New Decision Metrics

The previous article in this series showed where attribution breaks at the technical level. This article takes up the question that follows: if the scope of classical measurability has contracted permanently, what can an online retailer actually measure when traffic arrives from AI search engines, and what remains beyond reach?

The answer is more layered than the current discourse tends to suggest. Some things are measurable. Some are interpretable. And some remain, for the foreseeable future, genuinely invisible. The task is to separate these three categories cleanly, because confusing interpretation with measurement produces decisions built on ground rather less solid than it appears.

During the writing of this article, Stefan Wenzel's "Agentic Commerce" was published, which addresses several of the metrics discussed here (Share of Answer, Crawl-to-Referral, AI Referral Traffic). At various points in the text, Wenzel's assessments are therefore referenced directly where they allow a comparison with the independently researched primary sources. The aim is not a review of the book (a separate review can be found here), but a transparent cross-section that brings perspectives together and differentiates where the assessments diverge.

What GA4 actually shows today

Google Analytics 4 can, in principle, capture AI referral traffic, though not automatically and not completely. In its default configuration, a visit from ChatGPT appears as a referral from "chatgpt.com," provided the user clicked a link embedded in the ChatGPT interface. Perplexity shows up as "perplexity.ai." Gemini traffic, however, is frequently absorbed into regular organic traffic because the referrer structure within Google's own ecosystem is not cleanly separated.

The first limitation concerns channel grouping. GA4's default setup assigns AI referral traffic to the generic "Referral" category, where it mingles with every other referral source. Anyone wishing to see the AI share in isolation must create a custom channel group and manually assign the relevant domains (chatgpt.com, perplexity.ai, claude.ai, copilot.microsoft.com and others). The configuration takes a few minutes. In practice, most shops have not done it, which means the actual AI share remains invisible in standard reporting.

The second limitation is more consequential. An analysis by Loamly covering 446,000 visits found that over 80 per cent of traffic genuinely originating from ChatGPT arrives at the shop without referrer information.¹ These visits land in GA4 as "Direct Traffic" or "(not set)" and are indistinguishable from regular direct visits. The reason lies in how users interact with AI systems. A considerable proportion do not click the link embedded in the AI response. Instead, they copy the product or brand name, open a browser and search independently on Google. In GA4, this visit appears as organic search or branded search. The actual trigger remains invisible.

The Visibility Labs study, covering 94 e-commerce shops and twelve months of GA4 data, puts the measurable ChatGPT referral share at roughly 1.5 per cent of non-branded organic revenue, rising to 2.2 per cent in the second half of 2025.² The study itself notes that the true influence of ChatGPT on revenue is likely considerably higher, because the detour through branded search is systematically uncaptured.

What GA4 shows, then, is a fragment. It captures the portion of AI traffic that arrives with a clean referrer. That portion is growing rapidly (1,079 per cent year-on-year growth in ChatGPT visits according to Visibility Labs) but remains a small slice of the actual influence.

The conversion rate question and its complications

Among the most frequently cited data points in the current discourse is the observation that traffic from AI search engines converts better than conventional organic traffic. The figures, however, vary rather more than one might wish.

Visibility Labs reports a conversion rate of 1.81 per cent for ChatGPT traffic against 1.39 per cent for non-branded organic traffic, a 31 per cent advantage.² The study is methodologically transparent: it explicitly excludes homepage traffic (which inflates conversion rates) and blog traffic (which deflates them), focusing on commercial-intent pages.

Adobe and Similarweb arrive at markedly higher absolute values (11.4 per cent conversion for AI-mediated traffic versus 5.3 per cent for organic search), though the methodological underpinnings are less transparently documented and the definition of "AI-mediated" appears to be broader.³ In Stefan Wenzel's "Agentic Commerce," it is precisely this Adobe/Similarweb figure that is adopted, without addressing the discrepancy with other sources or explaining the methodological difference.⁴ This is symptomatic of a field in which numbers are frequently passed along uncritically because they support a desired thesis.

More extreme still are the figures from Authority Tech, reporting an 18 per cent conversion rate for LLM referral traffic, though here a different conversion event appears to be measured (lead generation rather than completed purchase, most likely).⁵

The honest assessment is that AI-mediated traffic almost certainly converts better than average organic traffic, because users arrive with a purchasing intention already substantially formed. But the precise magnitude of this advantage cannot be reliably determined from the available data, because the studies measure different metrics, compare different traffic segments and use different definitions of conversion. Treating any single percentage as established fact confuses a directional signal with a point forecast.

A further detail that is frequently overlooked: Visibility Labs finds that the average order value for ChatGPT traffic is 14.3 per cent lower than for organic traffic ($204 versus $238).² Revenue per session is consequently only 10.3 per cent higher, not 31 per cent. The conversion rate alone, in other words, does not tell the full story.

Share of Answer: a concept caught between ambition and operability

Among the new metrics proposed in the context of AI search is "Share of Answer," the proportion with which a brand appears in the responses of generative AI systems. Wenzel defines the metric in "Agentic Commerce" as "the share of a provider's presence within the answer generated by an AI agent" and describes it as "a prerequisite for agentic decisions."⁴

The concept is analytically sound. It describes a real shift, namely that visibility is migrating from the results list into the answer. What is not part of the answer does not participate in the decision process.

