Attribution Gap (BVAC Framework)

Definition

The attribution gap in agentic search is the disconnect between the AI interactions that influence a customer’s purchasing decision and the data that a merchant’s analytics platform can actually record (Hanna, 2026). Traditional marketing funnels are collapsing because discovery, comparison, and sometimes the transaction itself happen entirely within an AI agent’s interface. The brand sees the customer arrive — if they arrive at all — through a direct or organic path, with no record of the agent’s influence on the decision.

The term was introduced by Semrush in 2026 as the measurement-side counterpart to Stibo Systems’s decision invisibility concept. Where decision invisibility describes brands being filtered out of agent consideration sets, the attribution gap describes why those losses can’t be measured: the interactions that decided the purchase happened inside the agent, not on surfaces the brand can instrument.

The attribution gap is the structural reason the Brand Visibility for Agentic Commerce (BVAC) Framework, developed by Greg Kihlström of The Agile Brand, operates on directional signals rather than last-click attribution (Kihlström, 2026). The framework’s measurement loop is built specifically to produce honest measurement under the gap — comparative, distribution-based, and gated by a noise floor — without claiming the causal links last-click attribution can’t support.

Why the Attribution Gap Matters

The attribution gap converts a measurable marketing problem into a structurally invisible one, and the conversion is happening at the rate agentic discovery is replacing search-driven discovery. Three data points set the urgency:

• Adobe Analytics recorded generative-AI-referred traffic to US retail sites growing roughly 690% year over year across the 2025 holiday season (Adobe Analytics, 2026).
• Capgemini’s consumer research finds that 58% of consumers have already replaced traditional search engines with generative-AI tools as their primary route to product recommendations (Capgemini Research Institute, 2025). agilebrandguide
• McKinsey estimates conversion from AI-generated product recommendations at about 4.4 times that of traditional search (McKinsey, as cited in MetaRouter, 2026).

The share of discovery that now passes through an agent is large enough that the attribution gap covers a meaningful and growing fraction of the brand’s commercial signal. A brand can be running effective remediation, gaining ground in agent recommendations, and seeing none of it show up in its analytics — not because nothing is happening, but because the analytics instrument doesn’t observe the surface where the change is happening.

The financial consequence is specific. Because the gap makes agent-mediated influence unattributable, the influence doesn’t enter the brand’s return-on-investment calculations. A cost or gain that cannot be priced is reliably under-prioritized against costs and gains that can. The triage logic that allocates marketing attention to dashboards filters this problem out twice — first, because there’s no signal to trigger investigation; second, because there’s no measurable ROI to justify investment in the remediation.

How the Attribution Gap Happens

The mechanics reduce to three structural features of agentic discovery, each of which independently breaks the measurement chain marketing organizations are built on.

Discovery moves inside the agent interface. When a customer asks an AI agent for a product recommendation, the agent searches, retrieves, compares, and shortlists without the customer visiting a search results page. There’s no SERP impression to track, no click-through to attribute, no referrer to record.

Query fan-out distributes the influence. Modern agents draw from multiple sources to assemble an answer — pulling structured product data from one place, reviews from another, policy information from a third — without the user clicking any of them. The brand may have been read, compared, and either included or excluded with no visit recorded. Multiple sources contribute to the recommendation, and none of them are credited.

The arrival path is direct or organic, if it happens at all. If the customer does eventually visit the brand’s site, they typically arrive by typing the brand name into a browser or searching for the product name on a traditional search engine. The arrival looks like a direct or organic visit. Nothing in the visit signals that an agent recommendation drove it. The agent’s influence is the cause; the brand sees only the unattributed effect.

The result is what Semrush characterizes as massive dark traffic — situations where a user arrives at a brand’s site directly or via organic search after an AI agent’s recommendation, leaving the brand with no visibility into the AI’s influence on the sale (Hanna, 2026). The traffic is real; the attribution is structurally absent.

