The economics of AI commerce: what's measurable, what isn't, and what that means for budget

Reference reading: For the executive-pitch version of this argument, see building the business case for AI commerce optimization. For the analytics setup that surfaces AI-referred sessions in the first place, see analytics for AI agents.

The economic question every operator asks about AI commerce in 2026 is the same: is this a channel that’s worth funding now, or is it theatre. The honest answer is that the channel is small but structurally undercounted, the available numbers are noisy enough that confident claims should be treated with suspicion, and the investment thesis depends more on the marginal cost of catalog work than on a confident revenue projection.

This post is the shape of what can be said honestly about the economics of AI commerce, what stays in the hypothesis column, and which decisions don’t need a confident revenue estimate to be worth making.

The size question

AI-referred sessions don’t have a settled industry-wide measurement. Standard analytics setups (GA4 default channel grouping, segment configurations) don’t have a first-class category for AI agents — a shopper who clicks through from a ChatGPT recommendation typically arrives with no referrer or with a referrer that classifies as “direct” or “other.” A meaningful share of AI-influenced sessions never produces a click at all because the agent answers in-conversation with information pulled from its retrieval layer.

What that means in practice: the numbers operators see in their own dashboards underrepresent the channel’s real contribution. By how much is unknown — it depends on the catalog’s category, the analytics configuration, and how the agents in question handle referrers (which itself changes as the products evolve).

For a published external benchmark on the underlying mechanism, see Searchviu’s October 2025 testing on AI agents and JSON-LD, one of the few public studies that probes how AI agents fetch and read product pages. The mechanism for the attribution gap follows directly from how agents appear to interact with content.

The selection-effect hypothesis

The plausible mechanism for AI-referred traffic having different behavior than other organic traffic is selection. By the time an AI agent recommends a specific product, the shopper has typically had a short conversation that filters by their articulated criteria — budget, fit, use case. A shopper arriving from that interaction arrives with more pre-decision context than a shopper arriving from a generic Google search.

If that mechanism holds, three downstream behaviors would follow:

• Higher session-level conversion rate (the comparison work happened before the click)
• Higher AOV (the agent’s filter biases toward articulated, higher- consideration products)
• Higher long-term value per acquired customer (the matched recommendation selects for fit, which correlates with retention)

These are hypotheses about behavior, not measured constants. Any operator considering channel investment should measure these for their own catalog before building a financial case on them. Lumio does not have published industry-wide multipliers for AI-referred session behavior, and the ones circulated in 2025-26 marketing content frequently lack a citable source.

The actionable form of the hypothesis: instrument referrer classification and cohort tracking before drawing economic conclusions. The catalog’s own data is the only reliable source.

What’s verifiable about the channel

Three structural facts about AI commerce in 2026 are concretely defensible without per-catalog measurement:

1. The channel exists and is growing. ChatGPT, Perplexity, Claude, Gemini, and Copilot all surface product information. Microsoft’s October 2025 ads guidance for AI search explicitly addresses optimizing content for inclusion in AI- generated answers — the surface is real enough that Microsoft is publishing operator guidance for it.
2. Standard analytics undercount it. ChatGPT and Claude have historically not passed referrer headers consistently, and a share of agent answers complete in-conversation without producing a click. GA4’s default channel grouping does not have an AI category. The undercounting is structural, not a measurement bug to fix.
3. The catalog work that makes a catalog visible to AI agents overlaps almost entirely with the work that makes it visible in Google Shopping, Microsoft Shopping, and other algorithmic discovery surfaces. Structured data, identifier coverage, and feed quality serve all of these surfaces. The marginal cost of adding AI commerce as a beneficiary of work the SEO team is already doing is small.

What’s not on this list: specific revenue percentages, conversion- rate multipliers, or LTV multipliers presented as industry standards. Those numbers are not reliably published.

What this means for budget

Working backward from what’s verifiable, the case for AI commerce investment in 2026 isn’t “this channel produces X% of revenue today.” It’s: the marginal cost of finishing the structured-data and feed work that’s already partly done is low; the surfaces that benefit are real and growing; the per-customer value may also be higher (worth measuring); and not finishing the work compounds — catalogs that wait ship under-specified data relative to peers and lose surfacing share over time.

That argument does not require a confident revenue projection. It requires honest measurement of the work’s marginal cost and an acknowledgment that the channel’s per-customer value is uncertain enough to be measured rather than asserted.

What to ship this quarter

Three concrete moves that don’t depend on a confident channel-size estimate:

1. Instrument AI-referred attribution. Add a custom channel grouping in GA4 that classifies known AI referrers (ChatGPT, Perplexity, Claude, Copilot, etc.) and a “likely AI direct” segment for referrer-less sessions matching AI shopping patterns. Run a 12-week cohort to compare behavior. The catalog’s own data is the only honest source for the behavioral claims this post hedges.
2. Close the structured data floor. Whatever the AI commerce strategy is, products without complete structured data produce zero surfacing. Audit the catalog, fix the gaps, validate. This is a low-effort high-leverage move regardless of channel-size uncertainty.
3. Track LTV by acquisition cohort. If AI-acquired customers do show higher LTV in the catalog’s own data, that’s the strongest argument for continued investment in the channel — and the only honest way to make the case is from the catalog’s own numbers, not borrowed industry multipliers.

The economics of AI commerce in 2026 aren’t a story that fits a single confident metric. They’re the shape of a channel that rewards early instrumentation and punishes catalogs that wait for the dashboards to catch up.