funnel loses explanatory power, stable KPIs increasingly mask opacity, a growing share of the purchasing decision forms outside measurable systems, classical attribution reaches its technical limits, new metrics must abandon the claim to certainty, and the system itself shifts its role from observation instrument to decision architecture.

What follows is an operational question, no further diagnosis. What do these findings mean in the daily work of a retailer who must take decisions today on investments, roadmaps and organisational design?

This closing piece of the series translates the strategic consequences into operational measures. It is deliberately not framed as a universal solution. The order of recommendations follows the structure of the series, because the consequences derive from the findings themselves; no abstract maturity model has been imposed.

1. From channel to plug-in capability

The first finding of the series concerned the decay of the funnel. The consequence is to shift the steering logic, well beyond any attempt to repair the funnel itself. As long as marketing and commerce organisations think in channels, they keep optimising something that plays an ever smaller role in how users actually arrive at decisions. The relevant question today is under what conditions a company's offers become part of a machine pre-selection in the first place. The route into a particular channel's funnel matters less than it once did.

In operational terms, this requires a controlled break with three deeply held assumptions. The first assumption is that visibility equals reach. In agentic systems, visibility is replaced by structural plug-in capability. The second assumption is that every additional platform represents another channel. In reality, each platform defines its own set of requirements for data structure, identifiers and interfaces, which need to be served selectively. The third assumption is that campaign optimisation is the lever that produces growth. Today, growth is increasingly created in a layer that sits well upstream of the campaign.

Concretely, this means marketing teams must begin to shift part of their attention away from channel budgets and towards data plug-in. A first concrete step is to take stock of which platforms today already access content agentically. These include at minimum the Google Shopping Graph with over 50 billion listings as the backbone of AI Overviews and Gemini, the OpenAI Merchant Program with the Agentic Commerce Protocol specification,¹ the Perplexity Merchant Program,² and the Shopify Catalog MCP interface.³ A second measure is a deliberate decision on which of these platforms is genuinely decision-relevant for the relevant category, and which are knowingly not served because their operational relevance does not justify the complexity.

This decision can be prepared today through three reproducible methods, each offering a different angle on platform relevance.

First, a systematic mapping of AI referral traffic. In GA4, a dedicated channel called "AI Assistants" can be configured, grouping hostnames such as chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com and claude.ai as referrer sources.¹⁴ Since many AI platforms do not send standard-compliant referrer headers, and mobile app traffic frequently lands in the direct category, this mapping should be supplemented by a quarterly assessment of so-called dark traffic on product detail pages.¹⁰ A meaningful shift in the share of direct visits with clear product intent on categories that are frequently recommended in agentic systems is a robust indicator that a given platform genuinely produces volume for the assortment in question.

Second, a category-specific Share of Model test. This involves defining a set of 30 to 50 typical purchase-intent prompts for one's own category and querying the three to five most relevant LLM systems on a regular cadence. The measurement captures how often one's own brand and the relevant competitors are mentioned, and in what position within the response.¹⁵ The methodology aligns with established 2026 industry benchmarks, in which the average Share of Model for category leaders sits at around 31 per cent and the top three brands capture roughly 67 per cent of all mentions.¹⁶ Platforms in which one's own category is structurally underrepresented, for instance because they handle few product-near queries, can be deselected early on this basis.

Third, a controlled conversion and demand check. Platforms that show relevance in the first two steps should be tagged with unambiguous UTM parameters on the relevant entry points over a defined window, typically six to twelve weeks, and compared with the remaining channels on conversion rate, average order value and repeat purchase rate.¹⁷ This data makes visible whether agentic traffic in the relevant category acts as a supporting research channel or as a closing purchase channel, which in turn determines how deeply the integration warrants investment.¹⁸

2. The data base as decision foundation

The second and third findings of the series concerned the illusion of measurability and the invisible phase of the purchasing decision. The operational consequence lies in restructuring the product data base. Adding further tracking systems offers little in this situation. If a growing share of the decision forms outside measurable systems, then the lever shifts to the question of how cleanly one's own offerings can flow into those external decision processes in the first place.

