The Loss of Attribution

The earlier articles in this series traced the structural shift. This article changes perspective and examines the same erosion from the technical side, where it manifests in specific, documentable fractures running through the entire attribution infrastructure.

The answer, it turns out, is more technically specific and more immediately consequential than the abstract framing might suggest.

Where attribution breaks, technically

Attribution rests on a deceptively simple premise: that a sequence of touchpoints can be assigned to a single individual. In practice, this requires a set of technical conditions that have been quietly disintegrating for the past several years.

The most consequential disruption was Apple's introduction of App Tracking Transparency in April 2021. Since iOS 14.5, apps must obtain explicit permission before tracking user behaviour across other apps and websites. Global opt-in rates have settled at roughly 15 to 25 per cent, which means that between 50 and 65 per cent of all iOS conversions have become invisible to pixel-based tracking systems.¹

The impact was immediate and, for anyone relying on platform-reported performance data, rather alarming. Meta reported revenue losses running into billions of dollars, not because fewer conversions were occurring but because fewer could be attributed to the campaigns that had triggered them.² Reported ROAS collapsed by 30 to 50 per cent for many advertisers, even where actual commercial performance had barely shifted.

The second tectonic change concerns the demise of third-party cookies. Chrome, commanding roughly 65 per cent of global browser market share, has repeatedly delayed full deprecation whilst progressively restricting the technical infrastructure for cross-site tracking.³ Safari and Firefox have blocked third-party cookies for years already. The result is a landscape in which cross-session and cross-site attribution is possible for a diminishing share of users, and that share continues to shrink.

The GA4 problem

When Google Analytics 4 introduced Data-Driven Attribution in 2023, the promise was straightforward: an algorithmic model that would distribute conversion credit across touchpoints more intelligently than rigid last-click or rule-based approaches. Reality has proven rather more interesting.

A growing number of practitioners report that GA4 has inflated the revenue attributed to Google Ads by 50 to 100 per cent within a single year, quite independently of whether actual sales have grown at anything approaching that rate.⁴ The explanation lies in the model's mechanics: it tends to award disproportionate credit to channels for which it holds the richest data, and that channel is, with a certain inevitability, Google's own advertising platform.

Simultaneously, the proportion of traffic classified as "(not set)" or "Unassigned" in GA4 has been growing steadily. These are not edge cases. They represent user flows that the system can no longer attribute, whether due to missing cookies, blocked referrer information or interrupted session chains. In some accounts, they now account for 15 to 30 per cent of total traffic, a figure that rather undermines the confidence one might wish to place in the remaining data.

Cross-device and the identity gap

A further structural fracture runs along device boundaries. A user who researches a product on her smartphone in the evening and orders it from a desktop the following morning will, in most systems, appear as two entirely separate people unless she is logged in on both occasions.

GA4 attempts to address this through Google Signals and User-ID, but both require either a Google login or proprietary authentication. For a typical e-commerce shop, where the majority of visitors browse without logging in, cross-device attribution remains fragmentary at best. The consequence is that customer journeys appear shorter and simpler in the data than they actually were, because their constituent touchpoints are scattered across anonymous profiles that the system cannot reconcile.

AI search engines and the new referrer blind spot

The most recent, and quite possibly the most consequential, erosion of attribution concerns the growing use of AI search engines for product research. ChatGPT, Perplexity, Google AI Overviews and comparable systems are not merely changing how users find information. They are systematically withdrawing a growing share of the decision process from measurement altogether.

The figures point to a shift whose scale most analytics configurations have yet to reflect. AI referral traffic grew approximately 700 per cent year-on-year in 2025.⁵ An analysis by Visibility Labs across 94 e-commerce shops found that visits from ChatGPT alone increased by 1,079 per cent. ChatGPT accounts for 87 per cent of all measurable AI referral traffic. And users who arrive at a shop via ChatGPT convert 31 per cent better than non-branded organic traffic, which suggests that the decision is already substantially formed before the click occurs.⁶

Yet these figures describe only the visible portion of the phenomenon, and therein lies the difficulty.

The zero-click effect. Google AI Overviews increasingly answer product queries directly within the search results page. A study by Seer Interactive tracking 3,119 search terms across 42 organisations documents a 61 per cent decline in organic click-through rates for queries that trigger AI Overviews.⁷ The zero-click rate on Google exceeded 65 per cent in Q1 2026.⁸ Users research, compare and form opinions without ever visiting a shop. When they subsequently purchase, the entire research phase is absent from every analytics system.

The referrer gap. When a user asks ChatGPT which running shoe for overpronation under 150 euros is worth buying and then visits the recommended shop, the visit appears at best as a referral from "chatgpt.com." Which question was asked, which alternatives the AI discarded and why this particular product was recommended remains invisible. In many cases, even the referrer information itself is lost, and the visit lands as direct traffic or "(not set)" in GA4.

The detour through branded search. Perhaps the most consequential attribution distortion involves a mechanism that is nearly impossible to identify in the data: users receive a product recommendation in an AI search engine, then search for the brand name or product on Google and purchase there. In GA4, this conversion appears as "Organic Search" or "Branded Search." The actual cause, the AI recommendation, is entirely invisible. Visibility Labs explicitly notes that this effect systematically understates the true influence of AI search engines on e-commerce revenue.⁶

Pre-selection remains invisible. What distinguishes AI search engines from traditional search is not merely the absent referrer but the fact that the AI has already filtered, prioritised and evaluated before the user reaches a shop at all. The decision as to which three of 200 possible products are presented to a user is made within the AI, not within the shop. The retailer sees the click but not the curation that led to it. He sees the conversion but not the 197 alternatives that were already excluded.

