Fiverr Shed Buyers as Upwork Tracked Higher-Value Work
At 8:30 a.m. Eastern on July 29, Fiverr executives opened a quarterly earnings call with a marketplace problem. The company had fewer buyers, less revenue, and a weaker outlook than it had a year earlier. Founder and CEO Micha Kaufman was describing artificial intelligence as more than another category of work to sell. It was also pressure on the traffic and transactional jobs that had helped build Fiverr.
The figures in Fiverr’s second-quarter release made that pressure visible. Annual active buyers fell from 3.4 million to 2.7 million, a decline of 21.9%. Marketplace revenue fell 15.5%. Total revenue fell 10.0%. The company revised its full-year outlook and said recent AI-related demand and traffic headwinds had continued into the third quarter, alongside persistent weakness in categories most exposed to automation.
Then the same release pointed in another direction. Annual spend per buyer rose 15.6%, from $318 to $368. The number of clients completing projects worth at least $1,000 rose 13% over the trailing 12 months. Kaufman framed Fiverr’s response as a move from a transaction-oriented marketplace toward a work platform for higher-value projects.
Fifteen days earlier, Upwork had published a different view of the same pressure. Its Future Workforce Index 2026 found that generative AI and creative-production contract starts on its marketplace rose 90% year over year while earnings per contract fell 13%. At the other end of its classification, earnings for freelancers doing more complex AI-augmented work rose 45%.
Fiverr reported company financial and marketplace metrics through June 30. Upwork analyzed its own first-quarter job data and paired that analysis with a survey. Their definitions, time periods, customer mixes, and business models differ, so the disclosures support neither a platform ranking nor a measure of the whole freelance economy.
Read side by side, the disclosures expose a practical conflict for anyone buying or selling independent work. Cheap execution can become more abundant while losing value per contract. Larger projects can grow while the pool of active buyers contracts. A marketplace can announce an upmarket strategy; a worker still needs paid evidence to qualify for that work.
The move from output to outcome changes the brief, the evidence, the price, and the person who carries the risk. It also removes a familiar first rung. If routine projects once helped new freelancers earn a rating, learn a client domain, and build a portfolio, buyers and platforms need to decide who will pay for that learning when the market asks for senior judgment at the first serious engagement.
Two dashboards, one changing work market
Fiverr’s buyer figure can sound like a headcount. Its definition is narrower. An annual active buyer is an account that ordered a Gig on Fiverr during the previous 12 months, whether or not the order was later canceled. The metric fell from 3.425 million on June 30, 2025, to 2.676 million one year later. That is a decline of 749,000 accounts inside a rolling activity measure, with no direct count of business closures, freelancer losses, or AI-caused departures.
Annual spend per buyer has another boundary. Fiverr calculates it by dividing marketplace gross merchandise value over the previous 12 months by the active-buyer count on the measurement date. The increase to $368 says that the remaining pool, as measured, carried more spending per buyer. Distribution across buyers and any change in a typical freelancer’s earnings remain unknown.
The $1,000 project signal is narrower still. Fiverr said clients completing projects of at least that size grew 13% on a trailing 12-month basis. The metric tracks clients and leaves the number of projects, average duration, margin, and freelancer distribution unreported. It is an early indicator for Fiverr’s strategy, well short of proof that larger work has replaced the lost transactional volume.
The financial statement provides a useful check. Marketplace revenue was $63.1 million in the quarter, down from $74.7 million. Services revenue, which comes from Fiverr’s additional services rather than transactions between buyers and freelancers on Fiverr.com, rose 2.0% to $34.6 million. The platform can therefore add services and pursue larger engagements while its core marketplace remains under pressure. An upmarket product plan and a completed financial transition are different things.
Fiverr’s revised guidance makes the timing costly. It projected third-quarter revenue of $80 million to $88 million, a year-over-year decline between 18% and 26%. Its full-year range of $356 million to $372 million implied a decline between 14% and 17%. CFO Esti Levy-Dadon called the period an early stage of the transition and paired investment in the upmarket move with cost discipline. The company published no causal share for buyer or revenue losses attributable to AI.
