TL;DR. The first client is the hardest one because a buyer cannot tell a capable newcomer from a weak one, not because newcomers lack skill. On oDesk, a predecessor of today’s Upwork, ten percent of new workers outside the US without an agency behind them found at least one job, against 27 percent of those who entered under an established agency. A 2021 estimate puts the share of the world’s 163 million registered freelance profiles that have ever obtained work at 8.6%. A field experiment shows what lowers the barrier: one small paid job with a public rating took inexperienced workers from 12% to 30% employed within two months. A platform skill test, by contrast, produced one new project per approximately 63 credentials. Since 2023 there is a second problem: after an AI proposal writer launched on Freelancer.com, the link between a tailored cover letter and a callback fell by 51%. The CEO move: treat the first client as the buyer’s risk problem, and spend scarce hours on the proof that removes the most risk per hour. The Student move: run every piece of proof as a practice rep with a deadline, a standard and written feedback.
This piece sits between what actually signals skill in 2026 and the channel economics of Upwork, Fiverr and direct clients. If you are still deciding whether to go independent, start with the freelancing versus full-time decision framework. Once a client says yes, freelance pricing models compared covers how to charge.
The seven strategies are a CEOtudent editorial framework. CEOtudent ran no survey, holds no client data and tested none of them; each is graded by the public research behind it. This article is general information, not personalised financial, legal or tax advice.
Why is the first client the hardest one to get?
Amanda Pallais’s explanation in the American Economic Review is that a first job produces two things: output for the buyer and public information about the worker. The worker keeps the benefit of the information while the buyer carries the cost of producing it, so buyers hire too few beginners. Table 1 shows that barrier and what lowers it.
Table 1. What field experiments and platform studies measured (figures as printed in each source)
| Study | Setting and sample | What was tested or compared | Measured effect |
|---|---|---|---|
| Pallais (2014) | oDesk, 2010; 3,767 low-wage data-entry applicants, 952 hired at random for 10-hour jobs | A small paid job plus a public rating | Inexperienced workers with any job in the next two months: 12% to 30%; average earnings $10 to $27 |
| Stanton and Thomas (2016) | oDesk, August 2008 to December 2009; 83,029 new workers outside the US | Entering under an agency run by an established worker | 27 percent of affiliates found at least one job against ten percent of non-affiliates; median five days and three applications to the first hire against thirteen days and ten |
| Agrawal, Lacetera and Lyons (2016) | oDesk, January 2012; 14,617 job postings, 420,833 applications | Verified platform experience above the median of 4 prior contracts | Predicted chance of being hired per application: 6.7% to 8.3% for applicants in developed countries, 1.6% to 2.8% in less developed ones |
| Kässi and Lehdonvirta (2024) | A large freelance platform; 46,791 workers over nine years | Passing a platform skill test (a microcredential) | Earnings up 8.9 percent over two weeks; one new project per approximately 63 microcredentials |
Sources: the named papers; working paper versions were consulted where the Sources list says so, and the Stanton and Thomas figures are from the 2014 discussion paper.
Three readings matter. First, the barrier is information: in Pallais’s experiment, being hired changed nothing while the market could see only the hire, and outcomes improved once the evaluations became public. Second, someone else’s reputation is the fastest substitute for your own. Third, a certificate moves price more than entry: Kässi and Lehdonvirta conclude that microcredentialing “does not appear to be very useful to new workers struggling to land their first job on the market”.
