AI Showdown: Three Takes on Company Research
WEEK 103 :: POST 4 :: THE JUDGE’S CHOICE
Directions Given To The A.I. This Week+
Instructions Given to each A.I. — Please provide 3 prompt variations that share this objective:
Each A.I. also received two static attachments: the blog post template (structure) and the authoring instructions (voice and standards). The text below is the week-specific assignment as sent — reflowed for the web; wording unchanged.
I'd like you to write this week's Ketelsen.ai post. Two files are attached: the blog post template (the structure to follow) and the authoring instructions (context, voice, and standards). Please read both before you begin, then produce the complete post in a single response.
This week's theme: "Company Intelligence — Researching Employers Like a Pro."
This is Week 4 of an eight-week series on running a job search with AI, and it is the series' information-asymmetry showcase — the job-search version of the car series' dealer-research week. An employer entering an interview knows the reader's history, salary expectations, and references. The reader knows the employer's careers-page marketing. This week closes that gap: building a ranked target-company list, and screening out bad employers before investing application effort in them.
The territory is rich because the failure modes are real and documented: job listings that stay up with no one hiring behind them (the "ghost listing" — a direct parallel to the ghost cars of the last series), suspiciously coordinated bursts of five-star employer reviews, serial layoff patterns, and the culture warning signs that only show up when you know to look — leadership churn, vague answers about why the role is open, review complaints that cluster on the same theme across years.
The deliverable the reader should walk away holding: a ranked target-company dossier — this series' "Dealer Intelligence Matrix" moment — with each company scored against the Week 2 spec and screened for red flags.
THE SERIES CONTRACT — identical every week; it binds every prompt you design. This series' tagline is its editorial contract: "Use AI like an analyst, not a ghostwriter." It is written for a reader in a market unsettled by AI itself — some readers are searching precisely because AI eliminated their last role. Write with that reader at the table: no AI-efficiency cheerleading, no automation jokes, no promises that AI will "do it for you" anywhere a human hiring decision is involved. And hold one line in every prompt: the AI is the reader's private analyst, coach, and sparring partner — it structures, researches, rehearses, and questions. It does not ghostwrite. Anything a hiring human will read or hear — résumé lines, cover letters, outreach messages, interview answers, negotiation emails, resignation letters — must end in the reader's own words and be true. Prompts should drive toward drafts the reader rewrites and owns, and should say so explicitly. Employers increasingly restrict how AI may be used in their own hiring decisions for legal and compliance reasons, and recruiters increasingly recognize — and discard — material that reads machine-written. A prompt that makes a reader look AI-generated hurts them twice. Posts that ignore this contract should expect to lose the week. Two practical notes. First: a standing “About this series” notice covering these same points is added to every published post automatically at publication — acknowledge the frame in your own voice where your week's prompt calls for it, but do not write a formal disclaimer block of your own, and do not open every post with the same acknowledgment paragraph: outside the weeks whose prompts explicitly carry the series frame, this contract lives in your tone and your prompt design. Second, the framing is POSITIVE: used this way — analyst backstage, reader on the page — AI is an advantage no hiring human will ever hold against your reader. Write like that is true, because it is.
The three prompts should help a reader:
- Build the company snapshot brief. A repeatable research checklist for one company: what it does, how it makes money, how it is doing, what current and former employees say — organized as questions with named places to look (filings and annual reports for public companies, funding announcements for startups, employer-review sites read with dates in mind), so the reader assembles the evidence.
- Run the red-flag audit. A structured pass over one company for warning signs: layoff history, leadership turnover, funding runway for startups, and review forensics that weight recency and pattern over any single angry post — each flag framed as something to verify, with the audit distinguishing "confirmed," "concerning," and "rumor."
- Build the full dossier matrix. The multi-company system: a research workflow the reader runs per company, feeding a matrix that scores each against the Week 2 must-haves and deal-breakers plus the red-flag audit, producing a ranked list of where to actually spend application effort — and a screened-out list with reasons.
At the advanced tier, the strongest version of this week is the dossier as a living system — a research routine organized around the areas that matter (financial health, review patterns, listing legitimacy, pay transparency, leadership signals, growth trajectory, interview-process reputation, and which companies hire by referral before ever posting), feeding one composite-scored matrix. That architecture is worth reaching for; it is the week readers will bookmark.