Upwork’s dashboard begins somewhere else. The report, written by research leader Jennifer Brett and economist Teng Liu, analyzed marketplace jobs from the first quarter of 2026 against the same period in 2025. It identified AI-related work from job titles and descriptions with an LLM-based pipeline, mapped jobs into task clusters, and allocated contract volume and earnings with relevance weights. The placements were validated through content analysis and checked against LLM-generated task-cluster annotations.
Upwork states the limitation directly in its methodology: measuring AI work contains substantial uncertainty, and its categories may include errors. That matters because phrases such as “complex AI-augmented work” and “AI execution” sound like settled occupations. They are analytical groupings built from platform records. A freelancer may perform work in several groups within one contract, and a buyer may never use those labels in a brief.
The accompanying survey covered 2,400 U.S.-based skilled workers in March and April 2026. It limited the sample to people working above the administrative level and above an earnings threshold derived from Bureau of Labor Statistics data. Its reach ends with that selected population, far short of every gig worker, country, or buyer represented in Fiverr’s rolling account metric.
Keep the two releases beside each other without merging them. Fiverr shows a business confronting fewer active buyers and shrinking marketplace revenue while larger-project indicators improve. Upwork shows a platform-specific distribution of contract starts and earnings across categories its researchers constructed. One is a company operating result. The other is a labor-market signal from a marketplace and a selected survey.
Both companies are organizing their 2026 story around higher-value work, although their disclosures lack a shared magnitude. Buyers need to unpack that phrase before putting it into a procurement plan. Freelancers need to unpack it before building a career plan around it.
Fewer buyers, more spend per buyer
Fewer buyers alongside higher spend can signal concentration rather than simple health or weakness. Some occasional customers may stop ordering cheap tasks. Some retained customers may purchase broader projects. New services may move revenue outside the core buyer-to-freelancer transaction. Averages can rise because the denominator changed even when aggregate activity falls.
For Fiverr, the buyer decline was larger than the increase in spend per buyer. Marketplace revenue fell, and free cash flow fell from $25.0 million to $13.6 million. The company still produced positive cash flow and GAAP net income, but the operating picture forced management to invest in a different matching and delivery model while maintaining profitability.
One visible part of that investment is Fiverr’s proprietary Knowledge Graph. The company said it had deployed upgrades to improve matching quality and project outcomes for higher-value work. A taxonomy that once matched a buyer to a listed service now has to infer a broader problem, locate connected capabilities, and support a longer engagement. The product burden grows with the project.
Consider a buyer who once ordered a fixed set of product descriptions. The acceptance test was easy to state: count, length, tone, keywords, and deadline. A model can now draft that batch in minutes. The buyer may still hire a freelancer, but the valuable scope has shifted toward selecting products, diagnosing weak conversion, connecting catalog data, setting editorial rules, reviewing claims, running an experiment, and explaining what changed.
This broader scope reallocates responsibility. The buyer gives more context and access. The freelancer makes more decisions. The engagement lasts longer. Failure may touch customer data, brand promises, campaign spend, or a production workflow. A $1,000 threshold can capture a larger invoice while revealing nothing about whether those responsibilities were specified well enough to produce a good outcome.
The shift also changes acquisition economics. A small buyer can test a marketplace with a cheap, legible task. That first purchase creates trust at limited risk. If the useful entry point becomes a multi-week workflow project, the buyer must assess a stranger’s judgment before seeing the work. Fiverr must provide stronger matching, proof, and recourse. The freelancer must invest more time before a contract begins. The marketplace has to prevent that pre-contract work from becoming an unpaid consulting layer.
Higher spend per buyer can therefore coexist with a harder funnel. The platform wants fewer one-off transactions and more durable relationships, but larger commitments require more evidence. A buyer needs examples tied to business problems instead of a gallery of outputs. A freelancer needs enough context for diagnosis while keeping uncompensated proposal work within a reasonable boundary.
Missing from the company release are proposal hours, freelancer concentration, repeat purchase by project size, dispute rates, and earnings by experience level. Those omissions leave the workforce burden outside the financial picture.
A talent leader or founder should first decide which kind of work is being purchased. A brief for a bounded output, a diagnostic engagement, a system integration, and an operated business process require different matching signals and contracts. Calling all four “freelance help” hides the budget and the risk owner.
For a freelancer, the same distinction determines what to learn. Faster production may preserve margin for a while, then lose its advantage as competitors reach the same output with the same tool. Defensible evidence moves closer to problem selection, domain constraints, integration, review, and the business result that survived after delivery.
