The cost of trusting a stranger

When a brand books a creator it has never worked with, most of the cost of the deal is not the fee. It is finding candidates, checking that their audiences are real, confirming who they are and what they have done before, writing a contract that covers what could go wrong, and checking afterwards that the work was delivered. Ronald Coase called these the costs of using the price mechanism and argued that firms exist because, inside a firm, some of these costs are avoided (Coase 1937). In a later paper he showed that when such transaction costs are zero, the parties will bargain to an efficient outcome whatever the initial assignment of rights, and that therefore the costs of transacting, not the rules, are usually what matters (Coase 1960). George Stigler gave the economics of search: buyers keep searching while the expected gain from one more look exceeds its cost (Stigler 1961).

Checking candidates one at a time

# A brand looks for a creator whose audience is genuine. A share q of candidates is genuine.
# Checking one candidate costs k (hours of vetting, in money terms); checking reveals the truth.
# Candidates are checked one by one, so the expected number checked per genuine find is 1/q.
V = 1000.0                                  # value of a deal with a genuine creator
print("share genuine  cost per check   expected checking cost per genuine creator   net value")
for q in (0.5, 0.2):
    # manual vetting -> partly machine-readable -> verified data
    for k in (40.0, 10.0, 2.0):
        cost = k / q
        print(f"{q:13.1f} {k:15.0f} {cost:44.0f} {V - cost:11.0f}")

# Without any check (Akerlof): a brand pays at most the average value q*V. A genuine creator
# whose
# own outside option is above that price leaves, which lowers q further.
print("\nno checking possible: price a brand will pay = q * V")
for q in (0.8, 0.5, 0.2):
    print(f"q = {q:.1f}: {q * V:5.0f}  -> genuine creators with outside options above {q * V:.0f} leave the market")
# share genuine  cost per check   expected checking cost per genuine creator   net value
#           0.5              40                                           80         920
#           0.5              10                                           20         980
#           0.5               2                                            4         996
#           0.2              40                                          200         800
#           0.2              10                                           50         950
#           0.2               2                                           10         990
#
# no checking possible: price a brand will pay = q * V
# q = 0.8:   800  -> genuine creators with outside options above 800 leave the market
# q = 0.5:   500  -> genuine creators with outside options above 500 leave the market
# q = 0.2:   200  -> genuine creators with outside options above 200 leave the market

Two things are visible. When genuine candidates are scarce, checking costs dominate: at $q = 0.2$ and $k = 40$, a fifth of the deal’s value goes on finding the right person. And cutting the cost per check, from manual vetting to records that can be verified quickly, recovers most of that value.

The second part is George Akerlof’s “market for lemons”. When buyers cannot tell good from bad, they pay at most the average value; sellers of good quality whose alternatives are worth more than that leave; the average falls; and in the limit, the market can unravel until only poor quality is offered (Akerlof 1970). A creator market flooded with bought followers is a lemons market in this sense. Michael Spence showed the way out: a signal that is cheap for good types and costly for bad ones lets the good types separate themselves (Spence 1973).

Structured data and knowledge graphs

Search engines and, increasingly, large language models build their picture of people and organisations from what they can read and cross-check. Since 2011, the schema.org vocabulary, created by Google, Microsoft, Yahoo and later Yandex, has let websites describe people, organisations, books, products and their relationships in a form that machines can read (Guha et al. 2016). In 2012 Google introduced its Knowledge Graph with the slogan “things, not strings”: entities with properties and links, rather than pages that merely contain words (Singhal 2012).

In the terms of this chapter, structured data is a way of lowering $k$. A person described consistently across their own sites, with stable identifiers that point to independent records, such as an ORCID for publications, a DOI for each book, or a company registration, can be checked in minutes instead of hours. The identifiers are also a Spence signal of a modest kind: they are cheap to provide for someone whose record is real and costly to fake consistently across independent registries.

A digital reading of Coase

Coase’s insight suggests a way to think about reputation in the age of machine readers. When the cost of verifying a counterparty falls, more transactions can happen between strangers, without the protection of a long relationship or a large intermediary. Some of the value that intermediaries capture today is payment for verification; as verification becomes cheaper, that part of their margin should shrink, and the part that pays for real services, such as the productivity gains of Chapter 5, should remain.