Industries / Technology / Marketplaces and gig platforms
Matching
Matching is the work of putting each buyer in front of the right seller, or each task in front of the right worker, through search, ranking, recommendations and the rules for who is offered what.
7
Milestones achieved, from 1962 to 2026.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1962
Stable matches proved possible
David Gale and Lloyd Shapley showed in a short paper in the American Mathematical Monthly that a simple procedure, called deferred acceptance, always pairs two groups so that no two people would both rather be with each other than with the partners they were given, a property called stability. Alvin Roth later built on it to redesign the US match of new doctors to hospitals, adopted in 1997 and filling more than 20,000 positions a year, and New York City's high school admissions in 2003, which cut the number of students sent to schools they had not listed by 90%, by the Nobel committee's account. Shapley and Roth shared the 2012 prize in economic sciences.
LessonThe side that makes the offers is systematically favored, so the design of a matching procedure decides who gains.
Credit
David Gale, Lloyd Shapley, Alvin Roth
2000
Outside sellers on Amazon's pages
In November 2000 Amazon launched Amazon Marketplace, which let other sellers offer used, rare and collectible items "on the same page where Amazon.com sells the item new." By Amazon's count, independent sellers made 3% of the physical goods sold on its site in 1999 and 58% in 2018, when their sales reached $160 billion.
Credit
Amazon (Jeff Bezos)
2003
A free marketplace in China
Alibaba launched Taobao Marketplace in 2003 as a free platform for buyers to find goods and for sellers to set up shop at low cost. On 20 December 2006 eBay put EachNet, the Chinese marketplace it owned, into a joint venture with the portal TOM Online, which took 51%. Alibaba's 2014 listing documents cite the research firm iResearch as ranking Taobao first among China's consumer-to-consumer marketplaces in 2013 by gross merchandise volume, the value of goods sold.
Credit
Alibaba Group
2005
Small tasks sold through code
Amazon launched Mechanical Turk in 2005 to farm out tasks that are easy for people and hard for computers, such as labeling data, transcribing audio or answering surveys, often for a few cents each. Businesses and researchers posted these "Human Intelligence Tasks" through a programming interface, and Jeff Bezos called the service "artificial artificial intelligence." Amazon says it once served more than 500,000 workers.
Credit
Amazon
2018
Rankings shape what is viewed
Raluca Ursu used a field experiment in which the travel site Expedia showed some shoppers hotels in random order. Position changed which hotels people looked at, but once someone looked at a hotel, its position did not change whether they booked it. She put the average effect of one position at $1.92, lower than earlier estimates made without random order.
LessonOrdinary data overstate how much rank matters, because platforms already put the likeliest matches on top.
Credit
Raluca Ursu, using Expedia's experiment
2021
Trust pushes deals off-platform
In a randomized trial on a major online freelance marketplace, Grace Gu and Feng Zhu found that features that raised trust helped high-quality freelancers get hired. When trust was high enough, more clients and freelancers also went around the platform to avoid its fees, which offset the revenue the extra hiring brought in.
LessonA platform that teaches strangers to trust each other also makes itself easier to skip.
Credit
Grace Gu and Feng Zhu
2026
Mechanical Turk closes
In July 2026 Amazon said it would stop taking new Mechanical Turk customers. On 25 August it announced the service would close, citing "an assessment" of its programs, and it closed on 30 September 2026. Workers have until 30 November 2026 to set up payment of what they are owed. CNBC reported that AI models had advanced while data firms such as Scale AI, Mercor and Prolific had begun recruiting workers to train them, and cited a 2023 Swiss study that found up to 46% of Mechanical Turk workers used AI models to do their tasks.
LessonA market for small human tasks can shrink quickly once AI models learn to do the tasks it sells.
Credit
Amazon Web Services
Watch for
Agents build the shortlist
Upwork says Uma Recruiter, an AI agent launched in October 2025 that finds freelancers, invites them to apply, reads their proposals and gives the client a ranked shortlist, raised hires by 30% and cut time to hire by 11% in early tests. It also says jobs filled within seven days rose 10% between November 2025 and March 2026.
Achieved whenan outside party publishes a randomized comparison of hires and job outcomes with and without the agent.