The difficulty lies in operationalisation. The KnewSearch Benchmark Report 2026, which analysed 50,000 queries across 15 industries, provides early empirical orientation: the average "Share of Model" (a related metric) for market leaders sits at 31 per cent, the top three brands in a category capture an average of 67 per cent of all mentions, and Perplexity shows the highest source diversity with an average of 5.2 sources per response.⁶

These figures are a start, but they should be read with considerable methodological caution. First, the responses of generative AI systems vary not only by query but by timing, geography and user context, which means any measurement is a snapshot whose reproducibility is limited. Second, there is no standardised access to the responses of these systems. Measurement today requires either manual querying or specialised tools (such as Rankshift, Seenos or KnewSearch) whose methods and sample sizes are not always transparently documented. Third, the question of whether a brand is mentioned in a response is not equivalent to the question of whether that mention leads to a purchase. A mention in a structured comparison carries different weight from a passing reference in a contextual paragraph.

Monitoring Share of Answer as a directional signal is sensible. Treating it as a robust KPI is, at this stage, premature.

The crawl-to-referral ratio: a macro perspective

A second metric gaining currency is the crawl-to-referral ratio: the number of pages an AI system reads before sending a single visitor back. Wenzel references the Cloudflare data from Matthew Prince extensively and, for the most part, accurately.⁴ The figures are striking and substantially correct. Google operates at a ratio of roughly 100 to 1, OpenAI at approximately 3,000 to 1, Anthropic at 60,000 to 1.⁷

What Wenzel describes correctly is the direction. What he does not fully contextualise is the limited operative applicability of this figure at the level of an individual online shop. For a retailer, the crawl-to-referral ratio at macro level is informative: it shows that the implicit contract of the web (content in exchange for traffic) is dissolving. At the operative level, however, an individual shop can neither influence the ratio nor easily use it as a steering instrument. It is an indicator of a systemic phenomenon, not a KPI for one's own performance.

What a retailer can see in their own server log analysis is the crawl activity of specific AI bots (GPTBot, ClaudeBot, PerplexityBot, GoogleOther). An SEOmator analysis of Cloudflare Radar data from Q1 2026 shows that ClaudeBot operates at a ratio of 23,951:1, whilst GPTBot sits at 1,276:1.⁸ Whether one's own content is being crawled at all is a precondition for being considered in responses, but not a guarantee of it.

GEO and AEO: a useful distinction with a blurred boundary

Wenzel distinguishes in "Agentic Commerce" between GEO (Generative Engine Optimisation, optimising for mention in AI responses) and AEO (Agent Experience Optimisation, optimising for selectability by autonomous agents).⁴ The distinction is conceptually valuable because it separates two different objectives, namely being visible and being transactable.

In practice, the boundary is less crisp than the framework suggests. For the overwhelming majority of online retailers in the German-speaking market, AEO in April 2026 remains a theoretical construct, because fully autonomous agent transactions outside closed ecosystems (such as Amazon's Rufus) are not yet occurring at any meaningful scale. What is happening is GEO: optimisation for visibility in AI-generated responses. That is where the operative lever sits for the next twelve to twenty-four months.

The risk is that the AEO concept triggers premature investment reallocation. A retailer who channels resources into protocol compatibility and API provisioning for autonomous agents whilst the measurable traffic still derives from GEO mechanisms is courting misallocation. This does not mean AEO preparation is irrelevant, but the sequencing should reflect actual market conditions rather than projected ones.

What can be read from traffic that defies clean attribution

The fact that a growing share of traffic cannot be cleanly attributed does not mean no insights are possible. It requires a different analytical approach, one that relies more heavily on statistical patterns and historical comparisons than on deterministic attribution.

Three approaches have proved practicable.

The first concerns the analysis of direct traffic anomalies. If a shop's direct traffic rises over several weeks without a corresponding increase in brand awareness campaigns, PR activity or offline advertising, the inference is reasonable that a portion of this rise originates from non-attributable sources, among them AI recommendations, dark social and peer endorsements. The method is a diagnosis of exclusion, in which one identifies the explained causes and treats the remainder as a potential indicator of invisible influences. It is not precise. But as a directional signal observed over time, it is more informative than no consideration at all.

The second approach concerns post-purchase surveys. Visibility Labs explicitly recommends embedding a simple question on the order confirmation page, such as "How did you hear about us?".² Response rates are typically modest (5 to 15 per cent), but the answers provide qualitative signals that appear in no analytics system. If 20 per cent of respondents report that an AI recommendation brought them to the shop whilst GA4 shows only 1.5 per cent AI referral traffic, the discrepancy itself is a valuable piece of information.