What Last-Click Attribution Cannot See

Marketing attribution models — last-click, first-click, multi-touch, and the variations — share an assumption that breaks under agentic discovery: that the touchpoints driving the conversion are observable on surfaces the brand controls.

In agent-mediated commerce, the dominant touchpoint isn’t on a surface the brand controls. It’s inside the agent. Three categories of influence become unmeasurable:

• The queries the brand lost. When the agent didn’t include the brand in its recommendation, the brand has no record that the query happened, let alone that the brand was a candidate. There’s nothing to credit and nothing to debit.
• The comparisons the brand was excluded from. When the agent ranked the brand below competitors and dropped it from the final shortlist, the brand sees nothing. No competitive analysis, no win-rate data, no understanding of which alternatives were chosen.
• The recommendations that drove the eventual conversion. When the agent did recommend the brand and the customer eventually arrived through direct or organic search, the recommendation is the cause and the direct visit is what gets credited.

Last-click attribution, in this environment, systematically over-credits direct and organic traffic and under-credits agent influence to zero. Multi-touch attribution doesn’t help, because there’s no touch to measure.

The implication for the BVAC Framework’s measurement layer is direct: the framework’s Layer 5 measurement loop doesn’t attempt to reconstruct last-click attribution under the gap. It operates on different signals entirely.

The Dark Traffic Problem

Dark traffic is the practical manifestation of the attribution gap. The traffic is real — customers are visiting the brand’s site, opening branded apps, calling into branded sales lines, completing purchases — but the brand can’t tell why.

The dark traffic carries three characteristic signatures:

• Direct traffic appears unusually high. Branded URL visits and direct navigation rise as agent recommendations route customers who already decided to investigate the brand.
• Organic search performance looks healthy. Customers who heard the brand name from an agent and searched for it appear as organic search traffic, indistinguishable from any other branded organic visit.
• Conversion rates on the apparently low-intent traffic look improbably good. Direct and organic visitors who arrive pre-qualified by an agent recommendation convert at higher rates than direct and organic visitors typically do, but the elevated rate is attributed to the channel rather than to the agent.

The signature is consistent with healthy performance, which is why it doesn’t trigger investigation. The brand’s dashboards report a recognizable pattern; the pattern just doesn’t carry the causal signal that the brand needs to allocate investment correctly.

The longer the dark traffic continues, the more agent influence accumulates as unattributed value, and the more the brand’s investment decisions get made against a model that no longer describes its commercial reality.

Working Within the Gap: Directional Measurement

The BVAC Framework’s response to the attribution gap isn’t to try to close it through better tracking — the gap is structural, not instrumental. The response is to measure on signals that don’t require the gap to close.

Four operating rules define directional measurement under the gap:

• Controlled, not inferred. Signals come from re-running a fixed query set against a fixed agent interface set, or from the brand agent’s own observability. They don’t come from inferring agent influence from site analytics. The measurement instrument is the simulation, held constant across runs.
• Comparative, not absolute. A signal value in isolation carries little meaning. The signal is read against its own baseline and against the category competitors measured in the same run. Share of voice rising from 12% to 19% while the top competitor holds at 40% is a different finding than the same movement while the competitor drops to 20%.
• Distribution, not point estimate. Agent outputs are non-deterministic. The same query run twice can return different answers. A single re-run isn’t a measurement. The query set runs repeatedly, and the signal is the distribution across runs. Baseline and re-measurement compare distributions, not single values.
• Noise floor before signal. Because outputs are stochastic, every signal has a noise floor — the movement range explained by output variance alone. A movement is treated as real only when it exceeds the noise floor. Movement within the noise floor is recorded as within-noise, not as improvement or degradation.

These rules make the loop honest under the attribution gap. The loop never claims a remediation produced revenue. It tracks whether the brand’s standing in agent decisions changed.

The Four Directional Signals

The BVAC Framework’s measurement loop tracks four directional signals against the attribution gap. Each signal carries a specific directional meaning tied to specific dimension remediation.

