Peer Reviews and Search Behavior on a Job Platform in Mexico

The intervention is the display of a verified positive peer testimonial on a job posting on an online job-search platform in Mexico. On eligible vacancies, treated candidates see a card reading "Un chambeador opina:" ("A worker says:") followed by an italicized quote drawn from a verified, positive employee review — no star ratings, counts, or averages are shown. Two testimonial variants are tested: a "Group 1" variant emphasizing respectful treatment by the employer, and a "Group 2" variant emphasizing the employer's compliance with pay and benefit terms; which variant a candidate sees depends on their assigned group (see Experimental Design for the full assignment structure). Control candidates see the standard vacancy listing with no testimonial. The intervention targets active job-seeking candidates on the platform and is intended to test whether credible third-party signals about employer quality redirect search behavior toward signaled employers, and separately, whether such signals change how intensively candidates search overall (e.g., returning to the platform more, browsing more vacancies).

2026-07-09

2026-07-18

1. Applied to at least one vacancy during the trial observation window (binary, candidate-level)
2. Applied to a reviewed/eligible employer (binary, candidate-level)
3. Share of a candidate's applications directed to reviewed employers (continuous, candidate-level)
4. Search intensity and platform engagement following exposure: number of return visits/sessions and browsing activity (e.g., vacancies viewed, "see more vacancies" interactions) during the trial window (candidate-level, counts)

Outcome 1 is an indicator equal to 1 if the candidate submits one or more applications during the observation window, regardless of destination. Outcome 2 is an indicator equal to 1 if any of those applications is directed to a treatment-eligible employer (an employer with an active, displayed testimonial at the time). Unlike Secondary Outcome 1, this outcome is unconditional — measured across the full randomized sample regardless of whether a candidate viewed anything — and at the employer rather than the specific-vacancy level. Because it does not condition on a post-treatment behavior such as viewing, it remains a clean comparison across arms even if search intensity is itself part of how it comes about; the endogeneity concern raised for Secondary Outcome 1 is specific to conditioning on a mediating variable, not to unconditional outcomes that bundle all channels together. Outcome 3 is the ratio of applications to reviewed employers over total applications, defined among candidates with at least one application. Outcome 4 captures whether the intervention changes not just whether candidates apply, but how actively they search: return visits/sessions to the platform after initial exposure, and within-session browsing depth (vacancies viewed, "see more vacancies" clicks). The motivating hypothesis is that a credible positive signal raises a candidate's perceived returns to searching, leading them to search more rather than stop once one good option appears.

1. Applied to the specific reviewed vacancy, conditional on having viewed it (candidate-level, vacancy-conditional), examined two ways: (a) conditional on viewing at any point during the trial window, and (b) conditional on viewing within the candidate's first carousel of vacancies, in their first session after trial start — a stricter definition intended to be closer to treatment-independent
2. Interview attendance, where observed
3. Hire, where observed
4. Respect- vs. compliance-framed testimonial contrast on outcomes 1–3 above, restricted to the subset of treatment-eligible employers where the two dimensions are cleanly distinguished
5. Heterogeneity in Primary Outcomes 1–4 by candidate gender and by candidate occupation
6. Difference-in-differences comparing application rates to treatment-eligible employers versus non-eligible employers, across arms, to isolate redirection of applications toward reviewed employers from any general increase in application or search activity

Outcome 1 is measured only among candidates who viewed an eligible vacancy, and is expected to be more power-constrained than the primary outcomes because exposure is concentrated in a small number of employers. Viewing is not necessarily independent of treatment: if disclosure increases how actively candidates search (Primary Outcome 4), the set of candidates who view an eligible vacancy may itself differ by arm. Outcome 1 is therefore defined two ways — a broad version (viewed at any point during the trial window), better powered but more exposed to this concern and best read as descriptive rather than causal, and a narrower version (viewed within the candidate's first carousel in their first session after trial start), which is close to treatment-independent, since the first screen a candidate sees is set by the platform's feed logic before any reaction to a testimonial is possible, at the cost of a smaller sample. Outcomes 2–3 are observed for a partial, self-reported subset of candidates (historically ~27–30% coverage) and are treated as exploratory given this attrition. Outcome 4 is exploratory: the dimension contrast is available for only a minority of treatment-eligible employers. Outcome 5 examines whether effects on the primary outcomes differ by candidate gender or occupation; gender is expected close to an even split and reasonably well powered, while occupation is expected to be unevenly distributed, so heterogeneity for less common occupations is treated as exploratory and less reliable, with rare categories grouped rather than tested individually. Outcome 6 constructs two unconditional, candidate-level rates within each arm — the probability of applying to a treatment-eligible employer (as in Primary Outcome 2) and the probability of applying to a non-eligible employer — and compares their difference across arms to isolate redirection toward reviewed employers from any general increase in application activity captured separately in Primary Outcome 4; this is estimated via an arm-by-employer-eligibility interaction on a stacked candidate-by-employer-type panel, clustered by candidate, with the same comparison estimated in the control arm alone serving as a placebo check on whether eligible employers are inherently more attractive independent of any testimonial.

This is a randomized controlled trial testing the effect of disclosing verified positive peer testimonials about employers on job seekers' application behavior on an online job-search platform in Mexico. All candidates active on the platform during the trial period are randomly assigned, at the individual level, to one of three arms using a deterministic hash of their candidate identifier (1:1:1 ratio): a control arm that sees the standard platform experience with no testimonial; a "Group 1" arm that, on eligible vacancies, sees a testimonial emphasizing respectful treatment by the employer; and a "Group 2" arm that sees a testimonial emphasizing the employer's compliance with pay and benefit terms. Assignment is fixed for each candidate for the duration of the trial. Testimonials are displayed only on vacancies from employers for which a positive, verified testimonial is available at the time a candidate views the posting; this set of eligible employers is expected to number in the low tens and may fluctuate over the trial period as employers enter or exit the platform's active client base. All other vacancies are unaffected for all arms. For a minority of eligible employers, the testimonial content differs meaningfully between Group 1 and Group 2 along the respect/compliance dimension; for the remaining eligible employers, both treatment arms see the same general positive testimonial, and these observations contribute only to the pooled treatment effect. The trial runs for 10 days. Approximately 60,000 candidates are expected to be randomized, of whom a minority are expected to view at least one eligible vacancy during the trial and thus be exposed to a testimonial if assigned to a treatment arm. Outcomes are measured using platform administrative data on views and applications for the full population of randomized candidates. This trial may be followed by a subsequent wave, incorporating additional employer reviews and design refinements informed by this first implementation; any such wave would be registered separately, with its own dates, sample, and design. If a subsequent wave is conducted, data may be pooled across waves in the eventual analysis, with wave fixed effects and attention to candidates who appear in more than one wave.

Not available

Each candidate is assigned to a group using a fixed calculation applied to their unique candidate ID, rather than a physical or one-time random draw. The calculation (MD5, a standard method for turning a piece of text into a long number) converts each candidate's ID into a large number; the remainder after dividing that number by 100 places the candidate into one of three groups: a remainder of 0–32 assigns them to control, 33–65 to Group 1 (respect-framed testimonial), and 66–99 to Group 2 (compliance-framed testimonial). Because the same ID always produces the same remainder, the assignment is fixed and can be independently recomputed and verified by anyone with the candidate ID list, while still being unpredictable in advance and evenly spread out — producing the intended, approximately equal three-way (1:1:1) split.

Individual (candidate_id). Assignment is fixed for each candidate for the duration of the trial: a candidate sees the same arm across every eligible vacancy they view during the trial window.

No

Approximately 60,000 candidates.

Approximately 60,000 candidates.

Control: ~20,000. Group 1 (respect): ~20,000. Group 2 (compliance): ~20,000.

Using platform data on candidate activity, the baseline application rate (share of active candidates who submit at least one application) is approximately 30%. At 80% power and a two-sided 5% significance level, with approximately 20,000 candidates per arm, the minimum detectable effect for Primary Outcome 1 (applied to at least one vacancy) is approximately 1.1–1.3 percentage points (roughly 4% relative to baseline), depending on whether the comparison pools both treatment arms against control or compares a single arm against control. This calculation covers Primary Outcome 1 only; the vacancy- and employer-conditional secondary outcomes remain more power-constrained, as noted in the Secondary Outcomes explanation.