Three truths from a single month

The technical fracture becomes most visible when one compares different systems covering the same period.

Consider a mid-sized online retailer running Meta campaigns, using GA4 for web analytics and consulting its shop backend for revenue verification. At month's end, the three systems present noticeably different figures. Meta Ads Manager reports 420 conversions attributed to its campaigns. GA4 assigns 310 conversions to the source "Paid Social." The shop backend records 380 orders that fall within a plausible temporal window around the campaign activity.

None of the three is lying, exactly. Meta includes view-through conversions that GA4 does not capture. GA4 loses a portion of iOS users who declined tracking consent. The shop backend, for its part, has no visibility into ad creative interaction and can establish only temporal correlations.

The result is three internally consistent yet mutually contradictory accounts of the same commercial activity. And the question that actually matters, namely how many of these purchases were genuinely caused by the campaign rather than merely touched by it, cannot be answered from any of the three.

Server-side tracking as a partial remedy

The most widely adopted technical response to data loss has been server-side tracking. Rather than relying on browser pixels, which are vulnerable to ad blockers, cookie restrictions and Intelligent Tracking Prevention, conversion data is transmitted directly from the server to advertising platforms. Meta's Conversions API (CAPI) is the most prominent example.

The numbers suggest it works, to a point. Implementations report recovering 20 to 40 per cent of conversions that client-side pixels miss. In optimal configurations, capture rates can be lifted from 60 to 70 per cent to approximately 95 per cent.⁹

There is, however, a rather important caveat. Server-side tracking improves technical capture within the chain, that is, the attribution of a purchase to a click or an impression. What it cannot do is capture influences that occurred outside the chain altogether. A user who sought advice in a Slack community, subsequently performed a branded search and purchased directly will be tracked with greater precision, but the decisive impulse remains invisible. The accuracy of measurement increases whilst its explanatory scope stays exactly where it was.

Dark social and the disguise of influence

A particularly instructive case of technically invisible attribution concerns what the industry terms dark social: recommendations and links shared through private channels such as WhatsApp messages, Slack groups, emails or direct messages on social platforms.

When someone opens a product link from a WhatsApp message, the visit typically appears as direct traffic in analytics, because referrer information is lost in the transition from app to browser. The same applies to links shared in closed Facebook groups, Discord servers or private Instagram messages.

Current estimates suggest that up to 80 per cent of all social shares occur through these dark channels.¹⁰ The implications for attribution are considerable. A retailer observing a rise in direct traffic may not be witnessing the fruits of SEO strategy or brand awareness at all, but rather the echo of a viral recommendation spreading through private networks, which the tracking system has classified as sourceless.

Why repair alone is not enough

The cumulative effect of these technical fractures is a situation in which each individual system produces plausible results within its own frame of reference, whilst the composite picture grows steadily less complete.

The prevailing remedies, server-side tracking, Consent Mode v2, Enhanced Conversions, Data-Driven Attribution, each address a specific aspect of data loss and can, within their scope, deliver genuine improvement. What they cannot do is reconstruct a process whose most consequential phases take place entirely outside any technical capture mechanism.

The distinction matters. One class of problem concerns technical precision and can be resolved through better implementation. The other concerns jurisdictional scope and cannot be fixed with the tools of the existing infrastructure, because the relevant information never entered a measurable system in the first place.

What follows from this

The previous articles in this series described the shift from the perspective of user behaviour. This article has shown that the same shift is occurring in parallel at the technical level, independently and in addition.

iOS restrictions, cookie deprecation, cross-device gaps, dark social and the growing influence of AI search engines are not temporary disturbances awaiting a technical patch. They are expressions of a permanent reconfiguration in which the assumption of comprehensive measurability, the very foundation upon which classical attribution was built, no longer holds.

The question that follows is not about the next technical optimisation. It is more fundamental: if the scope of measurability has contracted permanently, what new metrics and evaluation frameworks might enable sound decision-making under these altered conditions?

The next article, New Decision Metrics, takes up precisely this question and examines what a retailer can actually measure when AI search engines send traffic to their shop, and where the line runs between reliable measurement and unavoidable interpretation.

Sources

¹ Flurry Analytics: iOS ATT Opt-in Rates, continuously updated. Global opt-in rates vary between 15 and 25 per cent depending on region and app category.

² Meta Platforms: Earnings Call Q4 2021 and Q1 2022. Meta estimated the revenue impact of ATT at approximately $10 billion for 2022.

³ StatCounter: Browser Market Share Worldwide, as of Q1 2026.

⁴ Widely documented across practitioner publications including Measure School, Analytics Mania and specialist GA4 forums. The inflation of Google Ads attribution through Data-Driven Attribution is a recurring theme in the analytics community.

⁵ Semrush: "AI Search Revolution," 2025. The study documents the growth of AI referral traffic across multiple industries.

⁶ Visibility Labs: "ChatGPT E-Commerce Traffic Study," 2025/2026. Analysis covering 94 e-commerce shops and twelve months of GA4 data.

⁷ Seer Interactive: "The Impact of AI Overviews on Organic CTR," 2025. Study covering 3,119 search terms across 42 organisations.

⁸ SparkToro/Datos: Zero-Click Search Study, continuously updated. The zero-click rate encompasses all search queries that conclude without a click on an organic result.

⁹ Meta: Conversions API (CAPI) Performance Benchmarks, documented in Meta Business Help Center and confirmed by independent implementation reports.

¹⁰ RadiumOne/GetSocial: "The Dark Social Report." The 80 per cent estimate is supported by several independent analyses and refers to the share of private channels in total social sharing.

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