1995: Turn participation into a market

Internet businesses can coordinate several groups whose interests overlap without being identical. A marketplace connects buyers and sellers; a search service connects information seekers, publishers, and advertisers. Amazon's1997 shareholder letter and Google's2004 prospectus are primary records of how two firms described their strategies and risks. The chapter's1995 opening marks an early commercial-Web period, not the founding date or launch of every company in the subtitle.

A network effect occurs when participation changes value for other participants. More sellers can improve selection for buyers; more buyers can make participation worthwhile for sellers. This is not the same as an economy of scale, where average production cost falls with volume. Nor is every additional user beneficial: congestion, spam, counterfeit goods, or irrelevant content can make a larger system worse. State the mechanism and the affected group before claiming a positive feedback loop.

Construct a small marketplace model. Ten sellers each offer a distinct item, while 100 buyers seek items from that assortment. Adding ten sellers with genuinely different items can improve match probability. Adding ten duplicate listings may instead increase search effort. The metric “number of listings” cannot distinguish those outcomes. A better evaluation might measure successful matches, time to find an item, defect rates, and repeat purchases alongside total participation.

Now add sponsored placement. Suppose an organic result produces a useful match for 80 of 100 visitors, while a more profitable sponsored placement produces one for 60. If the platform measures only advertising revenue, it can improve its dashboard while degrading the visitor's task outcome. This is a hypothetical counterexample, not a claim about a measured historical product. It explains why the payer and the user must be represented separately in an incentive model.

Advertising auctions also combine several objectives. A high bid does not necessarily create a good result if users rarely find the advertisement relevant or advertisers receive low-value traffic. Ranking, pricing, fraud detection, and measurement rules jointly shape incentives. Avoid reducing a real system to “highest bidder wins” without examining the documented mechanism and its version. Likewise, a click is an event, not proof of a satisfied user or completed purchase.

Trust requires operational work. Marketplaces need identity rules, dispute processes, payment handling, seller incentives, and mechanisms for removing abuse. Reputation can be informative yet vulnerable to fake transactions, selection effects, and retaliation. A high average rating does not eliminate the need to inspect how the observations were collected.

The durable lesson is to draw separate flows for information, goods or services, money, and authority. Then ask which metric each actor can manipulate and who bears the resulting cost. Company filings offer a dated view of strategy and risk; they should be read critically rather than converted into timeless laws of platform success.

Reconstruction exercise

A marketplace redesign raises sponsored revenue from 200 to 260 units per day but reduces successful buyer matches from 80 to 60 out of 100 visits. Compute the percentage changes. Propose two additional metrics that could test whether the redesign creates durable value, and describe one experiment design issue that could make a short-term comparison misleading. Explain why simply adding more listings does not prove a positive network effect.

Worked solution

Revenue rises 60/200=30%, while successful matches fall 20/80=25%. Useful additional metrics include repeat purchase or return visits over a defined horizon, and refunds or disputes per completed transaction. Random assignment, comparable traffic, seasonality, and delayed outcomes matter: a one-day spike can borrow from later trust or reflect a different visitor mix. More listings help only if they improve a relevant outcome such as selection or match quality; duplicates or spam can increase search cost instead. The evaluation should distinguish buyer, seller, and platform outcomes and predefine how delayed harm will be measured, rather than declaring success from the metric directly optimized.

Test the limits and transfer the idea

1. Expected-value calculation · allow 25 minutes

A revenue metric and a transaction metric can disagree

A fictional advertising slot receives 1000 views. Ad A pays 2 units per click, has a 10% click probability, and 20% of clicks become completed purchases. Ad B pays 3, has an 8% click probability, and 10% convert. Compute expected platform revenue and purchases for each. A team chooses B because revenue is higher. State exactly which objective improved and propose an experiment outcome that could reverse the decision under a durable-customer-value objective.

Get a starting hint

Expected clicks, payment per click and purchase conversion are separate factors.

Compare your reasoning with the worked solution

A yields 100 expected clicks, 200 revenue units and 20 purchases. B yields 80 clicks, 240 revenue units and eight purchases. Revenue rises 20%, while completed purchases fall 60%. The arithmetic supports only the specified short-run revenue objective. Repeat-purchase value, advertiser retention, refunds or later user abandonment could outweigh the additional forty units. An experiment should follow comparable cohorts over a defined period and measure those outcomes, including uncertainty and changes in user mix. Google's prospectus and Amazon's shareholder letter are primary corporate communications about strategy and risks; they are not observations of this invented auction or proof of a timeless best ranking rule.

Check your reasoning

• Computes clicks, revenue and purchases separately.
• Names the objective and both percentage changes.
• Designs a longer-horizon test rather than assuming revenue equals user value.

2. Selection-bias calculation · allow 25 minutes

An average rating can be selected evidence

Seller X has 100 completed transactions: 50 satisfied customers and 50 dissatisfied. Every satisfied customer leaves five stars, but only ten dissatisfied customers leave one star. Seller Y has 100 transactions: 80 satisfied and 20 dissatisfied, and everyone rates with the same five/one convention. Calculate observed averages and true satisfaction fractions. Which seller looks better by rating, and why? Propose an audit that could detect the selection mechanism without fabricating missing ratings.

Get a starting hint

The denominator of the displayed average is submitted ratings, not completed transactions.

Compare your reasoning with the worked solution

X displays (50×5+10×1)/60=260/60≈4.333 stars, despite 50% satisfaction. Y displays (80×5+20×1)/100=4.2 stars with 80% satisfaction. The observed average reverses the satisfaction ranking because response probabilities differ. An audit could randomly invite a sample of all completed buyers, compare response rates by complaint/refund history and investigate suppression or discouragement of negative reviews. Nonresponse remains a limitation; one should report it and examine sensitivity rather than assign guessed scores. A high average may reflect favorable experiences, selective participation or manipulation. Corporate strategy filings can disclose incentives and risks, but actual rating reliability requires transaction-level sampling and collection-process evidence.

Check your reasoning

• Uses rating counts rather than transaction counts for observed averages.
• Shows the reversed ranking and underlying satisfaction rates.
• Proposes an evidence-based audit while acknowledging nonresponse.

Check your understanding

Sponsored revenue rises 30% while successful user matches fall 25%. Which conclusion is defensible?

• The user metric cannot matter because users do not pay directly.
• The redesign is unambiguously better for all participants.
• One platform metric improved while a user outcome worsened; overall value needs a stated tradeoff and longer-term evidence.
• Every advertising-supported service must therefore fail.

Explanation

A multi-sided service needs separate measurements of who gains, who loses, and how effects persist.

A marketplace doubles listings by adding duplicates and spam. Why is this not sufficient evidence of a beneficial network effect?

• Participation must improve a relevant outcome for another participant, not merely increase a count.
• Only lower server cost defines a network effect.
• Listings are never relevant to marketplace value.
• Network effects require exactly two users.

Explanation

More participation can increase search cost or abuse; effects and economies of scale are different mechanisms.

Sources and further reading

• Google IPO prospectus, August 18,2004: Google filing, United States SEC
• 1997 letter to shareholders: Jeff Bezos, Amazon