← AI in online sellingC1 · LESSON 1 OF 7

C1, Lesson 1: The Funnel, and Where AI Sits in an Online Store

Tags: Sourced / General / Our view / Illustrative. Reading time about 12 minutes.

1. The story: the store that rearranged itself

In September 1998, Amazon engineers Greg Linden, Jennifer Jacobi and Eric Benson filed a patent application for "collaborative recommendations using item-to-item similarity mappings." [Sourced: US patent 6,266,649 B1, "Collaborative recommendations using item-to-item similarity mappings," priority date 1998-09-18; Google Patents lists it as granted July 24, 2001] The problem they faced was size. The common method of the mid-1990s was user-based: find shoppers similar to you, then suggest what they bought. That means comparing you with a huge crowd every time a page loads. Amazon's version flipped it. Build, ahead of time and offline, a table of "items people who buy this also buy," then at page-load just look things up. [Sourced: Smith and Linden, "Two Decades of Recommender Systems at Amazon.com," IEEE Internet Computing, 2017] In January 2003 Linden, Brent Smith and Jeremy York described it in IEEE Internet Computing under the claim that comparing items instead of customers "scales to very large data sets and produces high-quality recommendations." [Sourced: IEEE Internet Computing vol. 7 no. 1, pp. 76-80] By then the homepage, search results, shopping cart and order-confirmation page all carried recommendations. [Sourced: Smith and Linden 2017] A 2017 look back by Amazon's Brent Smith and Greg Linden is careful about the effect: it says a Microsoft Research report estimated that 30 percent of Amazon's page views came from recommendations. That counts page views, not sales. [Sourced: Smith and Linden, "Two Decades of Recommender Systems at Amazon.com," IEEE Internet Computing, 2017; the report they refer to is not named in the article] A 2015 Microsoft Research study of 9 months of Amazon browsing logs also counted roughly 30 percent of visits arriving through recommendation clicks. Its authors then estimated that at least 75 percent of that activity would likely have happened anyway, through search or other routes, so a click count overstates what recommendations add. [Per the verifier, not read by the drafter. Sourced: Sharma, Hofman and Watts, "Estimating the causal impact of recommendation systems from observational data," arXiv 1510.05569, abstract and Figure 2-3 text read Oct 7, 2026] The lesson for this course: for the first time, an online store did something a physical store could not. It computed, per shopper, what to show next. That is the pattern of everything in this course.

2. The one idea

An online store is a funnel: visitors enter at the top, a few reach "paid" at the bottom, and AI is just a set of tools placed at different points to change how many make it through, or to cut what each step costs. [General] Everything we add later should answer two questions: which step does it change, and what does it cost and break?

3. The kitchen-table version

Your market stall. A hundred people walk past. Twenty stop. Five pick something up. Two pay. Now hire a helper who (a) steers walkers to the right stall corner, (b) remembers regulars and puts their usual on top, (c) answers "do you have this in blue?" so you can keep serving, and (d) spots the person who says they paid but didn't. Four helpers, four different jobs, four different ways to go wrong. [General]

4. The funnel, step by step

  1. Find: a visitor arrives (ads, search, email, a link) and uses site search or browsing.
  2. View: a product page.
  3. Cart: they add to cart.
  4. Pay: checkout, with fraud checks and payment.
  5. After: delivery and returns (Commerce Modules 12 and 12B). Baseline fact: Baymard Institute, which averages 50 studies, puts documented cart abandonment at 70.22%. Different studies vary in quality, so read it as "roughly seven in ten carts are left," not a precise law. [Sourced: Baymard Institute, cart abandonment rate list; secondary average of 50 studies]

5. The four jobs AI does in a store

JobPlain meaningExample in the funnel
RankSort a list so the best items come firstSearch results, recommendation rows
PredictGuess a number or a yes/noWill this order be fraud? Will this person buy?
GenerateWrite or draw something newProduct descriptions, images, email copy
ConverseTalk with the shopperSupport chat, shopping assistant
[General; our framing]
Each job has its own failure: ranking can bury what you want to sell, predictions can be confidently wrong, generated text can claim things that are not true, and conversation can promise things the store cannot honor. Lessons 2 to 7 take them in turn. [Our view]

6. Worked example: what is one funnel step worth? (the $12 water bottle)

All numbers below are [Illustrative], computed by script. A small store gets 10,000 visits a month. Assume:

  • 55% view a product page: 5,500
  • 12% of those add to cart: 660
  • 55% of carts reach checkout: 363
  • 60% of checkouts are paid: 217.8, about 218 orders Overall conversion: 218 / 10,000 = 2.18%. Cart abandonment: 1 - 218/660 = 67% (close to Baymard's average, by design). Order value is $36 (a bottle, a cap and a brush) at 40% gross margin: revenue $7,841, gross margin $3,136 a month. Now suppose better recommendations lift add-to-cart by 5% relative (12% becomes 12.6%). Orders rise to 228.7, about 11 more. Extra gross margin: 10.9 x $36 x 0.40 = $157 a month. So what: a good-looking 5% lift is worth about $157 a month on this store. If the tool costs $300 a month, it loses money. [Our view] The skill we are teaching is asking "what is it worth, and what does it cost?" before "is it clever?"