Contract volume rises while execution pay falls
Upwork’s most useful finding lies in the divergence between volume and value inside AI-related work.
Across its marketplace data, freelancers performing AI work earned 34% more per hour than freelancers not incorporating AI. That broad premium can attract workers and buyers. Yet generative AI and creative-production work recorded 90% growth in contract starts while per-contract earnings fell 13%. Upwork also reported a 28% decline in earnings for AI-based execution tasks overall.
Contract starts differ from completed contracts, just as earnings per contract differ from hourly wages. A market can produce many more small engagements while each engagement pays less. Total marketplace earnings can still rise, remain flat, or fall depending on the starting values, completion, distribution, fees, and how much work each person wins. Upwork published too little in the release to calculate an individual worker’s outcome from those two percentages.
The pattern still matters. When a tool lowers the cost of producing a discrete output, buyers can commission more experiments. A marketing team may order several concepts instead of one. A founder may test multiple prototypes. A creator may hire for shorter editing passes. More contracts appear, but each carries less budget because the buyer expects faster production and can compare more substitutes.
That market rewards speed and tool fluency, but it can punish undifferentiated execution. A worker who cuts production time in half may initially earn a higher effective hourly rate on a fixed-price job. If buyers then halve the price, expand revision expectations, or split the work among more suppliers, the gain moves. Who keeps the productivity benefit depends on the brief, competition, platform design, and bargaining power.
Upwork’s complex-work figures point in the other direction. Earnings for more complex AI-augmented work rose 45% year over year. AI-augmented professional services grew 72% in volume and 22% in earnings. Brett describes an emerging “AI orchestrator” who connects tools to domain expertise, manages workflows, evaluates quality, and owns a business outcome.
Stanford economist Nick Bloom, a member of Upwork’s Economic Advisory Council, connected that distribution to a broader productivity puzzle. A separate survey he cited, covering nearly 6,000 executives, found little measurable AI impact at most firms. In Upwork’s data, the gains gathered in complex work where expertise, judgment, and business context sit on top of the tool.
That label is useful as a description, but dangerous as a shortcut. Writing “orchestrator” on a profile creates no extra value by itself. The buyer must have a problem that requires coordination and judgment. The worker must show evidence of those capabilities. The contract must give the worker authority, data, time, and access. Otherwise, orchestration becomes extra responsibility attached to an execution budget.
Potential supply is rising at the same time. In Upwork’s selected U.S. survey, 58% of full-time employees said they were considering freelancing, up from 36% a year earlier. Intention differs from opening an account or completing a contract. Still, more experienced employees may arrive with domain knowledge just as prices compress for the easiest projects on which a newcomer could build marketplace proof.
The split also appears within one profession. A designer may sell individual images in one contract and lead a brand-system migration in another. A developer may produce a script, then diagnose an unreliable support workflow and integrate several systems. A recruiter may generate outreach copy, then redesign a sourcing and interview process. People can move between levels of complexity as the unit of work changes.
That makes category-level advice incomplete. Telling freelancers to enter AI work ignores the 13% decline in earnings per contract in a fast-growing group. Telling buyers to purchase outcomes ignores situations where a bounded output is exactly what they need. The operating task is to match scope, evidence, and price to the level of responsibility.
A small fixed deliverable can remain valuable when its acceptance criteria are clear and its context is contained. It becomes exploitative when the buyer quietly expects diagnosis, strategy, integration, and open-ended correction inside the output fee. A larger outcome contract can pay well and still fail when nobody defines the baseline, supplies access, or accepts responsibility for decisions the freelancer cannot make.
Work packages are separating instead of moving along one clean line from cheap human work to expensive human judgment. Some outputs become cheaper and more plentiful. Some judgment becomes more valuable. Coordination can also be transferred to the worker without corresponding authority or pay. The percentages reveal the split; contracts decide where each project lands.
Higher-value work changes the job
The old marketplace brief began with a noun: logo, article, spreadsheet, video, list, translation, landing page. The new upmarket brief often begins with a business condition: conversion is weak, the support queue is growing, reporting takes three days, leads are poorly qualified, or an internal workflow fails at handoff.