The specifications that matter here are by now documented. The OpenAI Agentic Commerce Protocol Feed Specification defines its own required and optional schema for fields such as identifiers, prices, availability, shipping conditions, return options and contextual attributes. It is expected as a complete snapshot feed with recommended daily delivery.¹ The Klarna Agentic Product Protocol specification, introduced in December 2025, describes a comparable approach for around 100 million products across the Klarna network.⁴ At the semantic level, JSON-LD per Schema.org has emerged as the de facto standard, with a roughly 89 per cent share among deployed structured data formats and a 3.1-fold higher likelihood of citation in Google AI Overviews.⁵

Three operational requirements follow. First, every item in the assortment needs a stable, unambiguous GTIN or MPN. Products without these identifiers are systematically excluded from cross-platform matching, because agentic systems depend on being able to align identical items unambiguously across multiple sources.⁵ Second, price, stock and shipping information must be kept consistent across all internal data sources. A discrepancy between JSON-LD markup on the detail page and the merchant feed is enough to lower an agent's trust score and remove the product from recommendation sets.⁵ Third, product descriptions need a semantic enrichment that goes beyond classical SEO optimisation. An AscentCore study from March 2026 shows that with purely attribute-oriented descriptions, agents pick a suboptimal product in 52.6 per cent of cases, because a single ambiguous sentence is enough to discard a suitable option.⁶

In practice, this means running a PIM or product data audit against machine interpretability, beyond the usual measure of field completeness. The deciding question is whether an agent can interpret a field correctly without human mediation. A filled field on its own does not answer that question.

Three documented cases from European retail illustrate what machine interpretability looks like in operational terms. Zalando set up an internal tool called Content Creation Copilot as early as 2024, which uses large language models to extract structured product attributes directly from supplier raw data, images and free text. Machine interpretability is established at the entry point of the onboarding pipeline, well before an item ever surfaces in the assortment.⁷ Otto has described in its in-house tech blog how classical product search struggled with spelling variants and fuzzy queries and was replaced by a semantic representation in which items are organised by their intended meaning. Products are mapped through a vector representation to the problems they are meant to solve, so that even loosely worded queries return sensible matches.¹⁹ IKEA, in turn, operates a knowledge graph in which differentiating attributes such as function, material and room context are modelled explicitly, with recommendations emerging along explainable paths. The KGC presentation from 2023 sets out in detail how differentiators were lifted out of marketing copy into a machine-readable structure, so that recommendations remain transparent to external systems.²⁰

Over the next two years, this kind of investment moves from an optimisation measure to an infrastructural prerequisite. It determines whether one's own product data base serves as a starting point for agentic recommendations or remains structurally invisible to machine interpretation.

3. Selective plug-in capability over full integration

From the findings on attribution and on new decision metrics, a further strategic shift emerges, namely the deliberate refusal of comprehensive integration in favour of selective plug-in capability. The reflex reaction in many organisations to a growing number of agentic systems is to want to serve as many platforms, protocols and interfaces as possible at once. This strategy generates operational load while struggling to demonstrate its relevance.

The defensible alternative is an architecture that works from a unified base layer and adds specific integrations on top. The base layer consists of three components that are already stably documented. First, complete and cleanly maintained Schema.org/Product markup on every product detail page in JSON-LD, including price, availability, shipping conditions and review data.⁵ Second, a consistent product feed in the Google Merchant Center format as a lingua franca, which numerous agentic systems can process either directly or with minimal adjustments.⁸ Third, an llms.txt at the site root that gives AI crawlers a curated overview of the content relevant to them and thereby provides an additional signal without replacing the robots.txt mechanism.⁹

On this base, specific platforms can be plugged in selectively. Anyone selling into the US market for fashion-relevant assortments will prioritise the OpenAI Merchant Program, because ChatGPT is already gaining transactional weight there.¹ Anyone working in the DACH region with high-ticket or advice-intensive assortments should additionally examine Perplexity, because the search-to-purchase distance is structurally shorter there.² Anyone running on Shopify or a Shopify-Catalog-compatible stack can use the Catalog MCP Server to feed directly into the discovery layer of agents.³

A useful rule of thumb is to tie every new platform integration to a pre-defined threshold, for example a measurable share of agentic traffic in one's own category or a documented direct referral from the system in question. Platforms that do not meet this threshold remain on a watch list without immediately triggering integration effort.