7. A tiny recommender you can do by hand

Six orders, each a set of items [Illustrative]:

  1. bottle, cap 2. bottle, cap, brush 3. bottle, sleeve 4. bottle, cap 5. cap, brush 6. bottle, brush Among the 5 orders with a bottle: cap appears in 3 (60%), brush in 2 (40%), sleeve in 1 (20%). Is that "unusually often"? Compare with how often each item appears in all 6 orders: cap 67%, brush 50%, sleeve 17%. Ratio (lift): cap 0.9, brush 0.8, sleeve 1.2. Lesson: cap looks like the best match but is just a popular item. Item-to-item systems must compare against how common an item is, otherwise they recommend bestsellers to everyone. With six orders the sleeve result means nothing; real systems need thousands. [General; teaching example]

8. What it costs and what it breaks

  • Cost: tools, data work, review time, and mistakes. Count all four.
  • Breaks: a new tool changes who sees what. A merchandiser who cannot override a ranking loses control of margin.
  • Tell-tale sign of trouble: nobody can say which funnel step the tool is supposed to move. [Our view]

9. Try it yourself (no code, 30 minutes)

Pick any online store you run or use. Write down the five funnel steps. For each step, name one thing that annoys you as a shopper. Then mark which of the four jobs (rank, predict, generate, converse) could help. Finally, estimate the funnel numbers for a store you know (guess if needed) and compute what a 5% relative lift at one step is worth.

10. Self-check

  1. What did the Amazon item-to-item idea change compared with matching similar shoppers? It compared items offline in advance, so recommendations were lookups at page-load.
  2. Name the four jobs AI does in a store. Rank, predict, generate, converse.
  3. A store converts 2.18% of visits. What is cart abandonment if 660 carts yield 218 orders? About 67%.
  4. Why was "cap" a bad recommendation for bottle buyers in our example? It is bought with almost everything; lift was 0.9.
  5. A tool costs $300 a month and lifts margin by $157. Keep it? No, unless it has other benefits you can count.

11. Where this is going

Later lessons end with AI shopping agents (Lesson 7): software that browses for the buyer. We treat "superintelligence" only as a long-range label for that direction; nothing in the lesson depends on it. [Our view]

Sources

Review flags

  • RESOLVED (Oct 7, 2026): patent title. The exact title on Google Patents for US 6,266,649 B1 is "Collaborative recommendations using item-to-item similarity mappings"; inventors Linden, Jacobi, Benson; filed and priority Sept 18, 1998; Google Patents lists the application as granted July 24, 2001 (Google notes its legal data is not a legal conclusion).
  • RESOLVED: the "30 percent of page views" line. The 2017 article is by Smith and Linden (not Linden alone), says "a Microsoft Research report estimated 30 percent of Amazon.com's page views were from recommendations," and does not name the report. The lesson says so. A 2015 Microsoft Research study (Sharma, Hofman, Watts) found the same roughly 30% by raw click count but estimated at least 75% of that activity would likely occur anyway. We cannot confirm it is the report Smith and Linden meant, so the lesson does not say it is.
  • Baymard 70.22% is an average of 50 studies; the lesson presents it as approximate.
  • All funnel numbers, the order list and the $300 tool cost are illustrative.
  • No AI-first claims about Amazon sales impact are made.

Claim ledger (Lesson 1)

#ClaimSourceVerifiedStatusRe-check
1Patent filed with priority 1998-09-18, inventors Linden, Jacobi, BensonGoogle Patents US62666492026-10-06Read page header; title wording uncheckednone
2Mid-1990s CF was user-based; Amazon's was item-based with offline table and online lookupsSmith and Linden 20172026-10-06Readnone
3IEEE paper Jan 2003, vol 7 no 1, pp 76-80; quoted abstract lineIEEE/ACM page2026-10-06Readnone
4By 2003 homepage, search, cart, order-end carried recommendationsSmith and Linden 20172026-10-06Readnone
5~30% of page views from recommendations (Microsoft Research estimate)Smith and Linden 2017; Sharma et al. 20152026-10-06Secondary source explicitly attributed; underlying report unidentified; Oct 7 review resolved by attributionHistorical source limitation
6Cart abandonment 70.22%, 50 studiesBaymard2026-10-06Read; secondary average2027-01-06
7Funnel and recommender numbersOur computation2026-10-06Illustrative, script-checkednone
8Patent title exact; granted July 24, 2001Google Patents US6266649B12026-10-07Readnone
92015 MSR study: ~30% of visits via recommendation clicks; at least 75% likely to occur anywayarXiv 1510.055692026-10-07Read abstract and figures textnone