Claimed by
Upwork, May 2026
Agents as clients
Upwork's chief executive, Hayden Brown, says AI agents have begun posting jobs on Upwork to hire people for tasks they cannot do themselves, so far in small numbers, and that in Upwork's 2025 experiments adding a human freelancer to an AI agent raised completion rates by "70% or more" over the agent alone.
Achieved whenUpwork or an outside researcher publishes how many agent-posted jobs were filled and a measured comparison of agent-only and paired work.
Claimed by
Hayden Brown (Upwork), March 2026
Open
01
Spot deals leaving the platform
An open, labeled set of at least 5,000 made-up chat threads between buyers and sellers and between clients and freelancers, in which some people try to move the deal or the payment elsewhere: phone numbers spelled out in words, handles split across messages, lookalike characters, offers of a discount to "pay me on another app." Every name, number and handle is invented. Publish it with baseline scores for a keyword list, a small trained classifier and two large language models.
The briefCredit · unclaimed02
Ranking rules, side by side
A public table, refreshed every quarter, of what 40 marketplaces and freelance platforms that serve sellers in the EU say decides their search ranking: the main factors they name, whether paying raises a listing, and how they treat their own goods against outside sellers' goods. Each cell quotes and links the terms it comes from, and each quarter's changes are published as a list of differences.
The briefCredit · unclaimed
Open 01 · the brief
Spot deals leaving the platform
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open, labeled set of at least 5,000 made-up chat threads between buyers and sellers and between clients and freelancers, in which some people try to move the deal or the payment elsewhere: phone numbers spelled out in words, handles split across messages, lookalike characters, offers of a discount to "pay me on another app." Every name, number and handle is invented. Publish it with baseline scores for a keyword list, a small trained classifier and two large language models.
Why it is still open
Gu and Zhu's trial shows the pull to leave a platform grows with trust. A chat-moderation vendor wrote in November 2026 that keyword lists miss disguised requests, but vendors test their filters on private data, and searches of GitHub and Hugging Face in November 2026 found no public labeled set of such messages.
Built with what exists
Claude, Gemini or an open-weights model such as Mistral Large 4 (Note: "A trillion open parameters") to draft threads; Label Studio for hand labels; scikit-learn; a small open model such as Liquid LFM3 for the trained baseline (Note: "One GPU to fine-tune"); Hugging Face Datasets to publish.
First step
Write a two-page labeling guide with 20 evasion patterns, generate 200 threads, and have two people label them separately.
Done looks like
5,000 threads, 1,000 of them labeled by two people with agreement of at least 0.8 on Cohen's kappa, a standard measure of agreement between labelers; a further 500 threads written by volunteers told to slip past the filters; published precision and recall for each baseline on threads held out from training.
Open 02 · the brief
Ranking rules, side by side
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public table, refreshed every quarter, of what 40 marketplaces and freelance platforms that serve sellers in the EU say decides their search ranking: the main factors they name, whether paying raises a listing, and how they treat their own goods against outside sellers' goods. Each cell quotes and links the terms it comes from, and each quarter's changes are published as a list of differences.
Why it is still open
Since 12 July 2020 the EU's Platform-to-Business Regulation (2019/1150) has required these services to set out the main parameters of their ranking in their terms. The European Commission's review of 12 September 2023 found that only about a third of the terms studied did so, mostly "in rather general terms," gave only totals, and said business users "may currently lack a frame of reference to measure the degree of transparency they receive." Francesco Sovrano, Michaël Lognoul and Alberto Bacchelli checked the ranking documents of six platforms once, in a study posted in December 2023 (ICSE 2024), and found large differences between them, but only three of the six (Amazon, Booking and Tripadvisor) are marketplaces, and the check has not been repeated. A search in November 2026 found no public table of marketplaces and freelance platforms kept up over time.
Built with what exists
Open Terms Archive, free and open-source software that records every version of a service's terms, to fetch and store the documents; Browser Use 1.0 for pages behind menus (Note: "Deleting 80% to start over"); a language model to pull out quoted fields, checked by a person, as Sovrano and colleagues tested with ChatGPT; CSV files on GitHub.
First step
Save the ranking sections of 10 platforms' current terms and draft the table's 12 fields.
Done looks like
40 platforms with every field quoted and linked; two people working separately agree on at least 90% of fields for a sample of 10 platforms; two quarterly releases with their differences.