Table 2. How steep is the cold start? (CEOtudent analysis of the named sources)
| What was measured | Figure in the source | CEOtudent calculation |
|---|---|---|
| Inexperienced workers with any job within two months: control group, then after one small paid job | 12% and 30% (Pallais) | About 1 in 8; 2.5 times |
| New workers who found at least one job, without and with an agency | Ten percent and 27 percent (Stanton and Thomas) | 2.7 times |
| Registered freelance profiles that have ever obtained work | 14 million of 163 million (Kässi, Lehdonvirta and Stephany) | 8.6%; the authors’ range is 2.0% to 36.8% |
| Applications per job posting | 420,833 applications, 14,617 postings (Agrawal and co-authors) | 28.8 |
| Applications per hire, low-experience applicants | Hire rates of 1.6% to 6.7% (same study) | About 15 to 63 |
| A marketplace buyer’s yearly spend | Fiverr $342 per buyer; Upwork $5,129 of gross services volume per active client | The Upwork figure is 15 times the Fiverr one |
| Buyers on two SEC-reporting marketplaces | Fiverr 4,027 thousand (2023) to 3,135 thousand (2025); Upwork 851 thousand (2023) to 763 thousand (June 2026) | Down 22.2% and 10.3% |
CEOtudent analysis; derived cells were recomputed by script. The rows describe different platforms and years and do not add up to one rate. Applications per hire assumes every application has the same independent chance, an illustration only. Upwork and Fiverr metrics are company-reported in SEC filings.
Read it as an order of magnitude: in three datasets collected between 2008 and 2020, roughly one newcomer in ten gets any work. The buyer pool on both marketplaces in the last row is shrinking. Our channel comparison cited 785 thousand Upwork active clients at December 2025; the report for the quarter ended June 30, 2026 shows 763 thousand and links the decline partly to “the evolving impact of AI on certain categories of freelance work and on new client acquisition and retention”.
What is a buyer actually uncertain about?
A buyer with no history to read is pricing four separate risks. Each needs a different kind of proof, which is why one portfolio rarely answers all of them.
Table 3. The four buyer risks and the proof that answers each (CEOtudent editorial framework)
| Risk | The buyer’s unspoken question | Why it is real | Proof that reduces it | What does not |
|---|---|---|---|---|
| Skill | Can this person do this task to my standard? | Job-specific measures predict performance best in hiring research | A sample on a brief like theirs; a scored test | Years of experience; adjectives |
| Reliability | Will it arrive on time without chasing? | In Pallais’s experiment approximately 18% of hired workers did not return the file or log any hours | A small paid job delivered early; a person who vouches | A sample made with no deadline |
| Fit | Do they understand my problem? | Before AI writing tools, a proposal one standard deviation more tailored was worth as much as a $26 lower bid | Checkable observations about the buyer’s own situation | Tailored wording alone |
| Communication | Will working with them cost me time? | Tests show hard skills “but not soft skills and general trustworthiness” (Kässi and Lehdonvirta) | Visible process: questions, plan, updates | Polished prose |
CEOtudent editorial framework; the third column cites studies in the Sources list.
Hiring research supports the logic of proof-of-work. In the 2022 revision of the classic validity estimates by Paul Sackett and co-authors, work sample tests predict job performance at .33 and structured interviews at .42, against .19 for unstructured interviews and .07 for years of job experience. The authors note that their strongest predictors “are all job-specific measures”. Two cautions: the .33 is the corrected figure (the widely quoted .54 from a 1998 summary was revised down by .21), and these studies concern employee selection, not freelance buying.
Has generative AI made a polished portfolio worth less?
For written proposals there is direct evidence, all from Freelancer.com and all in working papers that are not yet peer reviewed. Anaïs Galdin and Jesse Silbert find that employers paid for tailored applications before large language models, and that afterwards those signals “no longer predict successful job completion”. In their simulation of a market where writing costs nothing, workers in the top fifth of ability are hired 19% less often and those in the bottom fifth 14% more often. Jingyi Cui, Gabriel Dias and Justin Ye study the platform’s AI Bid Writer, launched on April 19, 2023, across 5 million cover letters. The correlation between cover-letter tailoring and a callback fell by 51%, the correlation with an offer fell by 79%, and the correlation between callbacks and review scores rose by 5%. Time spent editing the AI draft was positively correlated with hiring success.
Star ratings do not fill the gap: on one large online labour market the share of workers receiving the top rating of 5 stars “went from 33% to 85% in just 6 years”, and the authors attribute more than 50% of the increase in scores to inflation. Clean writing is not the problem either. In an experiment with nearly half a million new jobseekers on an online labour market, those randomly offered algorithmic writing help were hired 8% more often, and the authors found no evidence that employers were less satisfied. Clarity helps a buyer read you; it no longer proves effort.
