How we got to a product call — and what's still unsettled.

00What this is

CORA is a Spain-first women's daily-nutrition brand. The deck had already converged on a single-SKU launch — a complete-protein coffee creamer (Option 4) — off a thin survey (n≈93, badly Brazilian-skewed; the de-biased Spanish core was n≈12). This project pointed a proven social-listening engine at the category to test that bet against thousands of real conversations rather than re-confirm it.

The whole exercise was run under one rule: answer CORA's open questions, don't re-derive the deck. That meant being willing to have the data contradict three sessions of our own thinking — which, twice, it did.

01The engine we re-pointed

We did not build from scratch. A listening + analysis pipeline was already built and proven on the GLP-1 category (13,886 items → a real product thesis). It scrapes four platforms (Reddit, YouTube, TikTok, Instagram via Apify), triages every item with a cheap model (the "census" — tagging concern, journey phase, emotion, geography), escalates the highest-intent items, and stores everything in a re-queryable Postgres corpus where the raw text is immutable and the tags can be re-run.

Re-pointing it at CORA was a configuration, not a rebuild: a new database (nm_cora_datalake), a new concern taxonomy, a new life_stage cohort axis, a new geo logic that separates Spain from LatAm-Spanish from Brazilian-Portuguese, and a connector re-weighting toward Instagram/TikTok (where Spanish women's conversation actually lives) over Reddit (which is English-dominated).

02The pivot: velocity → durability

The original brief specified a velocity-first analysis — rank concerns by growth, because CORA wanted to ride the fastest-growing opportunity it could still catch. We changed that, deliberately, when the founder chose to optimise for a durable need rather than the steepest riser (decision #01 below).

The reasoning: in a category that's filling fast (creatine-for-women, the gummy/stick explosion, Spanish brands already entering), the steepest slope can simply mean you're late. A large, stable, emotionally-hot need you can own for a decade beats a fad you catch at its peak. So the headline ranking became a durability composite — size × emotion intensity × concentration at a buying moment — and velocity was demoted to a health-check column: "is the durable thing we want to anchor on rising, flat, or fading?"

03The 21 assumptions we put on the table

Before spending a euro, we inventoried everything we were quietly treating as settled, so we could challenge it. LB marks load-bearing — wrong here and the run points the wrong way, not just noisier.

04The four decisions we forced before scraping

Four strategic calls were settled in writing first, so the data couldn't be cherry-picked to fit whichever answer we secretly wanted.

1B reshaped the whole extraction. It also, combined with 4A, committed CORA to the slower category-builder road — launching into the cold, price-sensitive Spanish core and consciously demoting Rafa's warm BR network from launch market to seeding asset. A defensible bet, made knowingly.

05The scrape, step by step

The engine was cloned, the new schema applied, the read-only role proven SELECT-only, and — the non-negotiable lesson from GLP-1 — nm_cora_datalake was added to the backup script and a real test dump run before any scraping. Five Spanish creator anchors were verified (Boticaria García, Marta Marcé, Cristina Mitre, Ismael Galancho, brand Woments) and three French ones; no handles were fabricated.

What went right and what bit us

• Reddit comment-tree blowout. We set Reddit to maxItems=15 to keep it a modest English reference. The actor caps posts, not comments, so fetchPostComments pulled full trees → Reddit ballooned to 13,954 items (69% of the corpus). We trimmed it back to ~2,738 (all posts + top threads by engagement) so it stayed reference, not driver.
• Instagram scraped thin. The hashtag-search input format was wrong for the actor; it returned ~385 items. IG — our intended primary Spanish channel — underperformed. TikTok (2,200) carried the social load instead.
• Cost discipline held. First run ≈ $23 combined, under target, after the Reddit trim.

06The geo gate — honoured, not waved

The pre-registered rule: don't draw a Spain conclusion unless ES-Spain clears 25% of geo-identified items. First run came in at 23.4% — just under. We did not wave it through despite a favourable early ranking. Instead we ran a cheap, surgical Spain top-up: comment-scraping the five Spanish anchors' posts (where Spain-Spanish conversation concentrates) plus a Spain-only YouTube re-query.

Final geo: ES 3,732 · LatAm-ES 3,485 · unknown 2,735 · US 2,440 · FR 277 · UK 140 · other 134 · BR 82 (quarantined). LatAm-ES is held as a Spanish-language reference tier — a far better proxy for Spain than English Reddit, because concern structure transfers across dialects better than across languages.

07What the corpus actually says

The Spanish read (ES only), in durability order, with the velocity health-check beside it:

The model also proposed concerns we hadn't seeded — calm_energy (promoted to a real tag), plus caffeine_non_responder, task_paralysis, hormonal_rage, hair_loss_adverse_event — the whitespace worth watching.

08Two reversals — where the data told us we were wrong

Net: the corpus splits into two coherent territories, not one fused need. The original Option 4 straddled both; pulling creatine narrowed the product onto the harder-to-defend emotional side.

09The reads closed — and reading the language corrected us done

The three diagnostics and a full read of all three analysis layers (skeleton, verbatims, threads) closed the open questions and corrected three things the numbers alone overstated:

1. Buyable vs. ventable settled it. Creatine 0.95 buyable (171 of 275 seeking) — the substance. Stress 0.47, mostly venting (157 of 402) — the wrapper. Stress is not a hero.
2. Creatine demand is anxious, not confident. The verbatims show the "seeking" is safety-seeking — "¿no engorda, no hincha?", "¿se puede tomar en café caliente?", "me empezó a dejar calva", drug-interaction and breastfeeding fears. The hero job is removing fear, not selling strength.
3. Perimenopause debiased out. 30.4% of ES peri mentions are anchor comments (~10× creatine's 2.9%). Its Spain-strength was a seeding artifact. Discount it.

The language also surfaced the fork the numbers hid: creatine demand skews 40+; the 40+ cohort is quitting coffee ("café no tomo," matcha "sin tembladera ni bajón"); Rafa reaches sub-40. Substance, format and channel point three ways. And one external check remains: Dan's ES search cross-check — the only clean growth read, since the velocity slopes are untrustworthy (sources only switched on mid-2025).

10What's settled — and the one decision left cohort × format

Most of the product is settled by evidence; one strategic fork is deliberately left to the founders. Settled: creatine + protein is the buyable substance (reassurance-led demand); stress is the emotional wrapper, not a SKU; brain fog / "claridad mental" is the through-line benefit; single SKU, not a range; perimenopause discounted; the deck's creatine-out decision reversed (false premise).

Full reasoning in the Strategic Report v2; present-live version in the Dashboard v2 — all three documents consistent on what's settled and on leaving §07 open.

11Next steps

1. Close the three open reads (§09) + Dan's search cross-check.
2. Lock the product call: single SKU vs. straddle, hero substance, hero emotion, format.
3. Build the strategy report and the dashboard (GLP-1 shape) on the settled call — deliberately held until the call is locked, so the two artifacts never contradict each other.
4. Re-open the deck's creatine-out decision with the founders, given Reversal 1.
5. Run-2 hygiene: a clean perimenopause read without peri-creator over-weighting; more Spain-targeted TikTok; bump the Apify plan off the $29 Starter cap; add the DNS A record for the dashboard.

12Lessons (carry into the next category)

• Anchor-seeding inflates whatever the anchors talk about. Seeding creator comments to get Spanish volume imports those creators' agendas. Always debias by anchor flag. We caught creatine; we nearly trusted an inflated peri number. This is the new general lesson.
• Reddit's maxItems caps posts, not comments. "Modest Reddit" needs a comment cap or fetchPostComments turns a reference into the whole corpus.
• Instagram hashtag-search is unreliable; comment-scraping known anchors is the high-yield Spanish lever.
• LatAm-ES is a real reference tier. A better Spain proxy than English — same language, closer culture.
• Honour the gate when the answer is the one you wanted. We were just under 25% with a favourable ranking — exactly when a rule becomes a decoration if you wave it through. We didn't.
• A scrape ranking is a starting point, not a verdict. The two findings that mattered most (creatine debias, weak cluster) came from interrogating the ranking, not reading it.

13Infrastructure & cost