Diagnose Your Funnel: Which Stage Is Failing?

Forty applications, three assessments, one interview, no offers. It is easy to conclude "I'm not good enough" or "the market is impossible". Both conclusions are too broad to act on, and both are usually wrong.

Your tracker holds a more useful answer. If almost nobody responds to your applications, the problem is in what happens before anyone reads your work. If you get assessments but rarely interviews, it is the assessments. If you reach final rounds and stop, it is something about how you come across in those rounds. Each of these has different fixes, and applying harder fixes none of them.

• How to read your tracker as a funnel and calculate stage rates.
• How to break rates down by channel, resume version and company group.
• The typical causes and fixes for each failing stage.
• How to use AI to diagnose from counts without sharing personal data.
• How to avoid overreading small samples.

Even well-prepared candidates are rejected far more often than they are hired, because each opening has many applicants and only one or a few succeed. A rejection can reflect a better candidate on the day, a role filled internally, a hiring freeze, or an interviewer's preference. One rejection tells you almost nothing.

Many rejections, sorted by stage, tell you a lot.

Using the "furthest stage reached" column from your tracker (article 5), count how many applications reached each stage:

The rate that matters most is usually the lowest early rate. Here the screen rate is low, and only three screens means you have had too few chances to judge the later stages. The first job is to get more applications to a screen.

A single overall rate hides useful differences. Compare groups:

• By channel: careers pages, job boards, referrals, platforms, cold outreach.
• By resume version.
• By company group (article 3): are you getting responses only from one kind of company?
• By role type: backend versus full-stack versus data.

A short program can do this from a tracker exported as CSV. The data here is a made-up sample.

# File: funnel_report.py# Turn a tracker export into funnel rates, overall and by channel and resume version.# TRACKER is made-up sample data in the shape of the tracker from article 5:# one row per application, with the furthest stage reached.import csvimport ioTRACKER = """company,channel,resume,stageA,careers page,v1,appliedB,careers page,v1,screenC,job board,v1,appliedD,job board,v1,appliedE,referral,v2,interviewF,careers page,v2,appliedG,job board,v1,appliedH,referral,v2,screenI,careers page,v2,screenJ,job board,v2,appliedK,careers page,v2,interviewL,job board,v1,appliedM,referral,v2,offerN,careers page,v1,appliedO,job board,v2,appliedP,careers page,v2,applied"""ORDER = ["applied", "screen", "interview", "offer"]def reached(rows, stage):    return sum(1 for r in rows if ORDER.index(r["stage"]) >= ORDER.index(stage))def rate(part, whole):    return f"{part}/{whole} ({part / whole:.0%})" if whole else "n/a"rows = list(csv.DictReader(io.StringIO(TRACKER)))print(f"Applications: {len(rows)}")print("Screen rate:   ", rate(reached(rows, "screen"), len(rows)))print("Interview rate:", rate(reached(rows, "interview"), reached(rows, "screen")))print("Offer rate:    ", rate(reached(rows, "offer"), reached(rows, "interview")))for field in ("channel", "resume"):    print(f"\nScreen rate by {field}:")    for value in sorted({r[field] for r in rows}):        group = [r for r in rows if r[field] == value]        print(f"  {value:<13}{rate(reached(group, 'screen'), len(group))}")

Applications: 16Screen rate:    6/16 (38%)Interview rate: 3/6 (50%)Offer rate:     1/3 (33%)Screen rate by channel:  careers page 3/7 (43%)  job board    0/6 (0%)  referral     3/3 (100%)Screen rate by resume:  v1           1/7 (14%)  v2           5/9 (56%)

What this sample suggests, cautiously: job-board applications produced nothing, referrals produced a screen every time, and resume v2 did much better than v1. With only 16 rows these are hints, not proof, but they point to clear experiments: stop sending v1, put job-board time into referrals, and keep tracking.

A spreadsheet pivot table does the same job if you prefer not to use code.

• Few tries, wide uncertainty. One screen from ten applications could be bad luck. One from fifty is a pattern.
• A rough guide: treat a stage rate as a signal once at least 20 to 30 tries have reached that stage's start. Interviews are rarer, so interview and offer rates need longer to become meaningful; use your notes (article 43) alongside the numbers.
• Large differences matter more than small ones. 0 of 6 versus 3 of 3 deserves attention even in small samples. 20 per cent versus 25 per cent does not.
• Change one thing at a time where you can, such as the resume version, so you can tell what helped.
• Seasons and markets move. A sudden drop across all channels may be the market, not you. Keep going, and compare with what peers are seeing.

Your tracker contains other people's names and companies' details. For an AI diagnosis, send counts and categories only:

Here is my job search funnel data for the last [N] weeks (counts only).Applications: [n]. Furthest stage reached: screen [n], assessment passed [n],interview [n], final round [n], offer [n].By channel (applications, screens): careers page [ , ], job board [ , ],referral [ , ], platform [ , ], cold outreach [ , ].By resume version (applications, screens): v1 [ , ], v2 [ , ].By company group (applications, screens): [ ... ].Notes on rejections at later stages: [short, e.g. "2 interview rejections:both after a design question"].1. Which stage is my main bottleneck? Explain using the numbers.2. Which differences between groups are large enough to act on, and which are   probably noise given the sample sizes?3. What are the two most likely causes, and what single change should I test   first for the next two weeks?4. What should I track to know whether the change worked?Do not assume anything not in this data.

Check the assistant's reasoning against your own numbers. If it calls a 1-out-of-3 difference "significant", discount that.

Harish has applied for ten weeks. His tracker shows:

• 92 applications; 9 screens (10%); 6 assessments passed; 4 interviews; 0 offers.
• By channel: job board 55 applications, 2 screens; careers pages 30, 4 screens; referrals 7, 3 screens.
• Interview notes: two rejections after a machine coding round, one after a behavioral round, one pending.

Diagnosis:

1. Screens are the first bottleneck, and job-board applications drive it down (about 4 per cent against 13 per cent from careers pages and 43 per cent from referrals). He moves most job-board time to referrals and careers pages.
2. Assessments are fine: 6 of 9 passed.
3. Interviews: four is a small number, but two of them failed at the same round type, machine coding, which matches his notes. He adds weekly 90-minute machine coding practice (article 31).

He changes two things, aimed at two different stages, and tracks both for three weeks.

• Reacting to single rejections instead of patterns.
• Looking only at the overall rate, missing channel and version differences.
• Fixing a later stage first when an earlier one is the bottleneck.
• Treating tiny samples as proof.
• Changing everything at once, so nothing can be learned.
• Pasting your tracker with names into an AI tool.

Funnel (8 weeks): 48 applied; 5 screens (10%); 3 assessments passed; 1 interview; 0 offers.By channel: job board 30 -> 1; careers page 14 -> 2; referral 4 -> 2.By version: v1 28 -> 1; v2 20 -> 4.Bottleneck: screens.Change to test for 2 weeks: v2 only, and half of job-board time moved to referrals.Measure: screens per application by channel and version; target 15%+ overall.

• Next: Feedback When Companies Won't Give It: Reconstruction and Self-Diagnosis diagnoses individual interviews.
• For the tracker itself, revisit Your Pipeline Tracker and 10-Week Execution Plan.
• For using AI with data you can check, see the AI course's article Spreadsheets and Data with AI: Analysis You Can Check.