That change sounds small. It moves work from production into diagnosis. A buyer ordering a noun can inspect the output. A buyer presenting a condition needs the freelancer to ask whether the stated problem is the real one, choose an intervention, work across systems, and measure whether the condition improved.
As the job expands, context becomes part of the input. The freelancer may need customer research, analytics, process history, brand constraints, or examples of failed work. The project can cross tools and teams. Quality turns into a series of checkpoints. Meanwhile, the worker may be held to an operating result affected by decisions outside the contract.
AI makes this expansion easy to underestimate. Production appears faster, so a buyer assumes the whole engagement should be cheaper. Yet faster drafts can create more options to evaluate. An automated workflow can require access reviews, exception handling, monitoring, and recovery. A generated analysis can move the difficult work into source verification and decision-making. Execution time falls while coordination time rises.
The buyer’s role changes too. A hiring manager who outsources the definition of success leaves the supplier without essential business context. Someone inside the company must own the baseline, grant access, resolve conflicting priorities, review consequential decisions, and state what happens when the data or model is wrong. That work belongs in the budget even when no employee logs it as procurement time.
For talent teams, independent work starts to resemble a capability decision rather than a task purchase. Should the organization rent expertise for one diagnosis, retain a freelancer through implementation, hire the person, or develop the capability internally? The answer depends on recurrence, data sensitivity, integration depth, and how much institutional knowledge the work creates.
A one-time market-research synthesis may remain a freelance deliverable. A pricing system that must be updated every week needs an operating owner. An agent touching customer records requires permissions, monitoring, and incident response. A campaign concept can be accepted by a creative lead. A promise about revenue improvement requires agreement about attribution. The higher the claimed outcome, the more the buyer must expose its own operating system.
Platform matching becomes harder as this responsibility grows. Reviews based on past outputs can help with execution work. They are weaker evidence for diagnosis in a new domain. A large project needs proof of how the freelancer framed a problem, handled missing data, communicated uncertainty, managed stakeholders, and left the client able to operate after departure.
Fiverr’s Knowledge Graph may improve discovery across related skills. Upwork’s task analysis may improve the language used to describe complex work. Those product moves leave the paid discovery phase untouched. Matching can suggest who might fit; shared understanding still has to be built inside the engagement.
The contract should reflect that uncertainty. Instead of asking five freelancers for a complete strategy in a proposal, a buyer can pay one or two for a bounded diagnosis. That phase can test access, reasoning, communication, and evidence. It can produce a scoped implementation plan without pretending the final cost was knowable before the systems were opened.
For the freelancer, paid diagnosis is also career evidence. It shows more than an attractive output. It records assumptions, tradeoffs, rejected options, and the connection between work and result. Those artifacts become the basis for moving toward higher-value projects. Without them, “move upmarket” is a demand to arrive with experience that no client was willing to fund.
The missing rung between output and outcome
The career problem sits underneath both marketplace stories. Routine work is often low paid and uneven, but it can give a new worker what courses struggle to provide: a real brief, an imperfect client, revision feedback, a deadline, and a result that can be shown. Tools may absorb more of that production while platforms prioritize larger projects. The need to learn remains, with no obvious party funding it.
PwC’s 2026 Global AI Jobs Barometer found a related pattern outside freelance platforms. Its analysis covered more than one billion job advertisements across 27 countries and territories. In a subset of 2.4 million U.S. entry-level postings, the roles most exposed to AI were seven times more likely to ask for skills traditionally associated with seniority, including judgment and leadership.
PwC also reported that these “seniorised” entry-level roles grew 35% from 2019 while other entry-level roles declined 10%. The finding concerns job advertisements and required skills, leaving completed hires, promotions, and employer-funded training outside the dataset. Pete Brown, PwC’s global workforce leader, focused on the apprenticeship problem: routine work once developed judgment, while exposed entry-level roles now ask for it earlier.
Upwork’s orchestrator description has the same unresolved step. Domain expertise, workflow design, commercial judgment, communication, and accountability are valuable because they take time to build. A marketplace can identify demand for that bundle while leaving workers to acquire it alone between contracts.
Deel’s State of Global Hiring Report shows how wide one new work category has already become. Based on more than one million contracts across over 37,000 companies in more than 150 countries, Deel reported that cross-border hiring for general AI trainers grew 283% in 2025. The occupation covered more than 70,000 workers at over 600 organizations, ranging from basic annotators to subject-matter experts in medicine and economics.