4. From attribution to indication

The fourth and fifth findings of the series concerned the loss of attribution and the need for new decision metrics. The operational consequence is to shift the function of classical attribution models, well beyond the usual reflex of refining them further. They continue to provide useful signals about parts of the customer journey, while losing their role as the central steering benchmark.

In their place comes a set of indicators that point to shifts in decision formation without fully mapping them. Several studies from 2025 and early 2026 document three phenomena that should now be visible in any GA4 or Adobe Analytics setup. First, the share of so-called dark traffic is rising, that is unattributable direct visits, because many AI referrers such as ChatGPT, Claude or Gemini land in GA4 without a correct referrer and are automatically assigned to the direct category.¹⁰ Second, click-through rates on classical organic results are measurably falling in queries where Google AI Overviews are shown, in several agency-wide analyses.¹¹ Third, conversion rates rise on direct traffic with clear product intent, which can be read as an indirect signal of agentically pre-filtered traffic.¹⁰

Three concrete measures can be put in place today. First, the share of dark traffic should be reported systematically every quarter, broken down by sub-channels so that shifts become visible early. Second, it is worth building a Share of Model indicator, that is a regular survey of how often one's own products are mentioned in the three to five most relevant LLM systems for one's own core queries. The methodological setup that makes such a survey statistically defensible is described at length in section 6.¹² Third, internal reports should explicitly distinguish between measurable attribution and interpretive indication, in order to lower the expectation of certainty in places where it can no longer be met.

5. The interface as a product

With agentic protocols, the definition of what counts as the externally visible product of an e-commerce business also shifts. Until now, that product in the narrow sense was the online shop, supplemented by app, marketplace presences and where applicable proprietary brand touchpoints. In an agentic architecture, an additional layer comes into play, namely the interface through which product data and transactional capabilities are made available to external systems. The piece on the shift from tracking to decision architecture and the guest post on headless-oriented architectures supply the structural backdrop.

In December 2025, Klarna introduced the Agentic Product Protocol as an open standard that addresses precisely this interface layer and makes more than 100 million products from the Klarna network accessible to agents.⁴ Shopify has implemented a comparable approach with its Catalog MCP Server for around 5.6 million shops, which are turned into Agentic Storefronts by default and thereby appear in ChatGPT, Gemini and Microsoft Copilot without further configuration.³ Alongside this, the OpenAI Agentic Commerce Protocol specification describes the delivery path through which merchants hand over their feeds to OpenAI.¹

For retailers that do not run on Shopify, two practical consequences follow. First, it is worth tracking the dominant protocol standards systematically. UCP, ACP and MCP are emerging as the most relevant specifications today, and their convergence or divergence will largely determine the further investment requirements over the next twelve months. Second, the in-house product data architecture should be designed so that it can be exported into several of these protocols without fundamental rebuilds. A PIM optimised solely for one's own website becomes the bottleneck in an agentic environment.

6. The role of the brand in agentic discovery

A consequence of agentic discovery that is often underestimated concerns the role of the brand. When a growing share of pre-selection takes place outside one's own sphere, the function of brand communication shifts. It increasingly works on the picture an agent reconstructs of a brand from publicly available sources, with the direct effect on the buyer playing a smaller role. That reconstruction emerges from a combination of structured product data, editorial sources, reviews and semantic signals on the brand's own website. How visibility is reorganising beyond classical search engines is explored in our piece on Generative Engine Optimization.

In July 2025, Mango introduced the Mango Stylist, a conversational interface deployed in its own shop for styling advice, which simultaneously serves as a controlled data point for external agents.¹³ The logic aims at a brand-consistent layer of interpretation that continues to play a role even when the actual advice happens in a third-party system. Adding yet another chatbot was, by all accounts, beside the point.

