FishDogResearch studies · Research on researchv1.0 · frozen 2026-07-10 · changelog: initial release
Shared research study link · Fielded Jul 10, 2026
What Report Buyers Want
Three panels of senior research consumers — investment professionals, C-suite operators, and the marketers who commission studies — on what makes them read, trust, and act on outside research. This page is built in the format the panel specified. The report is the demo.
Lead every FishDog deliverable with the disagreement structure — who holds which view, how strongly, and what would flip them — anchored on a two-page decision brief with an exportable data pack. The interactive page is the audit layer behind it, never the deliverable itself.
Basis: 38/38 on divergence-first (Q5) · 36/38 on the brief as primary format (Q6) · 0/38 accept a web page as system of record.
Headline finding · Q5: consensus-first vs divergence-first
Thirty-eight buyers. One answer.
Shown the same study opened two ways — "78% of experts agree" versus "Ten professionals. Three positions." — every respondent in every panel said they would trust and act on the version that leads with the disagreement.
Investment professionals9 analysts · 4 advisors
13
13/13
C-suite & operators2 CEOs · 9 senior managers
11
11/11
Insights & marketing buyers14 research commissioners
104
14/14
34 choose divergence-first outright4 skim consensus to triage — but trust and act only on the split0 act on consensus-first
Q5 of 7: "Version A opens: 'Executive summary: 78% of experts agree…' Version B opens: 'Ten professionals. Three positions.' Which would you rather receive, which would you actually trust and act on, and why?"
Read with careA 38/38 result is rare and should raise an eyebrow — ours did. Four respondents independently noted that a ten-expert study is directional, not proof. The panel did disagree, just not here: the splits live two sections down, and the conditions below are what keep this result from flattering us.
The conditions · Q5, volunteered
Divergence only converts with the receipts attached
The unanimous vote came with strings. Respondents volunteered what the split must ship with before they will act on it — "If Version B turns into personality theater without data, I bin it too." Counts are exact tallies; a dash means the panel wasn't separately tallied on that condition.
Who holds which viewrole, segment, exposure — so views can be weighted, not counted
38/38 all panels
Conviction per camphow strongly each side holds its view, not just headcount
22/24 finance 13/13 · ops 9/11 · insights —
What would flip each camptriggers and falsification — the line buyers put on their own IC slide
19/24 finance 12/13 · ops 7/11 · insights —
Panelist track recordcalibration and hit rate of the people behind the positions
9/13 finance only
And the counterweight: 21 of 25 operators and insights buyers act on the split, then hand-build a clean consensus-style one-pager for the level above them. The consensus line isn't wrong — it's the output the reader manufactures, not the input they trust. Format v2 generates it for them: see the board copy below.
Consensus and contest
Where the panel actually split
Ranked by disagreement. The format question was a landslide; the fights are about trust mechanics and delivery around it.
How they verify before trusting — rebuild the math in Excel vs. call vendors and check reality vs. photos and checklists
The analyst walkthrough — most want a short post-read Q&A; a hard minority won't act on spoken numbers ("it evaporates")
The interactive page's real users — active second stop vs. mined once then frozen vs. near-rejection (offline, compliance)
Does consensus have any job at all? — 4 of 14 insights buyers want it as Monday-night triage and "political air cover"
Primary working format — the two-page brief, against one deck loyalist and one interactive-first analyst
Divergence-first vs consensus-first — the headline: no contest anywhere
The argument
Two ways of deciding whether to believe you
The deepest split in the study isn't about format at all — it's about what counts as verification.
The methods audit — 26 of 38
"If I cannot rebuild the key number in Excel in under 10 minutes and explain the caveats on one slide, I will not take it to my CFO."
Tania Rogers · financial manager, 48, Fishers IN — with the entire finance panel behind her: "If I cannot replicate one chart in under 5 minutes, I will not use it."
VS
The reality check — 9 of 38
"I did not read the methods section. If the lines on the chart match what my vendors are quoting, that is enough."
Clinton Cromley · construction manager, 54, rural NY — two phone calls beat an appendix. Jay Gutierrez goes further: "Photos and checklists are what move me."
Both camps need the same artifact: numbers that survive checking against something outside the report. The appendix serves the auditors; named sources, dates, and local specificity serve the reality-checkers.
Kill signals · Q2
The tells that read as machine-generated filler
Asked what makes them dismiss a report as low-effort or generated, the panels named the same fingerprints. The reframe that matters: "I do not care who wrote it — human or machine — if it is thin, vague, or unverifiable, it is not getting into my IC deck" (Destiny Torres, 31, investment analyst). The objection is verifiability, not provenance.
1
Uniform prose cadence"same sentence skeleton page to page… no fingerprints from a human analyst"
30/38
2
Too-tidy numbers"Real data is ugly. If it is all tidy, it is modeled fluff." Round deltas, insights in perfect threes
16/38
3
Phantom apparatuscitations that don't exist, "see Figure 14" with no Figure 14, dead links, copy-paste scars
13/38
4
Sanitized verbatims"Real verbatims have burrs." Quotes with no slang, no typos, identical lengths — read as composites
11/25 where tallied
5
Swap-the-sector test"swap the sector name and the paragraph still reads fine. Feels copied, not thought."
8/24 where tallied
Tell #4 is the one that cuts closest for synthetic research: buyers detect authenticity by quote texture. Every verbatim on this page is untouched — burrs, Spanish, raccoons and all.
The report's afterlife · Q4
Nobody forwards the report
All 38 described the same pipeline. The original document has a working life of about 48 hours; its afterlife is as fragments inside the buyer's own artifacts. Adoption is decided by how well a report survives strip-mining.
Step 1 · minutes
Triage
Date, N, region, first chart. Most reports die here: "most outside reports hit my inbox and die there."
Step 2 · same day
Verify
Rebuild a chart in Excel, call a vendor, check the CRM. 38/38 check against something they already trust.
Step 3 · 48 hours
Strip-mine
38/38 lift 2–3 charts and a few numbers; many rebuild them in house style and drop the vendor's branding.
Step 4
Re-author
38/38 never forward the original. CFO gets one page; specialists get the appendix; frontline gets "two sentences and a photo."
Step 5 · quarterly
Grade
Archived with dated filenames, checked against what happened. "If it diverges twice, I stop opening their emails."
Format v2 device · generated from the distribution
The board copy
Twenty-one respondents told us they hand-build a consensus-shaped one-pager for the level above them after deciding off the split. So the format now ships it: one page, liftable, print-ready, always labeled as the summary layer over the disagreement — never the substitute for it.
FishDogBOARD COPY · 1 of 1 · v1.0 · 2026-07-10
Research buyers want the disagreement first — and the numbers rebuildable
Across three panels of senior research consumers (n=38: investment professionals, C-suite operators, insights buyers), every respondent chose a research report that opens with the structure of expert disagreement over one that opens with a consensus statistic.
38/38 trust and act on divergence-first presentation. Consensus headlines read as smoothing; the dissenting minority is treated as the risk signal.
36/38 work from a two-page decision brief. Slide decks are quarried for charts; interactive pages are audit tools. 0/38 accept a web page as the record.
Trust is mechanical: visible N and field dates, exportable data, and numbers the reader can rebuild. Reports that fail the rebuild test do not get cited.
Action: FishDog deliverables lead with the split (positions, conviction, flip-triggers), anchor on a two-page brief plus CSV data pack, and version-freeze every artifact.
Summary layer over a contested distribution — full split, verbatims and limitations: fishdog-report-lab.pages.dev/report-buyers · Synthetic panel study, fielded 2026-07-10 · Directional; human validation pass scheduled.
Recruited, not generated, from a census-grounded synthetic population of 340,000 U.S. residents. Hover any respondent for their position in their own words.
Respondents
38 of 68
recruited then curated — 30 recruits cut for weak fit, disclosed below
Fielded
Jul 10
2026 · 7 questions · 266 curated responses
Finance panel
13 of 40
hand-picked investment professionals from a 40-agent over-recruit
Not recruitable
2
cohorts we wanted and could not honestly field — see the candor note
n=14 · commission and consume external research · study 1071
Katherine Eng 40Charles Smith 54Thomas Rahman 39Cynthia Molina 60Lavera Magana 40Michelle Shuford 42Walden Leiva 46Eric Pickering 41Kathy Yu 53Brenda Araujo 39Marcus Gonzalez 38Joseph Reynolds 50Jennifer Holm 42Tiffany Beltz 52
divergence-first outrightdivergence to decide, consensus copy for upward or triage
Recruitment candor. Two cohorts we wanted, we could not honestly field. "VP and up": the hydrated pool tops out at manager-level titles for marketing — the insights panel is managers who genuinely commission research, not VPs. Hedge-fund portfolio managers: a 40-agent over-recruit of finance occupations surfaced analysts and advisors, no fund PMs — we hand-picked the 13 real investment professionals and dropped the other 27. A third panel drifted on spec: "media & entertainment executives" returned general management; we reframed it as the C-suite/operator lens and made no media-vertical claims from it. Skews and substitutions are disclosed here, not dressed up.
Findings
What earns trust, with the evidence attached
38 OF 38
The rebuild test decides citation
"If I cannot rebuild the key number in Excel… I will not take it to my CFO."
Every panel demands visible N + field dates on page 1 and an exportable data pack. Several put a stopwatch on it: 5 to 30 minutes.
36 OF 38
The two-page brief is the working artifact
"Twenty-five slides to say 3 numbers."
The deck survives as a chart quarry; the brief carries the decision. One deck loyalist (Gonzalez), one interactive-first analyst (Reynolds).
0 OF 38
A web page is never the system of record
"What if your portal updates Sunday night and I walk into Monday quoting ghosts?"
But near-all want it behind the brief: exports on every view, version-locked snapshots, visible N, drill to verbatims, no login.
6 OF 38 · UNPROMPTED
Say where the AI was used, and how it was checked
"If a model summarized verbatims, a human signed off. No machine seams."
A small unprompted 2026 signal across panels. Cheap to ship, differentiating — this page does it two sections down.
Format v2 device · AI provenance
How this study was made
Respondents
Synthetic. Recruited from FishDog's census-grounded population of 340,000 U.S. residents; each carries a grounded biography, occupation, region, and an ingested media diet. Recruited, not generated to order — the skews above are real recruitment outcomes, disclosed.
Fieldwork
7 open-ended questions per panel, asked one at a time via the FishDog API on Jul 10, 2026. Two format stimuli described neutrally in text; FishDog never named. 266 curated responses.
Analysis
Three independent AI analysis passes (one per panel) under an exact-tally rule: every count is an enumerated per-respondent tally, never an impression. Human-directed synthesis followed.
QA
Headline verbatims spot-checked against raw transcripts; share links verified; all counts re-tallied from the curated panels only. One dispatch bug found and disclosed: removed agents still received questions; their answers are excluded from every number on this page.
Qualitative, n=38, synthetic. Unanimity this clean warrants suspicion; four respondents said so themselves. Next step is a human validation pass: render one live client deliverable in this format and ask the client which version they would act on.
Appendix
Full narrative
A1
Objective & design
FishDog's presentation workstream needed evidence for how research deliverables should be structured: what content senior buyers want, in what structure and format, and what builds or destroys trust. Three panels were fielded to map how preferences diverge by buyer type — the working hypothesis, from client conversations, being that report consumption differs between raw-data buyers, report-shaped buyers, and interactive-leaning buyers.
Battery: Q1 last influential report and how it was actually consumed (screening) · Q2 stop-reading triggers and machine-generated tells · Q3 trust requirements before acting · Q4 preferred delivery form and the report's afterlife · Q5 stimulus, consensus-first vs divergence-first opening · Q6 stimulus, deck vs brief vs interactive page · Q7 magic wand, design the perfect deliverable.
A2
The trust stack, cross-panel
Visible methodology on page 1 — N, field dates, sample frame (38/38). The single most-cited kill signal is having to hunt for the denominator.
Rebuildability (38/38). "If I cannot rebuild it, you are selling theater." The data pack is not a nice-to-have; it is the citation gate.
Ranges, never hero numbers (38/38). "Single-point optimism is useless to me."
Falsification stated plainly (33/38). What would change this call — the sentence buyers lift onto their own IC slide.
Limits and failure modes on the page (35/38). "If there is no scar tissue, I do not believe you did the work."
Segment fit (36/38). The finding must cut to the buyer's region, size band, or role, or it reads as someone else's study.
A3
Confidence grades
High: divergence-first preference (38/38) · brief-first working format (36/38) · rebuild-before-cite behavior (38/38) · strip-mine afterlife (38/38) · no-gating requirement (38/38) · uniform-cadence as machine-tell (30/38).
Low / watch: per-panel differences in interactive-page appetite; the deck-first and interactive-first outliers as possible segment seeds.
Grading rule: High = 7+ of 10 equivalent converging with specific language; Medium = 4–6 or concentrated in one panel; Low = under 4 or reliant on articulate individuals.
A4
The format rules this page follows
Scope strip above the fold: N, dates, method, funding, exports. (Kill-signal #1: hidden denominators.)
Decision statement before findings. (Page 1 leads with the decision, 14/14 insights buyers.)
The split leads, with conditions: who, conviction, flip-triggers. (38/38 + the volunteered conditions.)
Unanimity reported with suspicion, contested questions ranked honestly. (Too-tidy results read as modeled fluff.)
Verbatims untouched, tagged, placed next to the finding they evidence — one challenging quote where the data allows. (Sanitized quotes are a machine-tell.)
Exact tallies with denominators, including "where tallied" gaps. (No round-number smoothing.)
Board copy generated, labeled as the summary layer. (21/25 hand-build it today.)
Version-frozen, no login, everything exportable, print-safe. (38/38 anti-gating; grayscale-checked.)
The engineering asks that fall out of this — per-response stance + conviction + flip-trigger classification, CSV export on shared studies, frozen share snapshots, the board-copy generator — are in format-v2-spec.md alongside this site.
FishDog · Research without respondents. Study fielded Jul 10, 2026 · 38 curated respondents across three panels · 7 questions · 266 responses ·
v1.0, frozen 2026-07-10, no silent edits — changes get a changelog line in the header.
Built in report format v2, which this study itself specified. Stance classifications on this page are analyst-tallied from raw responses (per-response classification is the pending engineering primitive); quotes, demographics, and counts are from the study record.