← AI in online sellingC1 · LESSON 7 OF 7

C1, Lesson 7: Shopping Agents, Measurement, and Your One-Page AI Roadmap

Tags: Sourced / General / Our view / Illustrative. Reading time about 14 minutes.

1. The story: when the shopper is software

On November 4, 2025, Amazon sued Perplexity AI. Perplexity's Comet browser had an assistant that could log in to a user's Amazon account and shop for them. Amazon alleged "persistent, covert, and unauthorized access" to its computers, and said it had told Perplexity on at least five occasions since November 19, 2024 that agents should not access Amazon this way. [Sourced: Amazon complaint, N.D. Cal. No. 3:25-cv-09514; Reuters, Nov 4-5, 2025. The complaint is Amazon's allegations.] After a hearing on March 6, 2026, the district court granted Amazon a preliminary injunction. [Sourced: district court order, hearing date stated in order] On August 4, 2026 the Ninth Circuit vacated that injunction and sent the case back. The panel found Amazon was unlikely to win its federal computer-access claim because, on the record before it, it was the user who "accessed" Amazon with the help of the assistant, which the court called a tool. It also said what the opinion was not: "We do not establish a new legal regime governing agentic AI," and it did not decide whether Perplexity could avoid liability under other claims, such as tort. [Sourced: Amazon.com Services LLC v. Perplexity AI, Inc., No. 26-1444 (9th Cir. Aug. 4, 2026), summary and opinion text] On September 10, 2026 the Ninth Circuit denied Amazon's request for the full court to rehear the case, so the panel's opinion stands and the case returns to the district court. [Sourced: Ninth Circuit docket entry 71, Sept 10, 2026] Amazon filed an amended complaint on September 21, 2026, with new allegations; those are Amazon's claims, reported by trade press and not tested in court. [Sourced, secondary: AI Weekly, Sept 22, 2026, citing The Next Web; the filing itself was not read] Why it matters for a store: the first big legal fight over shopping agents has not ended. Whether an agent may shop your site, how, and under what terms is still being settled. The practical job is to decide your own stance, and to be ready either way. [Our view]

2. The one idea

Until now, your store was built for a human with a browser. Shopping agents, programs that search, compare and sometimes buy for a person, read your data, not your design. Two questions follow: what do you want agents to be able to do on your site, and how will you know whether any AI feature, yours or theirs, actually earned its keep? [General]

3. The kitchen-table version

A customer sends a helper to your stall with a list: "cheapest 24 oz bottle, ships by Friday, free returns." The helper does not care how pretty your sign is. They read your price tags, delivery note and returns card, compare with three other stalls, and report back. If your tags are missing or wrong, you are not on the list. And the helper might ask you a question about your return rule. Is your answer written down somewhere they can find it? [General]

4. What is happening in the market (verified dates)

  • April 3, 2025: Amazon began testing "Buy for Me," a feature that lets customers buy from other brands' sites inside the Amazon Shopping app, powered by agentic AI. [Sourced: Amazon announcement, Apr 3, 2025; TechCrunch, Apr 3, 2025]
  • September 29, 2025: OpenAI launched Instant Checkout in ChatGPT, with Etsy first and Shopify merchants announced, and the open Agentic Commerce Protocol co-developed with Stripe. OpenAI said merchants pay "a small fee on completed purchases" and that prices for users are not changed; it did not publish a percentage. [Sourced: OpenAI, Sept 29, 2025, https://openai.com/index/buy-it-in-chatgpt/; Stripe, Sept 29, 2025]
  • March 24, 2026: OpenAI wrote that "the initial version of Instant Checkout did not offer the level of flexibility that we aspire to provide," so it was letting merchants use their own checkout while it focused on product discovery, with the Agentic Commerce Protocol extended to share product feeds. [Sourced: OpenAI, "Powering Product Discovery in ChatGPT," Mar 24, 2026; CNBC, Mar 24, 2026, reports it as a pivot away from Instant Checkout after the feature failed to take off] A January 2026 trade report, citing The Information, said Shopify merchants would pay a 4% fee; OpenAI's own pages do not state that, so treat it as unconfirmed. [Sourced, secondary: PYMNTS, Jan 21, 2026]
  • Nov 2025 to Aug 2026: the Amazon v. Perplexity case above. Commerce Module 21 covers the protocols and dates in more detail (Google's Universal Commerce Protocol announced Jan 11, 2026, and others). Here we stay with what a store does about them. [Sourced: see Commerce Module 21 ledger]