Three operational guidelines follow. First, every brand should regularly check how it is described in the three to five most relevant LLM systems, because these descriptions carry a growing share of the initial perception. Second, review and reputation data should be treated as a strategic data source in their own right, well beyond their familiar role as a marketing measure, because agents read them in a structured way and incorporate them into recommendation logic. Third, editorial content on the brand's own website, such as guides or background pages, should be designed in a form that is agentically readable, that is, clearly structured, semantically unambiguous and clearly tied to product contexts.

The first of these guidelines warrants a brief methodological detour. In practice it is frequently underestimated and just as frequently pursued with the wrong tools, which leads to misreadings. A sufficiently documented methodology now exists to move the question out of the realm of intuition and into the range of reproducible measurement.

The starting point is a statistical property of generative answer systems. An identical query produces different responses on repeated runs in ChatGPT, Perplexity, Claude or Gemini, because model temperature, retrieval timing and internal confidence thresholds all vary. A SearchAtlas analysis from summer 2025 measured monthly citation drift and found a share of 40.5 per cent new or changed sources on Perplexity and 59.3 per cent on Google AI Overviews.²¹ Under these conditions, a single measurement is statistically meaningless. Its margin of error is wider than the signal it attempts to capture.

What follows is a methodology shaped along the lines of political polling. At least thirty sampling runs are executed per prompt and per platform, ideally spread across seven to fourteen days to capture temporal variation. The results yield a point estimate for Share of Model and a confidence interval, which at thirty runs typically spans around sixteen percentage points and shrinks to roughly nine percentage points at one hundred runs. Reporting the confidence interval is what makes the measurement interpretable in the first place, because it stops the familiar conflation of short-term noise with structural shift.²²

The platforms differ from one another so structurally that a blended average leads astray. A Qwairy analysis of 118,000 AI answers in the third quarter of 2025 shows that Perplexity cites on average 21.87 sources per answer, against 7.92 for ChatGPT. Domain overlap between the two sits at only eleven per cent. ChatGPT draws 47.9 per cent of its citations from Wikipedia, Perplexity 46.7 per cent from Reddit.²³ A single cross-platform headline figure therefore conceals precisely what would matter operationally. Each platform defines its own competitive arena, with its own source logic and its own pattern of competitive displacement.

In practice, a repeatable audit procedure can be derived from this, one that works without specialised tooling. Several independent sources from spring 2026 describe the steps in comparable form. First, a category-relevant prompt set is curated, typically thirty to fifty prompts grouped into awareness, consideration and buying-intent queries. Then three to five models are selected, in the EU context most often ChatGPT, Perplexity, Claude, Gemini and Copilot, each queried at least thirty times. The responses are coded for mention, position, sentiment and cited sources, after which the principal competitors are measured against the same set. The exercise is repeated on a rolling monthly basis, with thirty-day moving averages and two-sigma anomaly thresholds, in order to separate genuine shifts from statistical noise.²⁴

Specialised vendors have productised this approach. Profound, with a Series B round of 96 million US dollars in February 2026, has worked the method through B2B discovery in detail, complemented by a joint study with Bain on the role of generative systems in the buyer intent channel.²⁵ Comparable tools come from Peec.ai, Otterly.AI and Promptwatch, each with different emphases between mention tracking, citation analysis and prompt set management.²⁶ For mid-market organisations, a staged entry tends to make sense. In the first two months, a manual exercise with twenty to thirty critical prompts is enough to validate the methodology. Only after that does it pay off to automate gradually via the model APIs, with typical costs running between 90 and 450 US dollars per month for a setup of one hundred prompts across three platforms.

This methodology remains an indicator, no kind of proof. It shows in what light a brand appears in machine pre-selections and how stable that picture is across platforms, time and competitive set. It replaces neither reach measurement nor conversion analysis; it adds to both a layer that is simply missing from most current reporting.

7. Organisation and decision logic

Perhaps the most far-reaching consequence of the entire series concerns organisation more than technology. When a growing share of the decision logic takes place outside direct control, the internal task of companies shifts. It is less about refining processes further and more about the ability to take consistent decisions under uncertainty.