That label contains several possible career ladders. Annotation, translation, economic analysis, medical review, safety evaluation, and model behavior research carry different entry requirements and liabilities. Deel’s customer data leave progression between levels unmeasured. A fast-growing title can conceal separate markets that share a name but lack a common route upward.
An institutional program offers a useful contrast. Anthropic’s Claude Corps will place its first cohort of 100 early-career fellows in U.S. nonprofits in October 2026. Fellows are full-time employees of CodePath, receive an $85,000 salary and benefits, and spend five hours each week in ongoing training during the 12-month program. Anthropic plans three cohorts totaling 1,000 fellows.
Career outcomes from Claude Corps are still pending, and its geography, cohort size, eligibility, and nonprofit model limit any wider inference. Its relevance is structural: the program names the employer, salary, training time, mentor, host project, and duration. It pays people while they connect tool fluency to an operating problem. Those are the missing inputs behind a generic instruction to become an AI orchestrator.
A year-long fellowship sits beyond most freelance buyer budgets. Smaller choices can still share the learning risk: a paid trial, access to a domain reviewer, staged responsibility, and feedback the freelancer may preserve without exposing confidential data. A platform can distinguish verified project participation from self-written skill claims. An experienced freelancer can be paid for mentorship rather than expected to absorb it inside delivery.
A route upward should preserve the worker’s diagnosis as paid work. A speculative sample may be reasonable for a narrow creative test when its use is limited and clearly disclosed. Asking candidates to audit a live funnel, design an automation plan, or solve an internal process before selection extracts the highest-value part of the engagement without a contract. Upmarket demand can make unpaid proposal work more sophisticated and more exploitative.
Keep the first rung small in consequence while preserving real judgment. A junior freelancer can own a bounded research slice, document assumptions, compare outputs, run a controlled test, or maintain one part of a workflow under review. The buyer gets evidence. The worker gets practice. The reviewer remains accountable for decisions beyond the worker’s access or experience.
Picture a video editor who has delivered dozens of clean assets but has never been allowed into a client’s campaign analytics. A request for “performance creative strategy” now asks that person to diagnose audience fatigue, choose tests, and promise acquisition results. Editing ability is already proven. Experience with data and decisions remains missing because earlier contracts kept both outside the worker’s reach. A paid diagnostic milestone can expose that boundary without turning the whole campaign into an audition.
That arrangement costs more than ordering an isolated output and hoping for senior judgment. It also reduces the risk of awarding a broad outcome contract to someone whose profile has never been tested in context. Training, selection, and delivery share one budget when valuable work depends on knowledge learned inside the engagement.
A freelance value ladder
Buyers and workers need a common way to name the responsibility inside a project. Treat the ladder below as a scoping tool, not a salary scale or universal progression. A person may work on different rungs in different domains, and a project may move upward only after evidence supports the change.
| Rung | Buyer is purchasing | Acceptance evidence | Worker must be able to show | Paid bridge to the next rung |
|---|---|---|---|---|
| Bounded output | A specified asset produced from supplied inputs | Format, accuracy, count, deadline, and disclosed tool use | Reliable execution against a clear brief | A review session that explains errors, choices, and domain context |
| Reviewed deliverable | An output plus verification and revision | Source trail, test cases, review notes, and resolved defects | Judgment about quality, not only production speed | A paid comparison of alternatives with a named reviewer |
| Diagnosed problem | A finding about why an operating condition exists | Baseline, evidence, assumptions, rejected explanations, and recommended scope | Problem framing under incomplete information | A bounded discovery contract with access to stakeholders and data |
| Integrated workflow | A repeatable process connected to people and systems | Handoff success, exception path, permissions, monitoring, and recovery test | Coordination across tools, owners, and failure modes | Staged implementation with an internal owner and controlled access |
| Accountable outcome | An agreed business result over a stated period | Outcome metric, attribution boundary, quality floor, cost, and post-delivery owner | Commercial judgment and responsibility proportionate to authority | Shared-risk terms only after a paid baseline and operational trial |
A bounded output can carry real craft and fair pay. Its advantage is legibility. Both parties know what has been supplied, what will be delivered, and who decides whether it passes. A scarce specialist may command premium rates on this rung without becoming an integrator. AI use can be permitted, prohibited, or conditioned on review without making the worker responsible for the buyer’s business outcome.
The second rung makes quality work visible. A buyer who wants verified research, tested code, edited video, or reconciled data should pay for the verification. “Human reviewed” is not an acceptance criterion. The contract should name the sources, tests, reviewer, error threshold, and revision limit. The worker’s evidence is the reasoning that survived review, not a claim that a human was somewhere in the loop.
Diagnosis deserves its own contract because it changes the project. The worker needs enough access to distinguish a content problem from a distribution problem, a model problem from a data problem, or a recruiting problem from a compensation problem. The buyer needs to see how the worker handles ambiguity before approving implementation. Folding discovery into an unpaid proposal encourages confident answers based on missing evidence.
Integration moves the risk into the organization. Credentials, APIs, customer records, internal policies, and employee workflows may enter scope. Late access, unresolved executive decisions, and systems controlled by another vendor stay with the buyer. The contract needs an internal owner, a test environment where possible, an exception path, and a handoff that leaves someone able to operate the workflow.
Outcome pricing belongs at the top only when authority reaches the same level. A revenue guarantee requires influence over product, traffic, price, inventory, sales response, and measurement. Shared-risk terms can work after a baseline and trial establish what each party controls. Otherwise, the outcome language transfers uncertainty to the smaller party while the buyer keeps the decision rights.
This ladder also helps a marketplace interpret the upmarket move. More $1,000 projects say little unless the platform can distinguish a larger batch of outputs from a diagnosed or integrated engagement. Search tags provide weak proof on their own. Verified scope, milestone acceptance, repeat operation, and client-side ownership can improve matching without exposing confidential deliverables.
For workforce planning, the paid bridge column is the important one. It asks how someone gains the evidence needed for the next responsibility. A review session, paid comparison, discovery contract, staged implementation, or operational trial creates learning inside a commercial engagement. The buyer receives lower-risk evidence while the worker avoids financing the entire market transition through unpaid practice.
Pricing follows responsibility more reliably than tool novelty. A freelancer may use AI on every rung. The fee should reflect scope, skill, time, access burden, review, liability, and the value of the decision. Charging more because a project contains AI is as weak as charging less because AI made drafting faster. Buyers should price the work and risk that remain with the human.
Monday’s brief carries a different promise
On Monday morning, imagine a marketing manager opening a marketplace because a campaign has stalled. The old brief asks for 20 short videos by Friday. It supplies footage, dimensions, examples, and a revision limit. That can be a fair bounded-output contract. Channel strategy, performance diagnosis, and a conversion guarantee belong in a wider scope with a wider budget.
An upmarket brief starts differently. It states that paid-social acquisition cost rose over six weeks, identifies the available campaign and creative data, names the channels the freelancer may change, and admits what the team does not know. It funds a diagnostic phase before promising a production volume. A company reviewer is available. Customer data stays inside approved systems. The first milestone delivers a tested explanation before an asset quota.
The freelancer’s proposal changes as well. It avoids promising that more AI-generated variants will fix the campaign. It names the evidence needed, the hypotheses worth testing, the decisions that remain with the client, and the work that can be completed inside the first milestone. It shows one prior case without pretending that another client’s result will transfer intact.
If the diagnosis survives review, the next contract may include a production system, experiments, and weekly learning. A failed diagnosis still leaves the buyer with a documented finding and pays the freelancer for real judgment. Both parties can stop before integration risk and outcome promises exceed the evidence.
By Monday afternoon, the manager has a choice. The company can post the 20-asset brief honestly, or it can fund the diagnosis behind the stalled campaign. Fiverr’s shrinking buyer metric and Upwork’s earnings split cannot make that choice. They show why disguising one kind of contract as the other has become more expensive.
The marketplace product, the buyer’s brief, and the worker’s career path now meet at the same line item: paid evidence. Platforms need to see larger projects become durable activity and task-level value reach workers. Buyers need proof before widening scope; freelancers need paid projects that let them acquire and preserve proof of judgment.
Before the manager posts the brief, one field should remain impossible to skip: who is paying for the learning between this rung and the next? If the answer is nobody, the scope is incomplete. It has moved the cost to the person with the least information and the least bargaining power.