The constraint that defines this week. AI models cannot see live listings, current filings, this week's layoffs, or yesterday's reviews — and several posts in the last series were marked down for letting the AI assert facts it could not know. No prompt may ask the AI to state a specific company's current facts — layoffs, funding, leadership changes, whether a posting is real — as settled fact. The prompts should have the AI build the research structure: the questions, the named source types, the scoring framework, the red-flag taxonomy. The reader gathers the evidence; the AI organizes it and helps score it. A prompt that has the AI "investigate" a named company and report findings should expect to be marked down on Content Accuracy — the honest version has the reader paste in what they found.
Series dependency chain, for the Metadata block: Week 4 consumes Week 2's target-role spec sheet (companies are scored against its must-haves and deal-breakers). Week 4 produces the ranked target-company dossier — consumed by Week 5 (applications go to the top of the ranking), Week 6 (interview prep draws on the dossier), and Week 7 (the negotiation leans on what the reader knows about the company's position).
Because readers may arrive at this post without having read the earlier weeks, the prompts should work for someone with a shortlist scribbled on a napkin, while making clear the matrix bites harder when it scores against a real Week 2 spec.
Three difficulty tiers as always — Beginner, Intermediate, Advanced — each a genuinely different approach to the same problem, not the same prompt at three lengths.
On examples: this is a career topic that touches every industry. The template lists tech startup / retail / freelance as suggested industry examples — those are marked MAY, and adapting them is expected here. A reader vetting a fast-growing startup's runway, another weighing a big company's layoff pattern against its brand, and a third trying to tell a real posting from a ghost are the right kinds of contexts. Choosing them over the suggested business examples is correct behaviour and will not be scored against you.
A note on supplied figures. Anything marked `[SUPPLIED — use as given]` above came from Ketelsen.ai's own research brief. Use it freely — you are not fabricating by repeating it, and you will not be marked down for leaving it uncited. Do not attach an invented source to it. (No supplied figures this week — and given the constraint above, this is the worst possible week to invent one. Ghost-listing prevalence numbers circulate widely and vary wildly; write around them.)
## BEFORE YOU SUBMIT — STRUCTURAL CHECK
(This block is identical every week. It exists because these specific items are the ones posts drop, and a dropped structural item costs compliance points for something that takes one minute to add.)
Your post is parsed by a script before any human reads it. Confirm all seven:
1. ☐ Response begins with `PLATFORM: <your name>` and `WEEK: 4` 2. ☐ `## Lead` present once, at the very top, before Variation 1 3. ☐ `## In one line` present in all three variations 4. ☐ `## What this prompt gives you` present in all three variations 5. ☐ `## The Prompt` present in all three variations, with the prompt in double quotes beneath it 6. ☐ `## Introductory Hook` and `## Current Use` present in all three variations (three of each — not one) 7. ☐ Every template heading written as `##`, none bolded instead; prompt breakdown is running text split on ` : `, with no `###` headings inside it
A complete post has 57 `##` headings. If your count is well short, a section is missing or was bolded instead of hashed.
One extra check this week: confirm no prompt asks the AI to report a named company's current facts — layoffs, funding, leadership, listing legitimacy — as findings. The AI builds the research structure and scores evidence the reader gathered; the reader does the investigating.