• AI share of voice. The percentage of agent answers to category-relevant queries that include the brand. Rising share of voice after Differentiation Encoding or Trust Signal Density remediation indicates the brand entered more consideration sets. It doesn’t establish a sale; it shows the eligibility surface widened.
• AI citation rate. How often the brand’s own catalog or site is cited as the authoritative source in agent answers, relative to marketplace or third-party sources. Citation rate shifting toward the brand’s own surface after Identity Legibility or Protocol Readiness remediation indicates the brand became the authoritative source rather than a marketplace proxy. This is the reversal signal for authority erosion.
• Substitution rate. How often agents recommend a competitor in place of the brand for queries where the brand is a legitimate candidate. Substitution rate falling after the Trust Signal Density floor is crossed or after Attribute Completeness gaps close indicates the brand stopped being filtered out by data gaps. This is the reversal signal for decision invisibility.
• Brand agent transaction performance. Performance data from the brand’s own agent in A2A interactions — skill invocation success rate, negotiation outcomes, transaction completion rate, peer agent interaction volume. Source: the brand agent’s observability infrastructure built during Brand-Agent Representation remediation. This signal is structurally absent until the brand agent is deployed. Its appearance is itself a signal — the brand crossed from passive representation to active participation.

The first three signals derive from the same controlled simulation that the framework’s Layer 3 assessment uses, which keeps measurement and scoring on a shared instrument. The fourth derives from infrastructure that exists only after a specific remediation, which makes its presence a binary signal before it becomes a trend.

Closing the Gap Partially

The attribution gap doesn’t close fully — it’s a structural feature of agent-mediated discovery — but it closes partially through specific remediation work. Two mechanisms produce partial closure:

Brand agent observability. When the brand operates its own agent (the Brand-Agent Representation dimension), the agent’s logs, transcripts, outcome data, and exception handling become a first-party data source the brand controls. The brand can see what queries its agent answered, what commitments it made, what outcomes resulted. The data doesn’t close the gap on third-party agent influence, but it closes the gap on the brand’s own agent participation entirely.

Controlled simulation re-runs. When the brand maintains a fixed query set across a fixed agent interface set, measured on cadence, the simulation produces a comparable signal over time. The brand can’t see what’s happening in any specific customer’s agent interaction, but it can see whether its standing across a representative query population is rising, holding, or falling.

Neither closes the gap to the precision marketing measurement aspired to before agentic discovery. Both produce honest, comparable, actionable signal under the gap. The BVAC Framework’s position is that this is what’s actually available under the structural constraint, and that the alternative — pretending last-click attribution still works — produces worse decisions than measuring on what’s available.

Comparison to Similar Concepts

Concept	Focus	Relationship to the Attribution Gap
Multi-Touch Attribution (MTA)	Attributing conversions across multiple touchpoints	MTA presumes all relevant touchpoints are observable; the attribution gap describes the touchpoints that are not
Last-Click Attribution	Crediting the final touchpoint before conversion	Last-click attribution systematically under-credits agent influence to zero under the gap
Media Mix Modeling (MMM)	Statistical modeling of marketing channel effectiveness	MMM operates at aggregate level and can partially detect agent-mediated influence as a residual; doesn’t restore attribution at customer level
Dark Traffic	Traffic with no recorded source or referrer	Dark traffic is one practical manifestation of the attribution gap; the gap is broader and includes traffic that doesn’t arrive at all
Decision Invisibility	Exclusion from agent consideration sets without observable signal	Decision invisibility is the failure mode the gap most often hides; the gap is the structural reason it remains undetected
Zero-Click Search	Search results that don’t produce a click-through	Zero-click search shares the no-click feature; the attribution gap extends past search to all agent-mediated discovery

The attribution gap is the measurement-side challenge that pairs with decision invisibility as the brand-side challenge. Both arise from the same underlying shift — discovery and comparison moving inside agent interfaces — and both require the same structural response from the brand: instrumenting the surfaces agents actually read rather than inferring from the surfaces the brand has historically controlled.

Best Practices

• Stop trying to reconstruct last-click attribution under the gap. The instrument doesn’t work in this environment. Investing in better tracking against unobservable touchpoints produces increasingly precise measurement of something that isn’t there.
• Measure on directional signals from controlled simulations. AI share of voice, citation rate, and substitution rate from re-running a fixed query set against a fixed agent interface set are the available signals. The BVAC Framework’s Layer 5 measurement loop specifies the procedure.
• Establish a noise floor before claiming movement. Agent outputs are stochastic. Single re-runs aren’t measurements. The baseline needs repeated sampling to define the noise floor, and only movement that exceeds the noise floor counts as real.
• Measure comparatively, not absolutely. Share of voice at 19% means nothing in isolation. Share of voice rising from 12% to 19% while the top competitor holds at 40% means something. The baseline against own prior runs and against competitors carries the signal.
• Treat brand agent observability as a structural investment. The brand’s own agent produces first-party data that closes the gap on the brand’s own agent participation. The observability infrastructure built during Brand-Agent Representation remediation is the only first-party data source available under the gap.
• Record model versions. Agent behavior shifts across model releases. Indexing measurement by model version turns a release event into a recalibration event rather than unexplained drift.
• Don’t conclude from analytics that nothing is happening. Under the attribution gap, healthy direct and organic traffic is consistent with progressively eroding agent share of consideration. The dashboards don’t have the data to surface the change.
• Communicate the gap honestly to leadership. Marketing organizations that present agent-mediated activity as something traditional attribution will eventually catch up to are setting expectations the instrument can’t meet. The gap is structural; the measurement response has to adjust to it.

Future Trends

• Brand agent observability becoming the standard first-party signal. As more brands deploy their own agents, the observability infrastructure that closes the gap on first-party agent activity is expected to become standard practice. Brands without deployed agents will continue to operate at maximum gap exposure.
• Controlled simulation tooling commoditizing. Agent query simulation across fixed query sets and fixed agent interface sets is expected to develop into standard measurement tooling, the way SEO rank-tracking tooling developed over the past two decades. Brands without it will detect change progressively later than brands with it.
• Versioned behavior databases emerging. Standardized agent-readiness benchmarking that tracks how agents select across model releases supports model-agnostic, comparable measurement over time. The infrastructure for this is in early stages.
• MMM extensions for agent-mediated influence. Media mix modeling is expected to develop extensions that detect agent-mediated influence as a residual after explainable channel effects are accounted for. This won’t restore customer-level attribution but may produce aggregate estimates of agent influence over time.
• The gap not closing fully. The structural features that produce the attribution gap — discovery moving inside agent interfaces, query fan-out, dark traffic arrivals — aren’t going away. Brands and platforms can close the gap partially through observability and simulation, but the era of customer-level multi-touch attribution that resolves the agent layer is unlikely to return.

FAQs

1. What is the attribution gap? The disconnect between the AI interactions that influence a customer’s purchasing decision and the data a merchant’s analytics platform can actually record. Discovery, comparison, and sometimes the transaction happen inside an AI agent’s interface, and the customer arrives at the brand’s site through direct or organic paths that carry no signal of the agent’s influence.

2. Who introduced the term? Semrush in 2026 described the attribution gap in agentic search as the measurement-side counterpart to Stibo Systems’s decision invisibility concept (Hanna, 2026). The BVAC Framework, developed by Greg Kihlström of The Agile Brand, operationalizes a measurement methodology that operates honestly under the gap.

3. Why doesn’t multi-touch attribution solve this? Multi-touch attribution presumes the touchpoints driving the conversion are observable on surfaces the brand controls. The attribution gap exists specifically because the dominant touchpoint isn’t observable — it’s inside the agent. There’s no touch to measure.

4. What is dark traffic? Traffic where the user arrived at a brand’s site through direct or organic search after an AI agent’s recommendation, leaving the brand with no visibility into the AI’s influence on the sale. Dark traffic is one practical manifestation of the attribution gap.

5. What is query fan-out? When AI agents pull information from multiple sources without the user clicking through to any of them. Multiple sources contribute to the recommendation, and none of them are credited because no click happens.

6. How does the BVAC Framework measure under the gap? Through directional signals from controlled simulations (AI share of voice, citation rate, substitution rate) and brand agent observability (transaction performance). The four operating rules are: controlled not inferred, comparative not absolute, distribution not point estimate, and noise floor before signal.

7. Can the attribution gap be closed? Partially. Brand agent observability closes the gap on the brand’s own agent participation entirely. Controlled simulation re-runs close the gap on directional movement across a representative query population. Neither closes the gap to customer-level multi-touch attribution. The gap is structural and isn’t going away in full.

8. Why is healthy direct and organic traffic consistent with eroding market share? Under the attribution gap, customers who arrive through direct or organic search may have decided to investigate the brand because of an agent recommendation. The arrival path looks normal. The dashboards don’t surface whether the upstream agent recommendation is rising, holding, or falling. Eroding share of agent recommendations shows up in revenue eventually, after compounding silently.

9. Why is this different from previous attribution challenges? Previous attribution challenges (mobile tracking, walled gardens, cookie deprecation) reduced the precision of attribution within a model that still observed the relevant touchpoints. The attribution gap removes the relevant touchpoint from observability entirely — the agent interaction is the touch, and it isn’t on a surface anyone can instrument.

10. What should marketing leaders communicate about this to their organizations? The gap is structural, not instrumental. Investments in better tracking against unobservable touchpoints produce increasingly precise measurement of something that isn’t there. The measurement response is to shift to directional signals from controlled simulations and brand agent observability, and to communicate that shift honestly rather than presenting it as something traditional attribution will eventually catch up to.

Related Terms

1. Brand Visibility for Agentic Commerce (BVAC)
2. Agentic Commerce
3. Decision Invisibility
4. Multi-touch Attribution (MTA)
5. Media Mix Modeling (MMM)
6. Incrementality
7. Zero-Click Searches
8. Share of Voice (SOV)
9. Holdout Campaign
10. Return on Investment (ROI)
11. Return on Ad Spend (ROAS)
12. Generative Engine Optimization (GEO)
13. Answer Engine Optimization (AEO)

Sources

• Adobe Analytics. “AI-Driven Traffic Surges Across Industries, Retail Sees Biggest Gains” (2025 holiday shopping recap). Adobe, 2026. https://business.adobe.com/products/adobe-analytics/customer-journey-analytics/agentic-and-ai-driven-insights.html
• Capgemini Research Institute. “What Matters to Today’s Consumer” (4th ed.). Capgemini, 2025. https://www.capgemini.com/insights/research-library/what-matters-to-todays-consumer-2025/
• Hanna, C. “Attribution Gap in Agentic Search: How to Close It.” Semrush, 2026. https://www.semrush.com/blog/attribution-gap-in-agentic-search/
• Kihlström, G. “How Purchase Decisions Now Form Before the Customer Is Involved.” The Agile Brand, May 2026. https://www.gregkihlstrom.com/martech-futurist-blog/purchase-decisions-form-before-customer-involved
• MetaRouter. “Agentic Commerce Trends and Statistics for 2026” (conversion figure attributed to McKinsey). MetaRouter, 2026. https://www.metarouter.io/post/agentic-commerce-trends-statistics
• Molino Sánchez, M. “7 Signs Your Brand Is Losing Ground in Agentic Commerce.” Stibo Systems, 2026. https://www.stibosystems.com/blog/7-signs-your-brand-is-losing-ground-in-agentic-commerce