No study we found tests portfolios directly, so the next step is an argument, not a finding: if cheap text lost its signal, cheap polished samples probably lose theirs too. In an ILO survey, 79 per cent of freelance-platform respondents named reviews and feedback from previous clients as key to obtaining new work, 54 per cent named previous work portfolios and 53 per cent the number of completed jobs. Proof that still works has three properties: someone other than you can verify it, it shows process as well as output, and it is specific to the buyer. Professional reputation in the AI era develops that argument.
Which seven proof-of-work strategies have evidence behind them?
The grading rule is fixed in advance. A: a randomised field experiment in a freelance market tested the mechanism, with getting hired or earning as the outcome. B: tested by large quasi-experimental or observational studies in freelance markets, or by experiments and meta-analyses in employee hiring that transfer by analogy. C: evidence exists but is indirect, mixed or small for newcomers. D: no study found; editorial reasoning only.
Table 4. Seven proof-of-work strategies (CEOtudent editorial framework; time costs are editorial estimates, not data)
| Strategy | What it proves | What the buyer sees | Time cost | Evidence grade and basis | Main risk | When not to use it |
|---|---|---|---|---|---|---|
| 1. Paid micro-pilot | All four risks, at small scale | A finished paid job and a rating or reference | 2-4 hours to pitch, 3-10 hours of paid delivery | A: Pallais, 12% to 30% | Anchoring a low rate | The problem cannot be cut small |
| 2. Borrowed credibility | Reliability and fit, vouched for | A referral, or an established name on the work | 3-6 hours to list and message contacts | B: Stanton and Thomas; Pallais and Sands | The proof stays with the intermediary | You cannot name the work you want |
| 3. Reproducible work sample with visible process | Skill and communication | Brief, decisions, drafts and files to inspect | 8-15 hours, reusable | B: Sackett and co-authors, by analogy | Process also exposes weak work | The skill cannot be judged from files |
| 4. Structured trial with a pre-agreed acceptance test | Reliability, with the buyer’s loss capped | A first milestone, a written test, a stop option | 1-2 hours to draft, 5-15 hours of delivery | C: Barach and co-authors; Agrawal and co-authors | A vague test invites disputes | The result cannot be tested objectively |
| 5. Public teardown of a prospect’s problem | Fit and skill | Three observations to check in minutes | 1-2 hours per prospect | C: tailoring was valued before AI | Sliding into unpaid consulting | The diagnosis is the whole job |
| 6. Verified skill tests and microcredentials | Hard skill only | A scored third-party result | 2-6 hours per test | C for landing work: Kässi and Lehdonvirta | Collecting badges, not clients | Before the first job |
| 7. Documented self-initiated project with a measurable result | Skill and initiative | A live result with before and after numbers | 15-40 hours | D: no experiment found | Weeks with no buyer in sight | Strategy 1 or 2 is available |
CEOtudent editorial framework. Grades describe how well the mechanism is evidenced, not a success rate; no success rate exists for any of these.
1. The paid micro-pilot
Sell a job small enough that saying yes is cheap: one landing page, one cleaned dataset, one week of inbox triage. Keep your rate and shrink the scope. Write three acceptance criteria before starting, deliver ahead of the deadline, then ask for feedback that names specifics. In the research reviewed here, this is the only strategy tested directly in a randomised experiment with employment as the outcome, and the experiment shows where the value sits. The jobs ran a maximum of 10 hours at $3 an hour or less, so each paid at most $30, while the effect of one job on later earnings grew from $19.72 after two months to $65.56 after six: the average later gain was more than twice the largest possible fee. Detailed evaluations added to earnings, but they “helped those who performed well and hurt those who performed poorly”. On Upwork a new contract carries a talent fee of 0% to 15%, disclosed when the contract starts, and bidding requires tokens called Connects, which the company sells.