5. What a store should change

  1. Machine-readable product data. Complete, accurate attributes (Lesson 3), price, availability, shipping and returns. Google's product structured-data docs list price, availability, shipping and returns as fields. [Sourced: Google Search Central]
  2. Policies in plain, consistent text. Returns, shipping times, warranties. A bot comparing stores reads the words, and so does your own service bot (Lesson 5).
  3. A stance on agents. Decide: allow, allow through an official feed or protocol, or limit. Write it in your site terms. Do not assume the legal position is settled.
  4. Fraud rules that expect agents (Lesson 6): legitimate agent-assisted orders may look "non-human."
  5. Measurement of agent traffic: mark it so you can compare conversion, returns and margin with human visits. [Our view]

6. Measuring whether any of this works

Every earlier lesson ended with "test it." Here is the method.

  • Holdout test: split visitors at random. One group gets the feature (treatment), the other does not (control). Compare orders per visitor. The only difference is the feature.
  • Incrementality: how many orders happened because of the feature, not those that would have happened anyway.
  • Decide the measure first: profit per visitor, not clicks.
  • Naive credit is wrong: if a tool says "customers who used it bought 9% of the time," that includes people who were going to buy anyway.

7. Worked example: the feature that looked like a win

All numbers [Illustrative], computed by script. You add an AI "find my size" helper. 40,000 visitors split at random: 20,000 get it, 20,000 do not.

  • With helper: 2.30% convert, 460 orders.
  • Without: 2.14% convert, 428 orders.
  • Difference: 32 orders, worth 32 x $36 x 0.40 = $461 margin. Is it real? The standard error of the difference in rates is about 0.147 percentage points; the observed difference is 0.16 points, so the z-score is about 1.09 and the 95% range runs from about -0.13 to +0.45 percentage points. It includes zero. You cannot tell this from luck. (Lesson 4 showed you need roughly 80,000 visitors per variant to detect a 10% lift at a 2% base rate.) Compare the naive view: 1,200 helper users converted at 9%, 108 orders, $1,555 "attributed." That is more than three times the measured effect, because people who use a size helper were already likely to buy. What to do: extend the test, or accept a smaller claim ("not proven"). [Our view]

8. Your one-page AI roadmap (capstone)

Fill in this table for a store you run or know. One row per lesson. Use the worked-example method to estimate benefit and cost.

LessonWhere it hits the funnelIdea to tryEst. monthly benefitEst. monthly costRisk to checkTest
1 FunnelWholePick the weakest step
2 SearchFindFix top zero-result queries
3 PagesViewFacts-table drafts, tell-tale phrase scan
4 PricingCartOne price or offer test, planned sample size
5 ServiceAfterBot with handoff rules and weekly read
6 CheckoutPayMeasure false declines
7 AgentsFindFeed, policies, agent stance
Rule: do the cheapest, most measurable one first. Starting points from our examples: the zero-result review in Lesson 2 (about $461 against about $260 of review time), and shipping-cost clarity at checkout (Lesson 6).

9. What it costs and what it breaks

  • Attention: seven ideas is too many at once. Pick one.
  • Dependence: agent shopping channels can send you customers and take a fee (when OpenAI launched Instant Checkout it said merchants pay a fee on completed purchases, and it has since moved away from that feature; other agent channels may charge too). Compare against your own margin.
  • Legal uncertainty: the Amazon case is a preliminary ruling in one circuit about one statute; it does not settle other claims.
  • Measurement drift: always keep a holdout for features you pay for.

10. Monday checklist

  1. Open your top three product pages as a "reader" with no design: are price, availability, shipping and returns all stated clearly in text?
  2. Write a two-line stance on shopping agents and put it on your team's list.
  3. Choose one feature you already run and set up a holdout for it.
  4. Fill in the roadmap table with rough numbers and pick one row.