The third approach concerns correlation analysis between Share of Answer trends and downstream metrics. If a retailer systematically tracks their visibility in ChatGPT responses (whether through monthly manual querying or specialised tools) and correlates this with trends in brand search volume, direct traffic and conversion rate, patterns may emerge over quarters that point towards a causal relationship. The connection remains correlative rather than causally provable, but it is the best available approximation of a reality that eludes direct measurement.

The three categories of the new measurement landscape

What emerges from this analysis can be summarised in three categories.

Measurable is AI referral traffic that arrives at the shop with a correct referrer, provided GA4 is configured accordingly. Measurable are the conversion rates and average order value of this segment, even if the statistical samples at many shops are still too small for robust conclusions. Measurable is the crawl activity of AI bots in server logs. And measurable is the trend in the share of "(not set)" and "Unassigned" traffic in GA4, which serves as a proxy for the growing attribution gap.

Interpretable but not provable are the causes behind changes in direct traffic and brand search volume that cannot be explained by known campaigns. Interpretable is the relationship between Share of Answer trends and revenue trends. And interpretable is the question of how much of the visible conversion advantage of AI traffic stems from pre-selection and how much from self-selection, that is, whether AI systems increase purchase probability or primarily forward users whose buying intent was already high.

Invisible remain the specific decision criteria by which an AI system recommended or discarded a particular product. Invisible remains the full extent of the detour through branded search, which systematically understates the true influence of AI recommendations. And invisible remains the entire domain of peer recommendations in closed channels, whose volume, by current estimates, accounts for up to 80 per cent of all social shares.⁹

What this means for decisions

The temptation is to wait for better data. That would be a mistake, because the completeness that classical attribution once promised is not coming back. The systems that influence purchasing decisions are not structurally designed to return transparency to the retailer.

What is emerging instead is a steering logic that operates on several levels simultaneously. At the measurable level, this means configuring GA4 correctly, isolating AI referral as a dedicated channel, checking server logs for bot activity and monitoring "(not set)" shares. At the interpretive level, it involves contextualising direct traffic anomalies, deploying post-purchase surveys and observing Share of Answer trends correlatively. At the strategic level, it means accepting that a portion of decision formation will remain permanently invisible, and orienting one's data strategy not towards gap-free attribution but towards probability statements and pattern recognition.

Wenzel's "Agentic Commerce" names the right metrics (Share of Answer, Crawl-to-Referral, AI Referral Traffic) but treats them at points with a matter-of-factness that runs ahead of these instruments' actual maturity.⁴ The direction is correct. The precision the text implies is not yet available. And it is precisely in this gap between direction and precision that the real operative work takes place: not with finished answers, but with well-founded approximations under uncertainty.

Glossary

AI Referral Traffic Traffic demonstrably forwarded from AI search engines (ChatGPT, Perplexity, Gemini) to an online shop, appearing in analytics tools as a referral with the corresponding source domain.

Share of Answer The proportion with which a brand or provider appears in the generated responses of AI systems. Measures visibility in the new paradigm where formulated answers, not search result lists, create the initial contact.

Crawl-to-Referral Ratio A metric describing how many pages an AI bot reads before sending a single visitor back to the crawled source. Illustrates the asymmetry between AI content consumption and traffic return.

GEO (Generative Engine Optimisation) Optimisation of content with the aim of being mentioned and referenced as a source in the generated responses of AI search engines. The operative successor to classical SEO for AI-mediated visibility.

AEO (Agent Experience Optimisation) A concept extending beyond mere visibility, aiming to prepare products and offers so that autonomous AI agents can independently select and transactionally process them.

Zero-Click Search A search query where the user receives the answer directly within the search engine or AI interface without visiting an external website. Reduces organic traffic for content providers.

Dark Social The sharing of links and recommendations through non-public channels (messaging apps, email, private groups) whose traffic is not identifiable as social traffic in analytics systems and typically appears as direct traffic.

Post-Purchase Survey A brief questionnaire on the order confirmation page that collects qualitative data on the origin of the purchase decision, closing gaps in automated attribution.

Sources

¹ Loamly: "Does ChatGPT Send Traffic to Websites? (Real Data From 446K Visits)," November 2025.

² Jeff Oxford / Visibility Labs: "ChatGPT Traffic Converts 31% Better than Non-Branded Organic Search (94 eCommerce Sites Analyzed)," February 2026.

³ Adobe / Similarweb: AI Commerce Traffic Analysis, cited in various industry reports, 2025.

⁴ Stefan Wenzel: "Agentic Commerce," 2026.

⁵ Authority Tech: "LLM Traffic Converts at 18%," March 2026.

⁶ KnewSearch: "2026 AI Search Visibility Benchmark Report," 2026.

⁷ Cloudflare / Matthew Prince: Crawl-to-Referral data, presented at SXSW March 2026 and documented in Cloudflare Radar.

⁸ SEOmator: "GEO Data Report 2026: Which AI Crawlers & LLM Bots Take the Most and Give the Least?," 2026.

⁹ Industry estimates on dark social sharing, aggregated from various sources (RadiumOne, GetSocial), 2024-2025.

This article is part of the Decision Journey Analytics series, examining how e-commerce decisions evolve beyond the classical funnel.