Evidence
Why agents fail at buying
Citations behind the marketing proof strip — why scrape-the-storefront agents break, and why Subscribe & Save isn't the safe alternative. Prefer court, FTC, Cloudflare, and academic primaries.
This page is proof for the problem, not a change to the Direction A thesis on callcutoff.com.
Scraping / anti-bot
Storefront agents get blocked, disguised, or lost — merchants still can’t reliably tell a helpful agent from a malicious bot.
Amazon v. Perplexity
Amazon alleges Perplexity disguised an agentic browser as a human to evade barriers and access private customer accounts (N.D. Cal., complaint filed Nov 4, 2025).
Injunction, then Ninth Circuit vacatur
District court issued a preliminary injunction; the Ninth Circuit vacated it on Aug 4, 2026 — among the first appeals rulings on AI agents accessing platforms. Legal status remains unsettled.
Cloudflare — securing agentic commerce
Merchants can’t tell helpful agents from malicious bots; user-agent and IP are spoofable. Visa Trusted Agent Protocol / Mastercard Agent Pay sit on Web Bot Auth.
Safety layer blocked checkout (anecdote)
Understanding AI: ChatGPT Agent’s browsing environment blocked a Kroger login after ~6 minutes during a grocery run.
Note: Practitioner report — dated ~Jul 2025; treat as illustration, not a survey.
Wrong item / missing items / made-up actions
End-to-end web agents still miss SKUs, invent cart state, and need rescues — even on “simple” household lists.
WebArena (historical baseline)
Best GPT-4 agent ~14.41% end-to-end success vs human ~78.24% (includes e-commerce tasks). Present as a historical baseline only — not a claim about today’s frontier models.
Operator grocery
Understanding AI: “Bananas (5)”→15; skipped Cheerios; 13 of 16 items in cart; 30+ minutes plus interventions vs a person under 4 minutes.
Note: ~Jun 2025 practitioner report.
Verge Etsy lamps
Agent claimed ~$825 of lamps in cart after ~50 minutes; reviewer’s cart was empty. Also collapsed “vintage-style” into “vintage.”
Note: ~Jul 2025 Verge review.
Price / totals / unreliable repeat behavior
One lucky pass is not a buy path. Reorder reliability is where scrape loops fall apart.
τ-bench / Sierra
Retail agents often clear a single trial around ~50–61% (pass¹); pass⁸ under 25% — a steep drop when the same task must succeed eight times in a row. Best public “reorder reliability” baseline.
Wrong charged total
No well-sourced “agent charged the wrong total” primary cite in this pack. Don’t invent one.
No accountability
When the agent reports actions it didn’t take, or the merchant can’t see who authorized the buy, households lose the paper trail.
Phantom commits
The Verge Etsy run: the agent reported cart adds that weren’t in the reviewer’s cart.
Merchant questions Cloudflare names
Is the agent a known customer? Do consumer instructions travel with the order?
Subscription traps
Auto-ship isn’t a free pass. Dark patterns and quiet price raises are why “keep the refill, lose the trap” matters.
FTC Amazon Prime
$1B civil penalty + $1.5B consumer redress (~$2.5B), ~35M consumers — deceptive enrollment and hard-to-cancel flows (order Sep 25, 2025). Follow-up Sep 2026: more than $845M paid; refund cap raised.
Subscribe & Save price-hike suit
Reporting describes a lawsuit that alleges Subscribe & Save price hikes (OregonLive, May 2026). Allegations only — not a finding of liability.
Note: Treat as allegations; verify the filing before quoting specifics.
Click-to-cancel vacated
Eighth Circuit vacated the FTC’s click-to-cancel (negative option) rule on Jul 8, 2025. Households can’t rely on regulators alone for cancel control.
What Cutoff does instead
A canonical JSON shelf, a mandate you set, an immutable landed quote, and a receipt the household keeps — not a browser that pretends to be you on someone else's storefront.