11. Try it yourself (no code, 45 minutes)

Complete the roadmap table for one store. For your top row, write: the funnel step, the number it must move, how many visitors you need (Lesson 4, Section 6), and the date you will decide.

12. Self-check

  1. What did the Ninth Circuit decide on August 4, 2026 in Amazon v. Perplexity? It vacated the preliminary injunction: on that record the user, not Perplexity's tool, "accessed" Amazon, so Amazon was unlikely to win its computer-access claims. The case went back to the lower court.
  2. How did the Ninth Circuit describe the computer-access law's reach? It said the law contemplates access by a person, and that the tool is not a person for statutory purposes.
  3. Two things a store can do to be readable by shopping agents? Complete product attributes and plain, consistent policies (shipping, returns).
  4. In our example the helper's holdout test showed 32 extra orders. Why is that not proof? The 95% range includes zero; luck could explain it.
  5. Why is "9% of helper users bought" a bad measure? Those users were already likely to buy; it counts orders the feature did not cause.

13. Where this is going

Some people use the word "superintelligence" for a far-future AI that outperforms people at nearly every task. We do not need it for this course. What matters for a store is nearer: software that shops on a person's behalf, today, with rules still being written. [Our view]

Sources

Review flags

  • District court order: dated March 9, 2026 (Judge Chesney, N.D. Cal. No. 3:25-cv-09514-MMC), after a March 6 hearing; stayed the injunction 7 days and denied a stay pending appeal (verifier-confirmed).
  • RESOLVED (Oct 7, 2026): the Ninth Circuit ruling is narrow and at the preliminary-injunction stage. The Ninth Circuit's docket (Justia mirror, read) shows the Aug 4 opinion (entry 61, text read), Amazon's Aug 18 petition and the Sept 10 order denying rehearing en banc (entry 71: no judge requested a vote). The Sept 21 amended complaint is verified only through trade press (the filing was not retrieved) and is described as allegations. District-court status after Sept 21 not checked. Not legal advice.
  • Amazon's complaint contains allegations; Perplexity's position was not read in full.
  • RESOLVED by rewrite (Oct 7, 2026): Instant Checkout. OpenAI's launch post (read) says only "a small fee"; no percentage is published. A 4% figure is reported by PYMNTS and eMarketer citing The Information (secondary, unconfirmed by OpenAI). OpenAI's Mar 24, 2026 post (read) says merchants can use their own checkout while it focuses on discovery; CNBC calls it a pivot away from Instant Checkout. The lesson now states those facts. There are no "current fee terms" to confirm because the launch feature was replaced; any new ChatGPT merchant terms should be read from OpenAI at the time.
  • Worked-example numbers (conversion rates, 1,200 users at 9%) are illustrative; the z-score and range were computed by script.

Claim ledger (Lesson 7)

#ClaimSourceVerifiedStatusRe-check
1Amazon sued Perplexity Nov 4, 2025; alleged access; told Perplexity at least five times since Nov 19, 2024Complaint excerpt; Reuters2026-10-06Read excerpts; allegations2027-01-06
2Preliminary injunction order dated Mar 9, 2026 (Chesney, 3:25-cv-09514-MMC), hearing Mar 6; stayed 7 days, stay pending appeal deniedDistrict court order2026-10-06Corrected by verifier2027-01-06
39th Cir. vacated injunction Aug 4, 2026; CFAA "contemplates access by a person... it is a tool, not a person for statutory purposes"; en banc denied Sept 10, 2026; amended complaint Sept 21, 2026Opinion No. 26-14442026-10-06Read summary and key passages2027-01-06
4Buy for Me test Apr 3, 2025Amazon; TechCrunch2026-10-06Read excerptsnone
5Instant Checkout, ACP Sept 29, 2025; "small fee"; Mar 24, 2026 shift to merchant checkout and discoveryOpenAI Sept 29, 2025 and Mar 24, 2026; Stripe; CNBC2026-10-07Read; 4% fee only in secondary reports2027-01-06
6Google product data fieldsGoogle Search Central2026-10-06Read excerpts2027-01-06
7Holdout arithmetic ($461, z 1.09, $1,555)Our computation2026-10-06Illustrative, script-checkednone