---
id: "16.6"
title: "Insight repository"
dept: research-and-insights
status: draft
owner: Param
branded: false
last_verified: 2026-10-26
sources:
  - "memory: brew-go-live-2026-10-17, brew-absorb-all-learnings, brew-audit-2026-10-15, brew-qa-audit-2026-10-20, brew-qa-sweep-2026-10-19, brew-qa-round2-2026-10-24, brew-model-compare-2026-10-20, brew-live-stores"
  - "brew origin/main a559de6: docs/audit-2026-10-24-design-partners.md:7-30"
  - "brand-directions/_research/product-audit.md:7, 177, 255, 322-335"
  - "brand-directions/_research/sources.md:25, 31-35, 39-43"
  - "handbook: product/partner-feedback-log.md"
---

# Insight repository

## Answer

Every insight we hold, with its evidence and what it changed; all of them so far come from our own testing, code reads and audits on partner stores, none from a partner interview, so each is a finding about the product or the stores, not about how founders feel.

## Today

### About the stores we read

| # | Insight | Evidence | Date | What it changed | Source |
|---|---|---|---|---|---|
| 1 | A store's ad page name is often not its store name | Puma India and Scrubroom showed "No live ads" because the tracker searched the store name, while their pages are "PUMA" and "Scrub Room" | 24 Oct | The tracker reads the page id from the read; a tracker with no page is never 0 | [src: brew/docs/audit-2026-10-24-design-partners.md:12] |
| 2 | Not every store sells its own brand | Scrubroom: 74% other brands, 68 of 140 products resold stethoscopes | 25 Oct | assortment.js reads own label against other brands for every store; Scrubroom's Plays rewritten | [src: brew/docs/audit-2026-10-24-design-partners.md:28] |
| 3 | A store's country must come from the store, not a default | Cozy sells in the US in USD but was set to India, so all 65 of its Plays were written for India | 16 Oct | Country read from the store's text; no ₹ default, 20 Oct | [src: memory brew-audit-2026-10-15] [src: memory brew-qa-audit-2026-10-20] |
| 4 | A bare one word brand name finds the wrong news | The 15 Oct audit noted a bare one word brand name is weak for Heard | 15 Oct | Saved items re-checked on read; the brand's product words count as context. Arata Takebe news still listed as open on 16 Oct | [src: memory brew-audit-2026-10-15] |
| 5 | Some store sites lack their own social links | No Instagram or YouTube on the site made Films say none | 24 Oct | Accounts found by search, labelled so | [src: brew/docs/audit-2026-10-24-design-partners.md:11] |

### About what we make

| # | Insight | Evidence | Date | What it changed | Source |
|---|---|---|---|---|---|
| 6 | An image model copies any rival picture it is given | Probes: gemini-2.5-flash-image copied the rival's words, products, model and badge | 17 Oct | Make our version sends only a written breakdown of the rival ad plus our own product photo; the rival picture never goes to an image model | [src: memory brew-go-live-2026-10-17] |
| 7 | A plan from anyone is a learning | The Arata PDF draft skipped the whole campaign plan because the prompt ignored suggestions from us | 19 Oct | Learnings absorb every plan whoever wrote it, and update the Plays and cards it touches | [src: memory brew-absorb-all-learnings] |
| 8 | Card numbers drift from the accounts they name | Round 2 report on Arata | 23 to 24 Oct | A card naming a connected metric takes the account's number; the Fact Sheet and Brand Book are the source of truth | [src: memory brew-qa-round2-2026-10-24] |
| 9 | Unbacked figures were printed as fact on Play covers | Scrubroom's cover said ₹2,625 entry set price; sets start at ₹1,899; 7 flagged covers still printed | 24 Oct | stat-check.js checks covers against catalog and reviews, unbacked figures marked to confirm | [src: brew/docs/audit-2026-10-24-design-partners.md:13] |
| 10 | Opus writes more brand specific Plays than Gemini Pro, but the test was unfair | 54 pairs on Arata: Pro reused one stat on 8 Plays; Opus cited brand stats; Opus also had the team's 19 Oct audit as input | 20 Oct | Nothing yet; a fair rerun and a bake off are proposed, not run | [src: memory brew-model-compare-2026-10-20] |

### About the product's shape

| # | Insight | Evidence | Date | What it changed | Source |
|---|---|---|---|---|---|
| 11 | The build is too many places for one job | 89 surfaces; 16 ways into a jam; about 11 screens and up to 14 decisions before the first card | 25 Oct | Fed the redesign brief | [src: brand-directions/_research/product-audit.md:177] [src: brand-directions/_research/product-audit.md:255] |
| 12 | One label does two things | Send brief emails our team in one place and opens your own mail app in another | 25 Oct | Listed as a hotspot; no fix recorded | [src: brand-directions/_research/product-audit.md:331] |
| 13 | Speed breaks when checks run per block | The claims check normalised the whole catalog per block; brand pages took 17 to 50 seconds | 19 Oct | Memoised; time the seed after every change on that path | [src: memory brew-qa-sweep-2026-10-19] |

### From outside research

| # | Insight | Evidence | Date read | What it changed | Source |
|---|---|---|---|---|---|
| 14 | Game parts copied without a reason do nothing | Duolingo's borrowed mechanic was neutral | 25 Oct | Lead with meaning, progress and making; scarcity and loss only when real and kind | [src: brand-directions/_research/sources.md:31] [src: brand-directions/_research/sources.md:39-43] |
| 15 | People leave between phases | Octalysis summary of the four phases | 25 Oct | Design each join on purpose: the reveal, the first win, brand memory | [src: brand-directions/_research/sources.md:25] [src: brand-directions/_research/sources.md:43] |

### Gaps

- No insight comes from a partner's own words or from watching a founder use the product [src: handbook research-and-insights/interview-and-call-library.md]

## Tomorrow

| Item | Version | Tag | Notes |
|---|---|---|---|
| Insights from partner calls and usability sessions | v1 | New | Each with the call row it came from |
| A status per insight: acted, open or dropped | v1 | New | Open ones go to Next |
| Insights 10 and 12 settled | v1 | Changes | After the fair model rerun and the redesign |

## How we do it

1. Add a row when a test, call, audit or probe shows something that would change what we build.
2. Write the insight in one line, the evidence with numbers, the date, what it changed and the source.
3. If nothing changed yet, say "Nothing yet" and add it to the Open part of the document it belongs to.
4. Never write an insight about how founders feel without a call or session row behind it.

## Numbers

| Metric | Definition | Where measured | Status |
|---|---|---|---|
| Insights | Rows on this page | This page | 15 |
| From partners | Rows whose evidence is a call or a session | This page | 0 |
| Acted on | Rows with a change in the product | This page | 13 of 15; 10 and 12 not yet, 4 only in part |

## Open

1. **Should insights from QA count here, or only from people?** Proposal: both, kept in separate sections, so the people section shows how empty it is.
2. **Run the fair model rerun behind insight 10?** Proposal: yes, five Plays on Pro with today's inputs, about $1.
3. **Who owns this page?** Proposal: Param, with Sai adding rows after each QA round.