SEONIB SEONIB

Agents Fail 43% of Shopping Tasks — and Still Buy Anyway

Author: SEONIB Editorial Date: 2026-11-14 09:39:01
Agents Fail 43% of Shopping Tasks — and Still Buy Anyway

The first hard benchmark of AI shopping agents produced a number nobody expected to be that bad: across 22 models from 12 labs run through realistic shopping scenarios, 57.2% of test runs passed and 42.8% failed. And the failure mode was worse than the rate — in 54% of the failed runs, the agent completed a purchase anyway, choosing a product that did not meet the stated need.

That is the defining risk of agentic commerce in 2026: not that agents cannot buy, but that they buy confidently and wrongly. And the wrongness is often traceable to your own product data.

The trust picture, in numbers

Consumers have adopted assistants for research far faster than for transactions:

  • 71% of US consumers now use AI chatbots, up from 22% two years ago.
  • Among AI users, 56% use it for price comparison and 44% for deal-hunting.
  • Only 23% trust generative AI to handle payments for them; about 5% are comfortable with an agent buying without asking.
  • 76% of shoppers say they do not want an agent buying on their behalf.
  • Trust rises sharply with safeguards: 42% of US millennials would let an agent buy within a $250 budget if returns were allowed within seven days — versus about a third without that option.
  • What buyers ask for: spending caps (30%), instant revocation (29%), easy cancellation (28%).

So the market is not “agents replace shoppers”. It is “agents narrow choices, humans complete purchases — for now”. Design for both.

Where your site breaks an agent

Agents fail on ambiguity, and most product pages are ambiguity factories:

Ambiguity Agent behavior Cost to you
Price only behind a form or “from $X” Estimates, often wrong Lost order or wrong-margin order
Shipping cost discovered at checkout Abandons or misreports Comparison loss
Compatibility stated in prose only Guesses, sometimes wrongly Return, review, refund
Stock status unclear Buys unavailable item Cancellation, bad review
Variant attributes buried in images Picks wrong SKU Return rate spike
No structured return policy Cannot weigh risk Excluded from shortlist

Every row is the same underlying issue: a fact a human would resolve by reading context, an agent resolves by inference. Inference is where the 42.8% comes from.

The facts agents need, in machine-readable form

1. Exact price, including taxes and shipping thresholds. One number, one currency, one unit, per market. “From $49” is a decision the agent has to make, and it will make it badly.

2. Availability with a timestamp. “In stock, updated 2 hours ago” is actionable; “usually ships in 2–3 days” is not.

3. Compatibility as a list, not a paragraph. Explicit yes/no per platform, with the exception stated separately.

4. Variant matrix with attribute names. Size, color, capacity as fields — not as alt text on six photos.

5. Return, warranty and cancellation terms as discrete fields. This is what the trust data says buyers weight most, so agents will too.

6. Honest limitations. “Does not support multi-currency checkout” prevents a wrong purchase; silence invites one.

Guardrails you should build for the buyer

The data says buyers want limits enforced in code, not promised in copy. If you are building an agent-facing flow:

  • Spend caps the buyer sets, enforced server-side.
  • Instant revocation — a single call that cancels an in-flight mandate.
  • Confirm-before-commit on anything above a threshold, or on any attribute the buyer has not previously approved.
  • A human-readable audit trail of what the agent did and why. This is what makes the second purchase happen.

Measurement

  • Agent-originated orders as their own segment, with return rate tracked separately. If returns are 3× baseline, the data is the problem.
  • Attribute completeness score across your top 200 SKUs — percentage with price, stock, compatibility, return terms as discrete fields.
  • “Cannot determine” events, if you have any agent-facing logging. Those are the exact points where the 42.8% happens.

The takeaway

Agentic commerce rewards the catalog that answers questions in fields rather than paragraphs. That is unglamorous work — an enrichment project, not a campaign — but it is the difference between an agent completing the right order and completing the wrong one.

The same discipline applies beyond product data: every claim an automated reader might act on should exist as a discrete, dated field rather than an inference. SEONIB publishes structured fields and per-language metadata for grouped articles, so the facts a machine reads are the same facts a buyer sees — which is the precondition for an agent making a correct decision rather than a confident one.

Share Article

Related Articles

Recommended Reading

Ready to Get Started?

Experience our product immediately and explore more possibilities.