What the Delhi Experiment Actually Changed
To understand the results of the Delhi experiment, it is not enough to say that some people were given “information.” This chapter reconstructs the actual intervention: what every treated participant received, what personalization changed, and where the path from exposure to employment became uncertain. That reconstruction gives readers a precise way to interpret the comparison between generic and tailored messages—and to avoid mistaking access to a digital tool for a successful job match.
The intervention was designed around a practical problem. Job seekers may know that digital platforms exist but still fail to register, create a usable profile, return to the platform, identify relevant vacancies, or follow up with employers. The researchers therefore combined personal assistance with repeated reminders. The treatment was not a single piece of advice; it was a sequence intended to move people through several stages of search.
Three layers of treatment, one causal question
The study divided 130 Delhi polling-station clusters into a control group and two treatment groups. The control group received neither the in-person information session nor the phone messages. The two treatment groups received the same initial offer: an explanation of the benefits of digital job search, information about a collaborating mobile-app and web-based platform, assistance with registration, and guidance on building a profile, searching for preferred jobs, and finding skills resources.
The common in-person component mattered because it addressed the intention-to-action gap. A person may say that a platform sounds useful and still not complete registration. With assistance, the participant could move from hearing about a search channel to setting up an account and learning how to use it. The session also made the platform concrete rather than leaving “digital search” as a vague recommendation.
The partner platform connected mainly blue- and grey-collar workers with employers. It allowed registered users to search vacancies, filter by location and job type, submit applications, or contact human-resources representatives directly. It also offered skills tests and learning resources. These features are relevant here not as a product inventory but because they define what the intervention made possible: a job seeker could encounter a vacancy, present a profile, and initiate contact without relying entirely on a personal intermediary.
The platform was still an interface between workers and employers, not an employer itself. It could display vacancies and record applications, but the final decision remained outside the researchers' control. A platform can reduce the effort of finding and applying for a job while leaving the number of suitable vacancies, the employer's selection decision, and the worker's acceptance decision unchanged.
After the in-person session, both treatment groups were sent six SMS or WhatsApp messages at intervals of roughly 10 to 15 days. The messages were designed to support both extensive take-up—registering in the first place—and intensive use—searching regularly and using the available tools more efficiently. Their topics included how to search, the number of jobs currently available in Delhi, highly demanded jobs on the platform, English-language requirements, government skilling programmes, and a government job portal.
That common sequence answers one causal question: what happens when people are offered information and hands-on access to digital job search, compared with people who receive neither? But the study also asks a more specific question about relevance. Does the same amount of messaging work differently when its content is adapted to the recipient?
Personalization changed relevance, not the underlying intervention
The Non-Personalized group received the same general information regardless of personal characteristics. Its messages described broad job-search strategies, vacancy counts, common job types, language requirements, and public and private opportunities. The Personalized group received messages built from baseline information about gender, whether the respondent was currently looking or expected to look later, and whether they aspired to jobs listed on the partner platform.
The tailoring could change the kind of opportunity placed in front of a person. A prospective male job seeker interested in private-sector work might receive information about part-time or work-from-home roles and the returns to investing in skills. A prospective woman interested in public-sector work might receive information about part-time government openings or vacancies directed toward female candidates. These examples matter because relevance is not the same thing as volume. Two people can receive six messages, yet one may receive a set that maps more closely onto their schedule, sectoral preference, or desired type of work.
Personalization also applied to all respondents assigned to that group, whether or not they had registered on the partner platform. That choice is important. If only registered users had received tailored messages, the comparison would tell us mainly how personalization affected people who had already crossed the first access barrier. By messaging everyone, the researchers attempted to reach both people who had registered and people who had expressed no initial interest or had not completed onboarding.
The design therefore contains a layered comparison. The control-versus-treatment comparison captures the effect of the full outreach package: in-person explanation and assistance plus a stream of messages. The Personalized-versus-Non-Personalized comparison captures the incremental effect of tailoring within a package that already includes the same initial assistance and platform opportunity. It does not tell us what personalization would do on its own, without the common session, or what any single message contributed independently of the others.
The sequence makes the intervention's ownership boundaries visible. The field team controls the session and message content. The participant decides whether to register, search, apply, or call. The platform records some actions and passes contact toward employers. Employers decide whether to respond. The worker and employer determine what happens after that response. A measured increase at one boundary should not be silently carried forward as if it occurred at every later boundary.
The take-up funnel loses people at each transition
The first transition was interest. Among participants offered treatment, about 65% expressed interest in registering for the platform. Almost all respondents who expressed interest did register. For the full treatment sample, unconditional registration was about 65.0% in the general-message arm and 68.8% in the tailored-message arm. Women showed higher interest and registration overall than men. This is evidence that the offer of a new information route had substantial appeal, especially for people who had faced greater limitations in traditional search channels.
But interest and registration measure willingness to try a tool, not whether the tool becomes part of a person's search routine. The researchers' administrative platform data show that only 39% of registrants were classified as active users, meaning they submitted at least one application or called human-resources staff. A registered participant submitted an average of 4.7 applications, but only 13.7% of applications received a callback from HR. These figures have different denominators, so they should not be multiplied into a single conversion rate. Their value is diagnostic: movement is lost at each step.
The same logic applies to the messages. Delivery was nearly universal, but receipt and observed WhatsApp viewing declined across the six-message sequence. “Seen” rates are a lower bound because a participant can disable read receipts. Personalized messages were more likely to be received and seen overall, particularly among men, but that does not mean that every recipient read, understood, or acted on them.
This funnel changes how high take-up should be interpreted. The intervention succeeded in making the platform and messages reach many people. It did not ensure sustained use, a relevant application, employer interest, or a completed match. A registration rate can be high while the active-user rate is much lower. A high number of applications can coexist with a low callback rate. And a callback still does not reveal the contract offered or whether the worker accepted it.
That attrition is not necessarily a flaw in the intervention. Some people may register simply to inspect an option, discover that the available jobs do not fit, or decide that another channel is preferable. Others may face costs that the platform cannot remove, such as family responsibilities, skills or language requirements, or employer-side screening. The data cannot always distinguish a well-informed decision not to continue from a failure of the platform to remain useful.
Why messages were sent beyond the platform's boundary
Sending messages to unregistered participants served two purposes. First, it created another opportunity to complete registration after the in-person visit. Second, it allowed the intervention to provide labor-market information even when the participant did not use the partner platform. The messages included information about government skilling and a government job portal, so the intended search environment was broader than one commercial or partner platform.
This matters for interpreting later survey outcomes. If a woman reports more digital search, that response may include the partner platform, a government portal, or another online route. The treatment can change awareness and channel choice without producing a proportional increase in activity on the partner platform itself. Conversely, a fall in partner-platform applications after a message about another portal may represent redirection rather than disengagement.
The messages also show why “information” has both a content and a delivery problem. Content must be relevant enough to change a belief or identify a plausible opportunity. Delivery must occur when the recipient can notice and use it. The study observes that personalized messages were more likely to be received and seen overall, but the result does not establish that personalization improved comprehension or that a viewed message caused a particular application. Those are further links in the chain.
The intervention was consequently trying to solve several different frictions at once: awareness, registration, navigation, repeated attention, and relevance. That breadth made the package more practical for a real job seeker, but it also made the experimental effects harder to attribute to one component. The bundle is a strength for studying a workable outreach model and a limitation for identifying the isolated contribution of each design choice.
The causal boundaries of the package
The first boundary is the distinction between assignment and receipt. Because clusters were randomly assigned, comparing everyone assigned to treatment with everyone assigned to control gives an intent-to-treat effect. It preserves the practical question of what happens when an outreach programme is offered in a real setting, including people who do not register or do not see every message. An analysis limited to message recipients would answer a different question and would require stronger assumptions about who receives information.
The second boundary is the difference between the two treatments. The personalized arm and the general arm received the same number of messages at roughly the same frequency and shared the in-person session. The additional contrast is relevance of content, not access to messaging itself. If the groups later differ, the evidence is consistent with a personalization effect within this broader package, but it cannot prove that tailoring alone would generate the same result in another setting.
The third boundary is the difference between platform activity and labor-market success. Applications and HR callbacks are intermediate outcomes. The platform does not record the final job offer, the nature of the contract, or whether an offer was accepted. It therefore cannot establish from its own records whether a participant moved into stable, satisfactory, or better-paid work. The survey data extend beyond the platform, but they still measure outcomes within the study's selected population and follow-up period.
Finally, the intervention should not be judged by a single activity metric. A generic message that prompts broad exploration may produce more applications over time. A tailored message may direct a person quickly toward fewer but more suitable vacancies. A participant who stops applying after learning that an ideal wage is unlikely may be updating a plan; another may be discouraged. The meaning of the observed action depends on the belief and access constraint that preceded it, and on what happens at the next stage.
The Delhi experiment therefore changed a set of opportunities to know, register, search, and apply—not a guaranteed path to employment. Once those stages are kept separate, the evidence can be read as a causal chain with several possible branches. The next task is to connect the starting beliefs and access constraints to that chain and ask what kinds of response each branch predicts.