Will AI agents hurt my e-commerce sales? Here's what we found

TL;DR. Anyone giving you a topline "AI will eat X% of e-commerce" number is making it up. The actual impact is asymmetric: replenishment categories (batteries, ink, consumables) are losing share silently to agent-ready competitors today; considered purchases (furniture, custom apparel) are barely affected; comparison-shopped categories sit in the middle. The defensible budget argument isn't "agents will be 30% of traffic next year, panic" — it's "the gap between agent-ready and agent-unfriendly sites is opening now, closing it is cheap, the asymmetry compounds." Sites that score below 35 on agent product findability lose share to competitors that score above 70. That gap is what the Head of E-commerce should care about, not a topline percentage.

Short answer

AI shopping agents are not uniformly hurting e-commerce sales today. They are reshaping share asymmetrically between sites that are structurally agent-ready (real <button> elements, schema.org Product data, accessible variant selectors, agent-allow robots.txt) and sites that aren't. The exposure is biggest in replenishment categories (batteries, ink, vitamins, repeat-purchase consumables) where the shopper delegates the full task to the agent. Considered purchases stay human-driven for now. Comparison-shopped categories sit in the middle. The right defensive move is not to forecast a percentage — the trajectory is too steep to extrapolate honestly — but to close the structural gap against your nearest two competitors while it's still cheap to close.

In this article:

• What we measured, and what we didn't
• The asymmetric finding
• Where this shows up commercially
• What we don't know yet
• What an honest answer looks like for your account
• What we'd push back on
• Frequently asked questions

---

A Head of E-commerce at a DACH retailer asked us this question in three different ways during a 30-minute call:

"Are AI agents eating our funnel? Like, today? Should I be worried about Q4?"

The breathless answer — "AI agents will replace 30% of e-commerce traffic by next year, you must act now" — is what most of the AI-commerce trade press is writing. It's also unfalsifiable, so it's not very useful.

The honest answer is more interesting. We've spent meaningful time instrumenting real AI agents (Claude Desktop with browser MCP, ChatGPT Operator, browser-use, Perplexity) attempting real purchase tasks on real e-commerce sites. We have a working view of what's happening today and a defensible read on what's coming. This post is that view, written for the Head of E-commerce who has a budget meeting in two weeks and wants something to defend.

What we measured, and what we didn't

We measured the conversion of the user → agent → eshop leg of the journey. That is: when a human shopper delegates a buy task to an agent, and the agent lands on a specific e-commerce store, what's the probability the agent completes the task?

We didn't measure the upstream leg (whether the agent picked your store in the first place — that's GEO/visibility territory). We didn't measure the downstream attribution (what share of human-completed sessions started in an AI conversation — that's broken in every analytics stack we've checked, including the ones selling themselves as fixing it).

The agent-commerce problem splits cleanly into three layers, and most of the existing AI-commerce tooling sits on either side of the middle one:

| Layer | What it measures | Who covers it today | |---|---|---| | 1. Visibility (GEO) | "Does ChatGPT mention us when asked?" | Athena, Profound, Scrunch, Peec, Otterly, Semrush AI, Adobe LLM Optimizer, Bluefish | | 2. Arrival + traversal | "Can an agent find and buy our products?" | Empty category. This is the layer we measure. | | 3. Attribution | "What share of revenue came via LLMs?" | Dreamdata, HockeyStack, Bizible — all partially broken for agent-mediated traffic |

The number we care about is the one in the middle: given an agent is on your site with the intent to buy, does it succeed?

We have direct measurement from active replay (a real agent running purchase tasks, with the full session captured) and indirect measurement from passive tracking (our serge.js snippet detecting agent-driven sessions in the wild). The pattern below holds across every category of e-commerce site we've tested.

The asymmetric finding

Agent task-completion rate is bimodal. We have a cluster of sites where agents complete the buy task on the first try, and another cluster where agents abandon before checkout. Very little in between.

The cluster split lines up with structural properties of the site, not category. Specifically:

• Sites that score 70+ on our scanner: agent task completion is 60-80% on a "find product X and add to cart" task.
• Sites that score 35-69: agent task completion drops to 15-35%. The agent typically gets to the product page but fails at variant selection (size, color) or at clicking the add-to-cart button.
• Sites that score below 35: agent task completion under 10%. The agent often can't even find the product because the catalog navigation is invisible to it.

Two things to notice. First, "70" isn't a magic number — it's just the threshold where most of the load-bearing structural fixes are in place (semantic add-to-cart, accessible variant selectors, schema.org Product, robots.txt allows agents, no bot-protection 403). Second, the median e-commerce site we've scanned scores in the 35-65 range. So most stores are in the middle cluster — agents sometimes succeed on them, which is exactly the failure pattern that produces the silent revenue hole.

Where this shows up commercially

Three patterns we've seen, all asymmetric. None of them apply uniformly.

Pattern 1 — Specific-product replenishment (worst affected). "Find me the same pack of AA batteries I bought last time, add to cart." The agent has a specific item in mind. It searches your site or browses to a category. If your catalog and variant selectors are agent-readable, it succeeds. If they're not, the agent goes to a competitor — and the competitor is usually a marketplace (Amazon, a specialist retailer with cleaner structured data). The user never knows their preferred store failed; they just see the agent succeeded somewhere else.

This is the silent loss we keep flagging. The lost sale doesn't appear in your GA4 as a churned customer — it appears as nothing. The session was never logged because the agent's failed click never produced a tracked event.

Pattern 2 — Discovery / browse (less affected today, more affected over time). "Find me a winter jacket under CHF 200." The agent does a multi-store search, compares, and recommends one to the user. Right now, the agent's recommendation is heavily weighted toward sites it can actually browse — same structural-readiness bias as Pattern 1 — but with the added human-decision step at the end. The user often overrides the agent and picks the brand they wanted anyway.

This pattern is less commercially urgent today because the human is still in the loop. We expect this to shift over the next 12-18 months as agents are trusted to complete more of the task, but we're guessing — anyone giving you a number here is making it up.

Pattern 3 — Comparison / verification (lightly affected). "Is this product cheaper on store X or store Y right now?" The agent reads product pages and compares prices. As long as your price is in machine-readable structured data (or a clean DOM element), you're fine. This pattern actually rewards sites that have invested in clean schema.org markup — they win the comparison even if their checkout is broken.

What we don't know yet

We don't have a defensible percentage for "what share of your revenue is at risk from agent traffic this year." Anyone giving you that number is making it up. Here's why we can't:

1. The denominator is moving fast. Agent-driven sessions were a rounding error in late 2024. They're measurable but small in mid-2026. The trajectory is steep but we don't have 18 months of stable data to extrapolate.
2. Substitution effects are unobservable. When an agent abandons your site and the user buys from a competitor, neither you nor we ever see the abandoned session. The "loss" is invisible to every analytics stack we know of, including ours — our snippet detects the agent arriving, but if the agent gives up before triggering any tracked event, even we don't see the full path.
3. Category-by-category variance is enormous. A site selling commodity replenishment items (consumables, ink, batteries) is at much higher exposure than a site selling considered purchases (furniture, custom apparel) where the human stays in the loop. Lumping them together produces a number that's wrong for everyone.

What we do know is that the failure mode is one-way. A site that's agent-ready today doesn't get worse over time. A site that's not agent-ready today loses more share to agent-ready competitors as the agent-mediated share of total traffic grows. That asymmetry is the basis for any defensible budget decision — not the topline percentage.

What an honest answer looks like for your account

If you're trying to defend a budget for this in a Q3 planning meeting, here's the structure of an honest answer:

1. Run the scanner on your top three storefronts (yours, your two closest competitors). Get the three scores.
2. If you're 15+ points behind a competitor, the silent-loss asymmetry is working against you today. If you're 15+ points ahead, you're benefiting from it.
3. The size of the current exposure is small — probably under 2% of total revenue, based on the agent-share trajectory we're seeing — but the size of the positional asymmetry is what matters. Closing a 20-point scanner gap is much easier in 2026 than in 2028, and the cohort of agent-ready sites compounds.

This is the actual budget argument. Not "agents will be 30% of traffic next year, panic." It's "the gap is opening now, closing it is cheap, the asymmetry compounds."

What we'd push back on

A few claims circulating that don't survive contact with actual agent behavior:

• "Publish your llms.txt — it's the most important AI-readiness step." The pitch conflates two layers. Anthropic and Perplexity do consult /llms.txt during their upstream retrieval/indexing — the layer that decides which URL to send an agent to. But once the agent is on your product page making click decisions, it reads the accessibility tree of that page, not a manifest you published months ago. So publishing an llms.txt is fine GEO hygiene; pitching it as your task-completion fix is selling the wrong layer. Full breakdown in our companion piece. (We demoted the llms.txt check from a failure to a low-weight optional in our own scanner for exactly this reason — see the merged PR if you want the rationale.)
• "Add an AI chat widget to your site." Different problem. AI chat is a human-help layer. It does not help an agent that is already trying to browse your product pages on a shopper's behalf.
• "Get listed in ChatGPT's answers — that's the AI strategy." Necessary upstream work (GEO), but it does not help you on the leg we're measuring. If the agent can find you but cannot buy from you, GEO did its job; you still lost the sale.

We don't think the GEO tools are wrong about their problem. We think they're describing a different problem than the one a Head of E-commerce should solve first.

What to do this week

If you read this far and you don't yet have a baseline scanner score for your store and your two top competitors: that's the cheap, defensible step.

Paste your domain at serge.ai — get a score in 30 seconds. Then paste your two competitors. If the gap exists, you'll see it. If it doesn't, you have your budget answer for the next quarter and you can move on.

The thing we'd suggest avoiding: spending money on "AI-readiness consulting" before you've measured. The fixes are usually small DOM-level changes that your existing front-end team can ship in a sprint, once they know which ones to ship.

Frequently asked questions

Will AI agents reduce e-commerce sales?

The impact is asymmetric, not uniform. Replenishment categories (consumables, batteries, ink, repeat-purchase items) are most exposed — agents complete those tasks autonomously and pick the easiest-to-navigate retailer, so structurally agent-unfriendly sites lose share silently. Considered purchases (furniture, custom apparel) are barely affected today because humans stay in the loop. Sites that are structurally agent-ready gain share over time at the expense of those that aren't. Anyone giving you a topline percentage for the impact is making it up — the denominator is moving too fast to extrapolate.

How much e-commerce revenue is at risk from AI agents in 2026?

There is no defensible topline percentage. Agent-driven sessions were a rounding error in late 2024, are measurable but small in mid-2026, and the trajectory is steep but unpredictable. The honest framing is asymmetric exposure, not a percentage of revenue. The exposure depends on your category (replenishment is worst affected), your structural agent-readiness score (sites below 35 lose share, sites above 70 gain it), and your competitive position (a 15-point gap behind a competitor is more urgent than the absolute level).

Do AI agents read llms.txt before buying?

It depends on the layer. Anthropic and Perplexity have publicly confirmed they consult /llms.txt during their upstream retrieval/indexing — the layer that decides which URL to send an agent to. But once the agent is on your product page making click decisions, it does not refetch /llms.txt — it reads the accessibility tree of that page. Publishing one is fine GEO hygiene; pitching it as the lever that protects your revenue from agent-mediated loss conflates the indexing layer with the task-completion layer. The actual lever for click-time agent success is your DOM structure on product pages and checkout.

Why don't AI agents show up in GA4?

Agents navigate via headless browsers and arrive without a Referer header, so GA4 categorizes their sessions as "direct." Even when an agent does trigger a pageview, the GA4 cookie attribution model is built for human session tracking and produces nonsense for agent-driven sessions. The result is an invisible revenue channel: sessions that arrived, failed at checkout, and never appeared in your funnel. Detecting agent traffic requires either server-side log analysis or a purpose-built snippet that fingerprints automation signals.

What e-commerce categories are most affected by AI shopping agents?

Commodity replenishment items (consumables, AA batteries, printer ink, vitamins) are most exposed. The shopper knows what they want, delegates the task to an agent, and the agent picks whichever retailer is easiest to navigate. Considered purchases (furniture, custom apparel, jewelry) are barely affected because humans stay in the loop and override agent recommendations. Comparison-shopped categories (electronics, appliances) sit in the middle — agents can read product specs from schema.org Product data and recommend the best-priced or best-spec option.

---

About the author. Yann Borie is the founder of Serge (Superstellar LLC, Zug, Switzerland). He spends his time instrumenting AI shopping agents on real e-commerce sites and turning what they fail at into shippable findings. Connect on LinkedIn.

This is part of a series on what AI agents actually do on e-commerce sites. See also: How AI shopping agents actually see your website and Simple steps to make your online store AI-ready. Or run a free scan on your store to get a baseline today.