2. Borrowed credibility
Three routes exist: a referral from someone the buyer trusts, an introduction through a weak tie, and subcontracting for an established freelancer or agency. In Stanton and Thomas’s data agency heads kept around five to six percent of affiliates’ wages, and after reweighting for observable differences affiliates were still 75 percent more likely to find a job. Read the terms, though: agencies had “the option to remove all feedback from a worker’s profile” earned under their name, so agree in writing what you may show afterwards. Ordinary hiring points the same way. At one US corporation referred candidates were 6 percent of applicants and 29 percent of hires. In LinkedIn experiments covering over 20 million people, moderately weak ties, measured by mutual connections, created the most job mobility, so a former colleague you rarely speak to may help more than a close friend. And in field experiments on oDesk, Pallais and Sands found that “little to none of the information contained in the referral was otherwise observable to the employer”.
3. The reproducible work sample with visible process
Pick a public brief or public data close to the work you want. Publish the brief, your constraints, a time log, the options you rejected, the result and the files. State where AI tools were used and what you checked by hand. A buyer should be able to rerun the analysis, open the source file or load the page. This is a freelancer’s version of the work sample test, and it answers the AI discount because process is harder to fake than output.
4. The structured trial with a pre-agreed acceptance test
Offer a first milestone with a written test of done, payment on acceptance and the right to stop there. Do not expect the guarantee itself to win the job. When a platform backed selected freelancers with a money-back guarantee, buyers switched towards them but total contracts did not rise, and a plain recommendation worked as well: a recommended seller could bid about 16% higher and be equally preferred. Buyers wanted information more than insurance. Visibility has better support: in the Agrawal study, monitoring tools partially substituted for verified experience. Offer tracked hours, a shared folder and a midpoint check.
5. The public teardown
Choose five prospects in one niche. Using only public information, write three observations each that the buyer can verify in two minutes, and show one fix. Publish a general version and send the specific one. Buyer-specific effort was a signal employers paid for, and AI destroyed the text version of it, so a teardown works only when its substance is checkable.
6. Verified skill tests and microcredentials
Use these after the first job, not instead of it. Returns were highest for the first few credentials (up to 16 percent) and indistinguishable from zero after the 12th. The workers who gained most had “successfully broken into the market” but lacked a longer history; their marginal income gain was almost 15 percent. The study’s design advice doubles as a selection rule: tests should be “cheap” to take but difficult to ace.
7. The documented self-initiated project
Build one thing with a number attached: load time, error rate or hours saved, before and after. Make it public and checkable. It has the weakest evidence and the highest time cost, so use it when nobody in your network buys what you sell.
Where is the line on unpaid spec work?
Proof-of-work is yours and reusable. Spec work is custom work for one buyer’s commercial use, done in the hope of payment. Three editorial rules: cap unpaid buyer-specific effort at two hours per prospect; never hand over files the buyer could ship; and when a buyer asks for a test task, propose strategy 1 or 4 at a small price. Unpaid time is already built into platform work: ILO survey respondents on freelance platforms spent about 23 minutes on unpaid work for every paid hour. We found no study comparing paid with unpaid trial tasks for freelancers.
Which strategy should you start with?
Table 5. First-move chooser by skill type (CEOtudent editorial framework)
| Skill type | Hardest risk for the buyer | Start with | Add next | Avoid as a first move |
|---|---|---|---|---|
| Writing and editing | Skill, because fluent text is now cheap | 1, with a named reference | 3, showing edits and sources | 7 |
| Design | Fit and taste | 3 on a public brief | 1 | Custom concepts on spec |
| Development | Skill and reliability, visible only at the end | 2, subcontracting | 4, with automated tests as the acceptance test | 6 alone |
| Data and analysis | Skill and correctness | 3 on public data | 5 | 7 without a real question |
| Operations and admin | Reliability | 1 | 2 | 6 and 7 |
CEOtudent editorial framework. The development row follows Stanton and Thomas’s observation that agencies concentrate in technical job categories.
Hours decide the rest. Under five hours a week, do strategy 2 only, then 1. With five to ten hours, add 3. Above ten, add 5.
What does a 30-day first-client plan look like?
The plan assumes 10 hours a week, 40 in total. Its targets are editorial planning numbers, not benchmarks.