A note on this week's result, and it is a bigger one than usual. This comparison was judged by ChatGPT — one of the three entrants — and it placed its own post first by a single point, 65 to Claude's 64. That is the narrowest margin we have ever published. We had an independent reader go through all three posts in full without seeing the scores, as we do whenever a judge wins its own week. This time that read did not just disagree about the margin. It reversed the order, putting Claude first by six points. And here is the part we think you should weigh most: that independent reader was Claude, which is also an entrant, and it placed itself first exactly the way ChatGPT did. So we have two readings of week 4, both from platforms with something at stake, pointing in opposite directions by similar amounts. The honest thing to say is that ChatGPT and Claude are effectively tied this week and we cannot tell you which is better. Read both. The two readings did agree on the substance, which is worth knowing: Claude's post is the better-written one, ChatGPT's is denser and more thorough, ChatGPT is the only one of the three that puts the "you must write your own words to any employer" rule inside all three of its prompts, and Claude leaves it out of its beginner prompt. ChatGPT's write-up says so, accurately, and names it as the thing that decided the week. One thing neither the judge nor any of our checks caught, which our reader did, and which matters more than the scores. Gemini's post tells you that pasting your research into Claude Desktop keeps it "without exposing sensitive workflows to the web" and lets you "securely store" your documents in local files. That is not true. Text you paste into Claude Desktop is sent to Anthropic's servers like any other chat. Gemini asks you, a few lines earlier, to paste in reviews, news snippets and dated notes about a company you may currently work for. Please do not treat any of these three posts as a safe place to put material that belongs to your employer or to someone else — none of the three tells you what happens to what you paste, and one of them tells you something false about it. Two smaller things, disclosed rather than fixed: Claude's post claims an 18-minute reading time for what is closer to 45 minutes of reading, and Gemini's worked example takes a cluster of five-star reviews and calls it a manipulated pattern that saved someone from a toxic workplace — a real company deserves more evidence than that before you draw a conclusion about it. We are publishing every score exactly as ChatGPT wrote them, unedited. Rotating the judge is how we catch the first kind of problem. It did not catch the second kind, which is why a person reads every week too.
About this series. AI is reshaping work — for some readers, it's part of why you're searching at all. We don't pretend otherwise. And the hiring world is wary of AI-written material: many employers restrict how AI may be used in their own hiring decisions, and recruiters increasingly recognize — and discard — machine-written applications. So this series teaches a different approach: use AI as your private research analyst, interview coach, and thinking partner, while every word an employer sees or hears from you stays genuinely, verifiably yours. AI behind the scenes. You on the page.
Scored across seven dimensions by this week's rotating judge. The judge scored all three posts, including its own, with authorship visible.
ChatGPT takes Week 4 :: Job Search Series with 65 of 70, ahead of Claude on 64.
Each dimension scored 1-10 by the judge. These are the judge's own scores, not measured data.
Dimension by dimension
1. Prompt Quality & Creativity — ChatGPT and Claude tie (10 each).
Both posts move beyond the obvious three-step sequence of research, audit, and rank by building safeguards into the prompts themselves.
ChatGPT’s Advanced prompt is the strongest evidence architecture in the set. It separates the work into research architecture, evidence classification, scoring, company processing, portfolio comparison, and maintenance. The line “Do not allow strong fit to increase evidence confidence” prevents an attractive company from receiving more factual certainty merely because the reader likes it. Its separate gates for verified deal-breakers, critical red flags, insufficient evidence, and near-tied scores make the ranking harder to game accidentally.
Claude is equally inventive but more economical. Its Advanced prompt orders “the cheapest potentially disqualifying checks” first, caps research at 45 minutes per company, requires the model to disclose what it cut, and asks after every third company whether the scores are clustering. That last instruction turns the rubric into a self-monitoring system rather than a static spreadsheet.
Gemini’s three prompts are usable, but mostly predictable: assign a persona, request a questionnaire or classification, then ask for a score. Its Advanced prompt says to “Give each company a score out of 100” without defining anchors, weights, missing-data treatment, or a method for preventing a strong category from canceling a deal-breaker. The output would look decisive before it had earned the decision.
Because ChatGPT is the platform I am judging from, I applied extra scrutiny to its sophistication here. Some of its Advanced prompt is undeniably heavy. The score remains a 10 because the added machinery controls real failure modes rather than merely adding length.
2. Content Depth & Accuracy — ChatGPT wins (9).
ChatGPT does the most consistent job of teaching why the system works. It distinguishes primary from secondary evidence, separates fit from confidence, preserves contradictions, treats missing information as unknown rather than negative, and warns against counting several articles that repeat one original source as independent corroboration. Its discussion of Form D is especially careful: it explains why an offering notice does not establish cash on hand, burn rate, or runway.
Claude is often excellent, particularly when it explains gates, null values, early-exit research, and observable scoring anchors. Several claims, however, are more confident than their evidence supports. The statement “a question cannot hallucinate a fact” ignores false assumptions embedded inside questions. Its startup example rates runway red largely from the time since the last announced raise, even though runway also depends on revenue, spending, debt, and undisclosed financing. Its pay-transparency advice—“The absence is the score”—is memorable but too absolute without first resolving jurisdiction, coverage, exceptions, and whether the posting is current.
Claude also misses the series contract in its Beginner prompt. It asks the AI to produce “three interview questions I could only ask well if I had done this research” but does not instruct the reader to rewrite those questions. The contract requires: “Anything a hiring human will read or hear … must end in the reader’s own words and be true. Prompts should drive toward drafts the reader rewrites and owns, and should say so explicitly.” The generated questions may be good, but the ownership boundary must be inside the prompt, not left to the surrounding article.
Gemini explains only a few prompt fragments per tier and repeatedly converts signals into conclusions too quickly. A three-year gap since a public funding announcement becomes “high risk,” and a tight group of positive reviews becomes a “manipulated pattern.” Both may justify further investigation; neither proves the conclusion stated. Its binary Go/No-Go framing also suppresses the assignment’s most important analytical state: insufficient evidence.
3. Template Compliance — ChatGPT and Claude tie (10 each).
Both carry the complete 57-heading structure, include every required repeated section for all three variations, keep the prompt breakdown in running : form, and use plain ## headings.
Gemini includes all 57 sections, so this is not a missing-content failure. It does, however, repeat two explicit structural violations throughout the article. The Week 4 check requires: “Every template heading written as ##, none bolded instead; prompt breakdown is running text split on :, with no ### headings inside it.” Gemini writes every heading as forms such as ## Lead and uses ### subheadings for every breakdown quotation. That is precisely the formatting pattern the pre-flight check was supposed to stop. The content is present, which keeps the score out of the seriously deficient range, but the structure is not compliant as delivered.
I also rechecked the ChatGPT post closely because a perfect compliance score for the judge’s own platform deserves more scrutiny, not less. I found no comparable missing or malformed required section.
4. Practical Utility — ChatGPT wins narrowly (9).
ChatGPT gives the reader a system that can be used immediately and maintained over time. The prompts name inputs, define evidence labels, include abstention states, preserve dates and source records, and produce next actions rather than merely summaries. The Advanced prompt’s separation of fit score, risk status, and evidence confidence is especially useful because three different questions should not be collapsed into one number.
Claude contains the single best practical design decision of the week: its 45-minute company-research cap combined with early exit. It also asks the model to identify what the time limit removed and what the reader loses by removing it. That is an unusually honest way to make a large workflow usable. The score falls one point because the Beginner prompt’s AI-generated interview questions omit the contract’s required rewrite-and-own instruction. A workflow is not fully practical if following it exactly can make the reader sound machine-written in front of the employer.
Gemini is easy to start, but several prompts are too underspecified to produce dependable results. The Beginner prompt’s “exactly 10 critical questions” imposes a number without establishing coverage. The Intermediate prompt does not require dates, source independence, contradictory evidence, or an unknown state. The Advanced prompt asks for an unexplained 100-point score and a binary verdict. A reader could act on all three today, but should not trust the apparent precision of the result.
5. Engagement & Readability — Claude wins (10).
Claude has the clearest and most memorable voice. “The Snapshot Brief takes a company name and produces questions. The Red-Flag Audit takes evidence and produces judgments. The Dossier Matrix takes criteria and produces a ranking.” That is the cleanest explanation in the set of why the tiers are genuinely different. Lines such as “That is not a shortlist, it is a recency effect with a spreadsheet next to it” turn an abstract decision-system failure into something readers recognize immediately.
ChatGPT is consistently clear and professional, but it is dense. Its thoroughness becomes repetition in several breakdowns, and the Advanced tier can feel closer to a system-design specification than a magazine article. The 34-minute reading estimate is honest, but also evidence that the post asks a great deal of the reader.
Gemini is the shortest and easiest to scan, which is a real strength. Its voice, however, leans on familiar AI-copy phrases—“flips the script,” “sinking ship,” “rigorous screening,” and “Take the next 16 minutes”—and the bold formatting inside every heading adds visual noise. It reads quickly but less distinctly.
6. Citation Quality — ChatGPT wins (9).
ChatGPT supplies relevant sources for every tier and generally matches the source to the claim: FTC guidance for fake and hijacked job listings, SEC and Investor.gov materials for public-company filings, Department of Labor material for WARN, and BLS data for industry and regional context. The named FTC job-ad guidance, Investor.gov filing guidance, and BLS 2026 QCEW researcher guide are real and support the uses described. The post stops short of a 10 because direct links and publication dates would make verification easier.
Claude’s cited sources are also real and mostly well chosen: SEC filing guidance, WARN material, the FTC review rule, state and city pay-transparency rules, and nonprofit filings. The score is lower because several consequential analytical claims exceed what those citations establish, and the broad citation to Anthropic’s own prompt-engineering documentation is less independent than the government and regulatory sources.
Gemini writes NOT APPLICABLE in all three citation sections while making current factual claims about ghost listings, layoffs, review manipulation, model context capacity, funding risk, and product security. This is thin sourcing, not fabrication; I found no invented citation because Gemini supplied none. The judge prompt explicitly says honest NOT APPLICABLE should score as thin rather than dishonest, which is why the score is 4 rather than 1.
7. Tier Differentiation — ChatGPT and Claude tie (10 each).
Both create three different kinds of work rather than one prompt at three lengths. Beginner defines what to research. Intermediate evaluates evidence and uncertainty. Advanced allocates effort across companies through a maintained system. Their inputs, outputs, decision stakes, and reader skill requirements all change by tier.
Gemini also differentiates the tiers clearly enough to earn a strong score. The Snapshot, Audit, and Matrix are distinct products. The gap is that its Advanced tier is mostly a larger comparison request rather than the living research architecture the theme explicitly invited. ChatGPT and Claude define how the dossier is built, calibrated, updated, and challenged; Gemini mainly asks the model to score the information already collected.
Winner: ChatGPT, by one point
ChatGPT wins Week 4 with 65/70, one point ahead of Claude’s 64/70. The result is close. What separates them is not voice—Claude wins that comfortably—but the reliability of the decision system. ChatGPT more consistently separates evidence quality from company attractiveness, preserves unknowns and contradictions, prevents deal-breakers from being averaged away, and supports its factual teaching with stronger citations. It also obeys the series contract in every prompt, including explicit reader ownership wherever hiring-facing language might emerge.
This is a self-win, so the burden of proof is higher. I looked specifically for places where ChatGPT’s length might be mistaken for depth and where its complexity might make the system less usable. Those weaknesses are real and cost it the Engagement dimension. They do not overturn the narrow advantage in accuracy, evidence controls, and sourcing.
The outcome would have flipped if Claude had added one explicit rewrite-and-own line to its Beginner prompt and softened the few claims that treat an indicator as a conclusion. It also would have flipped if ChatGPT had lost one more point for density or made its Advanced workflow materially harder to operate than it is to admire.
The honest counter-case
Claude did two important things better than the winner. It wrote the more engaging article, and its Advanced prompt is more conscious of the reader’s limited time. The early-exit research order, 45-minute cap, disclosed cuts, and score-clustering check may produce better real-world adherence than ChatGPT’s more elaborate architecture. A reader who values execution over completeness could reasonably prefer Claude.
Gemini’s strongest advantage is low friction. A beginner can understand its three prompts at a glance and start immediately. Its short comparison section—Beginner to screen, Intermediate to investigate, Advanced to decide—is clean and memorable. The problem is not accessibility; it is that the simplicity removes too many safeguards.
ChatGPT’s main weakness is scale. The Advanced prompt is powerful, but it asks the reader to maintain a dossier, evidence model, scoring system, portfolio view, refresh schedule, and change log. That is appropriate for a serious search, but more system than many readers will sustain. Claude is better at forcing the workflow to fit into an actual week.
Takeaway for readers
This week shows that the best research prompt is not the one that sounds most investigative. It is the one that makes the division of labor impossible to misunderstand.
The reader finds and dates the evidence. The AI defines the questions, labels uncertainty, exposes contradictions, applies a stable rubric, and identifies the next cheapest useful check. Strong prompts also separate gates from weights, confidence from fit, and missing evidence from negative evidence. Those distinctions matter more than a polished company summary because they prevent the tool from converting incomplete information into false certainty.
The final lesson is equally important: AI should remain backstage. Let it organize the dossier and challenge the reasoning. Anything a recruiter or interviewer hears should still sound like the person who did the research—because it should be that person’s own truthful words.
TAGS:
