← All case studies

Case study · Machine learning advisory

A machine learning analysis for Monster.com

A vice president asked for a written view on two open questions: what could be learned about the candidate journey, and whether machine learning would materially improve resume-to-job matching. Fifty-six pages, delivered while the company was mid-rebuild.

The candidate journey, and where the data existsSearchjob searchfilters, queriesViewposting openeddwell timeApplyresume submittedthe only clean eventScreenrecruiter actioninvisible to the candidateOutcomehired or notrarely fed backOnly one stage produces clean data. The recommendation had to work with that.
Most of the journey is unobservable. Any model has to be honest about which parts it is inferring.

1. The brief

A vice president at Monster.com asked for a written analysis. The company had been through a long decline and was being rebuilt, and two questions were open: what could actually be learned about the candidate journey, and whether machine learning would materially improve the matching of resumes to job listings.

The work was unsolicited in origin — a conversation that turned into a request — and it was delivered as a fifty-six page analysis.

56pages of analysis delivered
2problems examined
3competitor implementations documented
2recommendations, with next steps

2. Start with the questions, not the technology

Figure 1

What was actually being asked

Two questions the executive asked us to answer The customer journey What are candidates actually doing on the way to a placement? And can it be measured per person, not just in aggregate? Resume-to-job matching Can machine learning make the recommendations meaningfully better? And what would it actually take to build? Neither question had an obvious answer, and the company was mid-rebuild after a long decline. The brief was to find out what was true, not to sell a solution.

The two problems were treated separately because they have different data requirements, different success measures and different build costs. Merging them would have produced one confident answer to neither.

The section that made the rest credible

The analysis opened with a page of questions back to the client — what are you tracking now, how well is it working, which parts of the journey matter to you, what does “better” mean here. Not rhetorical questions. Real ones, because the answers changed which methods were appropriate. A proposal that arrives certain about a business it has not yet examined is a proposal that has stopped listening.

3. Look at what already exists before proposing to build

Figure 2

Three implementations, examined

Rather than assert what was possible, we documented what three others had already built Case 1 A professional network graph-based matching Case 2 A job aggregator ranking at scale Case 3 A cloud jobs API matching as a bought service The third case was the uncomfortable one: a large part of what was being considered as a build could be bought. A proposal that had not looked would have missed it.

Each was documented for what it does, how it appears to work and what it implies about cost and effort - rather than named as a competitor and left there.

The finding that argued against our own scope

One of the three was a cloud service selling job-matching as an API. That meant a meaningful part of what was being considered as an internal build was available to buy. Saying so reduces the size of the project being proposed, which is exactly why most proposals do not say it.

4. What was recommended

RecommendationWhat it involved
Understand the journey properly firstEvaluate the available journey-analysis platforms against the specific need, then work out where custom analysis has to fill the gap. Not a tool purchase — a fit test
Test matching where the data already existsRecommendations from referral keyword traffic, from partial search behavior on the site, and delivered by periodic email — three places where signal was already being generated and discarded

A further section listed opportunities beyond the brief: predictive analytics on the journey, keyword and search-structure analysis from server logs, and other applications of the same techniques.

The technical judgment underneath it

Visualizing individual paths through a site is a Sankey problem, and the analysis included working code for it rather than a description. Showing the actual method is a different kind of claim from naming it, and it lets a technical reader judge whether the person writing has done this before.

6. What this case is meant to show

The usual proposalWhat was done here
Arrives certainOpens with questions whose answers change the recommendation
Proposes to buildDocuments what can be bought first, even when that shrinks the work
Names the technologyShows the method, in code, so a technical reader can judge it
One big recommendationTwo, separated because they need different data and prove different things
Why this one is named

Every other case study on this site withholds the client. This one does not, because the work was a written analysis prepared at an executive’s request rather than an assessment of a company that did not ask to be assessed. The distinction matters and it is the whole reason the others stay anonymous.

Marketing Analytics Consultants

Considering whether AI is worth building for something specific?

We look at what already exists, tell you what can be bought rather than built, and give you the arithmetic — including when the answer is that it is not worth doing.

Prefer email? info@marketinganalyticsconsultants.com