In practice, this calls for a shift in the division of labour. Marketing, e-commerce, IT and product data management need to work more closely together, because the levers for visibility, conversion and margin quality no longer sit in any single function. How that collaboration is anchored organisationally depends heavily on the size and maturity of the organisation. The documented practice points to three distinct models, each suited to a different context.

By far the most common model in mid-market organisations is a hybrid one. A Gartner survey of more than 1,800 executive leaders found that 55 per cent of organisations now have some form of AI oversight body in place, typically staffed with existing leaders from operations, IT, legal, finance and compliance who take on the task as an additional role rather than as a newly created full-time position.²⁷ A March 2026 analysis of mid-market AI governance (1,000 to 10,000 employees) sets out the pattern explicitly: mid-market companies do not, as a rule, have the resources for a Chief AI Officer and a dedicated AI staff function, and instead distribute the responsibility across existing leadership layers, with a Governance Lead from operations or the COO's office acting as owner and Initiative Owners drawn from the affected business units.²⁸ An assessment of AI team structures in retail shows that 52 per cent of all retailers, and 63 per cent of retailers with annual revenues above 500 million US dollars, prefer this hybrid model, in which a small core handling strategy and governance works alongside specialists embedded in the business units.²⁹

Pure centres of excellence with dedicated, newly hired headcount tend to appear at very large organisations. Walmart has established such an AI Centre of Excellence with its own staff structure and direct executive sponsorship, documented as a template for groups with several thousand to several hundred thousand employees.³⁰ For mid-market contexts, this is not the usual pattern in the documented cases. Embedded models, in which AI specialists report directly into the business units, in turn only become viable above a certain scale and AI maturity, which mid-market structures rarely meet.

For the task described here, the agentic plug-in capability, the picture becomes clear. In the mid-market, the answer is not, as a rule, two to four newly created full-time roles, but a small cross-functional working group of typically two to four people drawn from existing functions, who carry this responsibility as an additional remit, with an estimated time share of between ten and twenty per cent of their regular working time. An internal owner from e-commerce or operations takes on the coordination, supported by named contacts from IT, product data management and marketing. A dedicated, newly hired position only becomes defensible once the documented share of agentic traffic in one's own category crosses a threshold at which the coordination cost exceeds the scope of a part-time remit. In most mid-market organisations, that threshold will only become relevant in the next twelve to twenty-four months, if at all.

The working group is permanently responsible for three tasks. It maintains a current overview of relevant platforms and protocols. It monitors the quality of the company's own product data against the most important external specifications. And it supplies a quarterly report on the indicators mentioned above.

This unit replaces no existing function. It supplements the classical channel and performance work with a layer that is structurally under-resourced in most organisations today and thereby becomes the quiet bottleneck, without this being immediately visible.

8. A realistic investment logic

A final consequence concerns investment logic. The temptation in a phase of structural shift is to invest either too much too early or too little too late. Both extremes generate avoidable cost.

The defensible alternative is a staged logic anchored in three quantities. First, the share of agentic traffic in one's own category, measured through the indicators outlined above. Second, the documented volume via the platform in question, for example through direct referrals, conversion data or qualitative signals from customer service and sales. Third, the maturity of the protocol concerned, because investments in not yet stabilised specifications carry a higher adjustment risk.

A pragmatic rule of thumb is to shift between five and ten per cent of the existing marketing and e-commerce budget into agentic plug-in capability over the next twelve months, with a focus on product data quality, interface architecture and internal capability building. In most mid-market contexts, this order of magnitude is enough to build a robust base without investing prematurely in full integration of every system. It is at the same time high enough to produce a measurable effect and low enough to keep misallocations contained.

Closing: decision as an architectural problem

The central consequence of this series is that decisions in digital commerce can no longer simply be taken; their conditions must be designed. The point lies in the structures under which decisions form, with the individual purchasing decision itself a secondary matter.

Decision-making capability thereby becomes an architectural question. It concerns data models, interfaces, organisation and investment logic in equal measure. Companies that take this perspective build structures that remain capable of action even under reduced visibility. Trying to recover lost transparency leads, in our experience, into a cul-de-sac.

Those who hold on to the idea of fully measuring and controlling decision processes will increasingly operate at the edges of a model that is already dissolving. The strategic task for the next twelve to twenty-four months is to develop a viable architecture for the time after this model's erosion.

This series therefore ends less with a recommendation than with a shift in frame of reference. Anyone who continues to think of digital commerce as a funnel will be left explaining phenomena the funnel no longer captures. Anyone who understands it as an architecture can begin to be effective at the few points where the organisation still exerts immediate influence today.

This article is the closing part of the Decision Journey Analytics series, examining how purchasing decisions are restructured beyond the classical funnel.

Glossary

Agentic Commerce Protocol (ACP) A specification maintained by OpenAI for the structured handover of product data to ChatGPT-based shopping features. Defines required and optional fields of a product feed, its delivery path and permitted data types.

Catalog MCP Server Shopify's interface through which AI agents can search the Shopify catalogue and retrieve product data. Follows the Model Context Protocol and is being evolved towards the UCP specification.

Universal Commerce Protocol (UCP) An emerging vendor-neutral specification for interaction between agents and commerce systems, covering discovery, checkout and payment functions.

llms.txt A text file placed at the site root that gives AI systems a curated overview of the content relevant to them. Complements robots.txt without replacing it.

Share of Model An indicator metric that measures how often one's own products or one's own brand are mentioned across a defined set of LLM systems for one's own core queries. Complements classical visibility metrics.

Dark Traffic The share of visitors arriving in web analytics systems without a correct referrer and recorded as direct visits by default. Grows structurally with agentic systems, because many AI referrers do not send standard-compliant referrer headers.

Agentic Storefront A storefront configuration in which a shop is present by default in agentic discovery systems such as ChatGPT, Gemini or Microsoft Copilot, without the need for a separate integration.

Sources

¹ OpenAI Developers: "Agentic Commerce, Products Feed Specification" and "Get Started Guide," 2025/2026. Available at developers.openai.com/commerce.

² WebFX: "The 5-Minute Guide to Perplexity's Merchant Program," 2025. Hashmeta: "Perplexity Merchant Program: How to Get Your Products Featured in AI Search Results."

³ Shopify Developers: "Catalog MCP Server," 2025/2026. Shopify Enterprise Blog: "Agentic-Ready Product Data: How to Get It & the Cost of Inaction," 2026.

⁴ Klarna Investor Relations: "Klarna Launches Agentic Product Protocol: The Open Standard That Makes 100M+ Products Instantly Discoverable by AI Agents," 15 December 2025.

⁵ Alhena: "Schema Markup and AI Search for E-Commerce," 2026. Stellagent: "Agent-Ready Product Data," April 2026.

⁶ AscentCore: "The Shift to Business-to-Agent (B2A) Commerce: Why Your Product Descriptions Are Now Your Most Critical Sales Asset," Research Report, March 2026.

⁷ Zalando Engineering Blog: "Content Creation Copilot, AI-assisted product onboarding," September 2024.

⁸ Skuanalyzer: "Is Your Product Feed Ready for Agentic Commerce? A Merchant Center UCP Checklist," 2025.

⁹ Aurora Commerce: "Aurora Commerce adds support for the LLMs.txt standard," 2025. AgentReadyHQ: "How to Implement LLMs.txt for E-Commerce," 2025/2026.

¹⁰ Loamly: "What Google Analytics Gets Wrong About AI Traffic," November 2025. Out Of The Blue: "Vanishing Clicks: Why 15 to 40 percent of Your Paid Traffic Is Going Dark in GA4," June 2025.

¹¹ SiteGrade: "Google AI Overviews 2026: Real Organic Traffic Impact," 2026. The Digital Bloom: "Organic Traffic Crisis Report, 2026 Update."

¹² Presence AI: "2026 GEO Benchmarks: AI Search Traffic Up 527% While Traditional Organic Drops 40%," 2026.

¹³ Digital Commerce 360: "Mango launches new Stylist assistant with conversational AI," 8 July 2025. Mango Fashion Group Pressroom: "Mango launches Mango Stylist," 2 July 2025.

¹⁴ Authority Tech: "AI Traffic Attribution: How to Track Visitors from ChatGPT, Perplexity & Gemini," January 2026. Surfacelocal: "Track ChatGPT Traffic: UTM Parameters Guide," 2026.

¹⁵ Aether AI: "Share of Model Benchmarking: How to Measure Your AI Visibility Against Competitors," 2026. CiteCompass: "Share of Model: The AI Era's Answer to Market Share."

¹⁶ KnewSearch: "2026 AI Search Visibility Benchmark Report." Loamly: "AI Visibility Benchmark 2026: 2,014 Companies Across ChatGPT, Claude, Gemini, and Perplexity," 25 January 2026.

¹⁷ AgentSurge: "How to Track ChatGPT and AI Referral Traffic All the Way to Revenue," 21 February 2026. Hikmah AI: "How to Track AI Traffic in Google Analytics 4, 2026 Setup Guide."

¹⁸ Wislr: "Your AI Channel Is a Consumer Research Engine, Assisting with the Final Sale." Smartersorting: "How We Tested 5 AI Shopping Assistants Across 2,500 Interactions," 28 January 2026.

¹⁹ Otto Tech Blog: "Rethinking semantic product search: a new approach to queries with unusual spelling variations," 2024/2025. Available at otto.de/jobs/en/technology/techblog.

²⁰ Metaphacts: "Building explainable and trustworthy recommendation systems: What we learned from IKEA at KGC 2023," 2023. Knowledge Graph Conference 2023, IKEA presentation on product knowledge graph and explainable recommendations.

²¹ SearchAtlas: "Why Do AI Search Results Keep Changing? Citation Drift Across Platforms, June to July 2025," 2025. Available at searchatlas.com/blog/why-do-ai-search-results-keep-changing.

²² MaximusLabs AI (Krishna Kaanth): "AI Share of Voice: How to Measure Brand Visibility in AI Search, The 30-Sample Methodology for Reliable SOV Data," March 2026. Available at maximuslabs.ai/ai-search-101/geo/measurement/share-of-voice.

²³ Qwairy: "Perplexity vs ChatGPT Citation Study: An Analysis of 118,000 AI-Generated Answers," Q3 2025.

²⁴ Pulsar: "How to Monitor Your Brand in ChatGPT, Perplexity and Gemini," April 2026. SearchSeal: "How to Audit Your Brand Across AI Agents in One Day," 2026. Senso.ai: "The Complete Guide to Tracking Your Brand in ChatGPT, Claude and Perplexity," 2026.

²⁵ The DataStory (Justin Evans): "Case Study: It's Profound, Profound raised 96M USD in February 2026," April 2026. Bain x Profound: "Influence Orchestration in the GenAI Era: How LLMs Are Reshaping B2B Discovery," 2026.

²⁶ Surferstack: "AI Visibility Platforms in 2026: Promptwatch, Profound, Peec AI and Otterly.AI Compared," April 2026. AllAboutAI: "12 Top Tools for Tracking Brand Visibility on GPT and AI Search," 2026.

²⁷ Gartner: "AI Ethics, Governance, and Compliance," survey of more than 1,800 executive leaders, 2025/2026. Available at gartner.com/en/articles/ai-ethics-governance-and-compliance.

²⁸ Assembly (AI Assembly Lines): "How Do Companies Structure AI Governance? A Framework for Mid-Market Leaders," 17 March 2026. Available at aiassemblylines.com/post/how-do-companies-structure-ai-governance-framework.

²⁹ Dialzara (Adam Stewart): "How Do Retailers Structure Cross-Functional Teams for AI Product Work?," January 2026 (update). With reference to Harvard Business School research on cross-functional AI teams.

³⁰ The CDO TIMES (Carsten Krause): "Walmart Case Study: Best Practices for Setting Up an AI Center of Excellence (CoE) in Retail," 7 June 2024.

For a deeper analysis of the data architecture behind one of the most aggressive platform plays in this space, see our piece on the Shopify Agentic Plan. For a sober reading of where the operational maturity sits today, see the Agentic Commerce reality check, and for a deeper protocol-level treatment the piece on UCP versus ACP.