- Week 1. Name one buyer type and one small job. List 30 contacts and 10 prospects. Start the work sample.
- Week 2. Send the referral request to all 30 contacts. Publish the work sample. Write two teardowns.
- Week 3. Send three more teardowns and five micro-pilot offers. Offer one established freelancer a subcontract trial.
- Week 4. Deliver the first pilot early. Request specific feedback and permission to quote it. Log what the buyer changed, then repeat week 3.
If day 30 arrives with no pilot, change the buyer type before the method. When a pilot works, the retainer playbook covers what to offer next.
What should you actually say?
1. Referral request to a weak tie
“Hi [name], we worked together on [project] in [year]. I now take on [specific job] for [buyer type] as a freelancer. I am not asking you for work. Do you know one person who has [specific problem] this quarter? If so, I would be grateful for an introduction, and I will send two lines you can forward.”
2. Subcontract offer to an established freelancer
“Hi [name], I follow your work on [niche]. When you are over capacity, I can take [narrow task] off your plate under your name and your review. To make that safe I suggest one paid test piece of [size] at [price], with your acceptance criteria. Two requests: feedback I can learn from, and agreement on what I may show afterwards.”
3. Teardown plus micro-pilot pitch
“Hi [name], I looked at [public asset] and noticed three things you can check in two minutes: [observation 1], [observation 2], [observation 3]. I fixed the first on a copy: [attachment]. I can do the other two as a fixed job: [scope], [price], delivered by [date], accepted when [test]. If it misses the test, you do not pay for that milestone.”
4. Feedback request after delivery
“Thank you for the project. Could you write three sentences on [speed], [accuracy against the brief] and [communication], including one thing to improve? With your permission I will quote it with your name and company.”
The fourth script exists because evaluations with objective detail moved earnings more than a bare rating. If AI tools are part of your process, say so early; billing and disclosure for AI-assisted work covers the wording.
What are the CEO move and the Student move?
The CEO move is allocation. A newcomer’s scarce asset is hours, and Table 4 ranks where they buy the most risk reduction: borrowed credibility and a paid pilot before anything that takes weeks. A refusal is information about the buyer type or the size of the offer.
The Student move is to make each rep count twice. Set a standard before starting, time the work, get feedback from the person who paid or vouched, and log what was corrected. After three reps the log is itself proof of process. The cold start problem for digital products applies the same sequence to selling a product.
What does this evidence not mean?
It does not promise income. The strongest studies come from one platform, oDesk, between 2008 and 2013, and Pallais’s sample was low-wage data entry; effects on senior design or engineering work may differ. The agency result is observational. The AI studies are working papers on one platform and a few skill categories, and one headline result is a simulation. The validity coefficients come from employee selection. Time costs, grades and the 30-day targets are editorial. No source gives a success rate for any strategy, and none is implied here.
Frequently asked questions
Can I get a first freelance client with no portfolio at all?
Yes, through routes that do not need one. A referral or subcontract borrows someone else’s record, and a small paid pilot creates your own.
Should I work for free to get the first client?
The experiment that worked used paid jobs. Free, reusable proof such as a teardown is different from unpaid custom work, where two hours per prospect is a sensible editorial cap.
Do certificates or skill tests help?
They raised earnings by 8.9 percent over two weeks in the best study but produced one new project per approximately 63 credentials. Take them after the first job.
Is a platform or a direct client better for the first job?
A platform gives a public record but heavy competition: 28.8 applications per posting in one 2012 dataset. Direct routes rely on borrowed credibility. Many beginners need both.
How many applications should I expect to send?
No current benchmark exists. In older data the median non-affiliated worker who found a job needed ten applications and thirteen days, and only ten percent found one.
Has AI made it harder for beginners?
For written proposals, yes: tailored text lost about half its link to callbacks on one platform, and employers shifted towards work history. That raises the value of verifiable proof and of anyone willing to vouch for you.
Sources
- Pallais, A. (2014). Inefficient Hiring in Entry-Level Labor Markets. American Economic Review, 104(11), 3565-3599 (NBER Working Paper 18917 consulted).
- Stanton, C. and Thomas, C. (2016). Landing the First Job: The Value of Intermediaries in Online Hiring. The Review of Economic Studies, 83(2), 810-854 (CEP Discussion Paper 1316, November 2014, consulted).
- Agrawal, A., Lacetera, N. and Lyons, E. (2016). Does standardized information in online markets disproportionately benefit job applicants from less developed countries? Journal of International Economics, 103, 1-12 (NBER Working Paper 18720 consulted).
- Kässi, O. and Lehdonvirta, V. (2024). Do Microcredentials Help New Workers Enter the Market? Evidence from an Online Labor Platform. Journal of Human Resources, 59(4), 1284-1318.
- Kässi, O., Lehdonvirta, V. and Stephany, F. (2021). How many online workers are there in the world? A data-driven assessment. Open Research Europe, 1, 53 (revised version with corrected head counts).
- Pallais, A. and Sands, E. G. (2016). Why the Referential Treatment? Evidence from Field Experiments on Referrals. Journal of Political Economy, 124(6), 1793-1828 (NBER Working Paper 21357 consulted).
- Brown, M., Setren, E. and Topa, G. (2016). Do Informal Referrals Lead to Better Matches? Evidence from a Firm’s Employee Referral System. Journal of Labor Economics, 34(1), 161-209 (Federal Reserve Bank of New York Staff Report 568 consulted).
- Rajkumar, K., Saint-Jacques, G., Bojinov, I., Brynjolfsson, E. and Aral, S. (2022). A causal test of the strength of weak ties. Science, 377(6612), 1304-1310 (abstract only).
- Sackett, P. R., Zhang, C., Berry, C. M. and Lievens, F. (2022). Revisiting meta-analytic estimates of validity in personnel selection: Addressing systematic overcorrection for restriction of range. Journal of Applied Psychology, 107(11), 2040-2068.
- Barach, M. A., Golden, J. M. and Horton, J. J. (2020). Steering in Online Markets: The Role of Platform Incentives and Credibility. Management Science, 66(9), 4047-4070 (NBER Working Paper 25917 consulted).
- Filippas, A., Horton, J. J. and Golden, J. M. (2022). Reputation Inflation. Marketing Science, 41(4), 733-745 (NBER Working Paper 25857 consulted).
- Wiles, E., Munyikwa, Z. T. and Horton, J. J. (2023). Algorithmic Writing Assistance on Jobseekers’ Resumes Increases Hires. NBER Working Paper 30886.
- Galdin, A. and Silbert, J. (2025). Making Talk Cheap: Generative AI and Labor Market Signaling. arXiv 2511.08785. Working paper, not peer reviewed.
- Cui, J., Dias, G. and Ye, J. (2025). Signaling in the Age of AI: Evidence from Cover Letters. arXiv 2509.25054. Working paper, not peer reviewed.
- International Labour Organization (2021). World Employment and Social Outlook 2021: The role of digital labour platforms in transforming the world of work.
- Upwork Inc. Annual Report on Form 10-K for the fiscal year ended December 31, 2025, US Securities and Exchange Commission. Company-reported.
- Upwork Inc. Quarterly Report on Form 10-Q for the quarter ended June 30, 2026. Company-reported.
- Fiverr International Ltd. Annual Report on Form 20-F for the fiscal year ended December 31, 2025. Company-reported.
Table 1 reports figures as printed in the named sources. Table 2 is a CEOtudent analysis of those sources. Tables 3, 4 and 5, the grading rule, the spec-work rules, the 30-day plan and the scripts are CEOtudent editorial frameworks; their time costs and targets are editorial estimates, not data. Deliberately not used: reply and conversion rates for cold outreach, which circulate without a primary source; referral shares from commercial, self-reported vendor surveys; the 1998 work sample validity of .54, since revised to .33; the uncorrected head counts in the first preprint of the online workforce estimate; and platform fee percentages absent from the filings. CEOtudent ran no survey for this article.
This content was compiled with the support of AI following in-depth research, then written and prepared for publication by the CEOtudent editorial team.
This post is also available in:
