Research Playbook — finding winners and studying competitors

Where to find ads that are winning, how to tell they are winning without spend data, how to study a competitor end to end (ads → landing pages → product pages → offers → quizzes), how to mine the voice of customer, and what each step costs. Everything runs on the VPS (ssh vps-claude), never in the Mac's browser.

---

1. Source map

---

2. TrendTrack recipes

Statics of a domain, newest and longest-running (/root/workspace/competitors/tt_statics.py <domain> <slug>; tt_dl_only.py <slug> re-downloads without credits):

search=[domain], searchType="domain", mediaType="image", status="all", sortBy="newest", order="desc"
→ newest 30
same + minDaysRunning=<threshold>  → the "most performing" proxy

Choose the threshold by probing totals (limit=1) at several minDaysRunning values until ~30–45 ads remain. Thresholds used: Primal Viking ≥ 200 days (32 statics), Mars Men ≥ 200 (39), Ancestral ≥ 200 (44), Primal Storm ≥ 135 (29 — young brand).

Why run time and not reach: sortBy=reach returns 0 for US ads and sortBy=longestRunning returns empty (API bug). Run time is the honest proxy.

MCP tools for a brand overview (Step 1B of the blueprint): search_advertisers / lookup (pages of the competitor AND our brand) · search_ads / get_brandtracker_scaling_ads (active, past, scaling) · get_brandtracker_transcripts (video transcripts already indexed) · analyze_tracked_brand / brief_competitor (strategic overview) · search_shops / analyze_brand_changes (traffic, trend, offers) · search_emails / analyze_shop_emails (email sequences).

Freshness filter that works (product discovery): don't trust shop createdAt. Use the weekly advertising.history: keep shops where ads ≈ 0–10 around 60 days ago and active ads ≥ 50 now; or search_ads with created_after = today − 60 d + sort_by reachDelta7d (ad firstSeenAt = real first run date).

Givenche's three hacks (ads > traffic; traffic data lags 30–60 days):

1. Recent shops sorted by active ads (many show 0 traffic but are scaling).
2. AppLovin pixel on shops < 3 months old (insiders scaling on a new network).
3. Recurring products (subscription apps installed: Kaching Subscriptions, Recharge, Seal…) = big LTV.

Scaling calculation (scaling-shops method): last 5 weekly points of advertising.history, ≥ 3 points, last > first, growth ≥ +50 %. Collapse mirror shops (identical ad history) into one card.

---

3. Meta Ad Library via Apify (the full-library pull)

Actor apify~facebook-ads-scraper, parameters used (/root/workspace/competitors/adlib/collect.py): active_status=all, ad_type=all, country=ALL, query = the landing DOMAIN, sort_data[mode] = total_impressions, sort_data[direction] = desc, limit per set (e.g. 1,200 for small brands = all ads; 220–320 for big ones). Second set per big brand: the 100 most recent launches (sort_data[mode] = time_active for "newest").

What you get: ad id, page, start/end, active flag, variants (DCO), copy, titles, link, media URLs. Downstream (adlib pipeline): media.py (download, B2 copy), transcribe.py (videos via API), analyze_ads.py (per-ad fields: hook, problem, root cause, mechanism, angle, proof, offer, summary), lpscan.py (landing pages), site/classify.py (fixed taxonomy), site/build_data.py (site data).

Budget share estimate: Meta's impression ORDER is real (checked 29/29 against the live page) but US numbers are hidden. Weight each ad 1/rank^0.8 and sum by angle → "≈ % of budget". Always label it an estimate.

Never: run Meta keyword searches on generic words (they return 50,000 unrelated ads); run two Meta scrapers from the VPS at once.

---

4. Radar — following one competitor day by day

Built for Resilia (/root/workspace/radar/, site radar.machineroom.app). Use when the goal is to learn a competitor's STRATEGY over ~20 days: winners, cut speed, iterations, angles per product, pages, funnels.

Rules learned:

• Never base the footprint on pages. Search by the landing domain (q=resilia.shop, cross-checked with the brand name), keep every ad whose REAL link (the l.facebook.com/l.php?u= href) points to the domain. The brand ran ~17–38 pages, most disguised as health-review sites ("Vascular Wellness Report", "The Blood Sugar Report"). Exclude resellers (Amazon, other shops) and homonyms.

• Dedupe by creative (video hash), not by ad id: Meta runs the same video under many ids ("N ads use this creative" = N ads).

• Product identification from the ad text + landing path (each product had its own description pattern and pre-landers).
• Meta gives lifetimes to the hour ("Total active time N hrs"); big advertisers launch in waves several times a day (Resilia ~150–220 new creatives/day) → at least 2 passes/day.
• Current setup: TrendTrack-only (cron 4×/day for new ads, daily cuts and pages), ~250 credits/day, caps 400/pass and 9,000/period, ledger data/<domain>/tt_ledger.json.
• Winner signals without spend: survival > 7 / 14 days, growing duplicates, number of pages carrying the creative, relaunches/iterations.

---

5. Signals that an ad (or a brand) is scaling

Our own winners (account data): ROAS above break-even and ≥ 10–15 % of monthly spend (20–25 % = very big), confirmed by month-2 rebill. Never judge before 72 h.

---

6. Voice-of-customer mining (always, for every new product)

Evolve's order: Reddit → YouTube → TikTok comments → Amazon reviews → post-purchase survey → AnswerThePublic. Actors that work from the VPS datacenter IP (no residential proxy):

Script: python3 /root/research_scrape.py <slug> [platforms] reading products/<slug>/research-config.json. Query patterns: desires "I want [outcome] reddit" · emotions "[problem] makes me feel reddit" · product experience "anyone tried [solution] reddit" (capture the OUTCOME) · demographics "at what age does [problem] start" · horror "[solution] gone wrong". Capture in 3 columns: Desire | Customer language (verbatim) | Ad idea. Report every scraping failure explicitly (blocked actor, aborted run, empty dataset); never hide it.

---

7. The market research report (Step 1 deliverable — exact structure)

1. Top 5 angles — table: # · Angle · The idea (2–3 sentences including the mechanism) · Weight (# ads · copies · run time).
2. 5 secondary angles — one line each (common leads: results timeline, origin/heritage, scarcity/authenticity, alternative mechanism, the expert who let the customer down).
3. Top 5 headlines — word for word, with source ad and angle.
4. Headline bank — by category: Pain & mechanism · Comparison · Product · Story · Urgency.
5. Video hooks — word for word.
6. Avatar — age, situation, visible problem, what they tried and spent, everyday shame behaviours.
7. What proves it to them — everyday proof objects they'd notice themselves.
8. Winning formats — static/video split, dominant formats.
9. Their offer — price, subscription, guarantee, social proof.
10. Their weaknesses — backed by numbers (traffic trend, live ads trend).
11. Recommendation — the 3 angles and 10 headlines to attack first.

Process (blueprint 1A → 1B → 1C): (1A) for EACH ad one row: format · exact headline · sub-headline · video hook (first 3 s) · angle · mechanism · visual format · proof · offer · CTA; group by angle; rank by weight — give the per-ad table first, then the synthesis. (1B) TrendTrack cross-check with every number sourced. (1C) the report above. Get sign-off before step 2.

---

8. Competitor study — full method

1. Scope: small advertiser → 100 % of ads (active + inactive). Big advertiser → top 150–250 by impressions + 100 latest launches. Statics study → newest 30 + top 30 by run time (+ more pages if an angle needs it).
2. Collect (Apify or TrendTrack), save raw JSON on the VPS, media to B2 (mr-archive).
3. Transcribe videos with mr-transcribe (API). Beat-by-beat analysis only if asked (it's expensive); full transcripts are enough for the global analysis.
4. Per-ad analysis (Workflow B table in SKILL.md): hook, problem, cause, mechanism, angle, proof, offer, landing.
5. Landing pages: read every one 100 % (method §9), map ad → LP → PDP.

6. Classify into BIG angles with a taxonomy fixed before classifying. The testosterone taxonomy used for 1,295 ads: angles = libido · size · belly · man boobs · face/jaw · energy · hair · muscle · 40+ · generation · digestion · GLP-1 · alcohol · comparison · generic T · other; mechanisms = aromatase · cortisol · deficiency · absorption · endocrine disruptors · SHBG · nitric oxide · DHEA/precursors · age decline · none · other. Promo ads get the angle underlying the promo.

7. Weight angles (rank-weighted share, run time, variants, LPs).

8. Routing: which angle goes to which landing page.
9. Strategy read: pillar, current push, tests, cadence, cuts, new angles, weaknesses, what to copy.
10. Deliver a site per brand (English, big, simple, all images as small thumbnails _m.jpg for slow networks, "Which landing page for which angle" section).

Example costs: 4-brand Meta library study = Apify $8.63 + OpenAI ≈ $38 (analysis with a GPT model) for 1,295 ads, 67 LPs, 21 PDPs, 238 videos. 4-brand statics study ≈ 330 TrendTrack credits.

---

9. Landing page teardown

For each landing page (advertorial, listicle, quiz, PDP, comparison, VSL):

1. Type and headline.
2. Flow, block by block: problem → cause (UMP) → failed solutions → mechanism (UMS) → bridge (story/authority) → proof → offer. Quote the key line of each block.
3. Which ads send traffic here (count) and which angle they carry → congruence check.
4. Where it sends next (PDP URL), and the PDP's prices, compare-at prices, bundles, subscription, guarantee, gifts, countdowns.
5. Persona call-outs used ("Men 45–65", "for men who've tried everything").
6. Compliance posture (site is stricter than ads). Captures: Playwright mobile 390×844 on the VPS (nice -n 19 ionice -c3), forcing US locale with Shopify cookies (localization=US, cart_currency=USD) when the VPS IP geolocates elsewhere.

Patterns seen: "one landing per villain" (Primal Viking) · one "10 reasons" listicle cloned per angle (Mars Men) · advertorial "I tested the 2 top-selling…" for comparison angles (Primal Edge, Elora) · persona-specific PDP headlines ("For Men With Man Boobs Who've Tried Everything") · proof ads → proof page · media co-branded listicles (don't copy).

---

10. Quiz teardown

Method (quizmap study): crawl the quiz end to end on a phone-sized screen; click EVERY answer of every question; screenshot every screen; draw the tree (a straight line if all answers lead to the same next screen, a branch otherwise); run several full profiles to the result and landing page; then compare word by word what each answer changes ("dead questions" = no visible effect). Intercept submissions so nothing is sent to the competitor.

What to record: opening promise · length (questions, loaders, proof screens) · question formats · branches · proof inside the quiz · result screen (verdict levels, named root causes, score, projection chart with a date) · offer screen · landing page and whether it uses the answers · email capture · data passed in the URL (persona) · live A/B tests (e.g. "SKIP QUIZ & BUY NOW") · mistakes.

Takeaways that generalise: ~10 one-tap questions; a verdict with 3 levels + a named root cause the product answers + a dated projection tied to the offer length; cheap "made for me" hooks (dream-body picture, event + date countdown); nobody passes the quiz; test a skippable email step; actually USE the persona on the landing page (both competitors collect it and throw it away); every question should change the result or feed a segment; avoid diagnosis wording.

---

11. Product discovery (when the brand is not chosen yet)

• Two strategies: traffic curve ("the curve that explodes", lagging) vs ads-first (Givenche, faster). Ask which one to run.
• Scaling-shops run (28 Sep): shops created 15 May–15 Jun with ≥ 10 active ads → 4,302 domains → 1,276 scaling (≥ +50 % in 4 weeks), 145 new in health; dominant new health angle: lymphatic drainage drops (5 brands).
• A watchlist product is never launched without an explicit go. Keep candidates in a list with the shop URL and TrendTrack id.

---

12. Budgets and etiquette (hard rules)

• Give every research agent a credit budget; keep a TrendTrack floor (≥ 1,000; the radar uses ~250/day; the monthly 10,000 recharge was burned in one afternoon once by 8 parallel agents).
• Apify: monthly cap is small — estimate first; reconfirm any amount the founder dictates (voice dictation is ambiguous).
• Meta direct: ≤ 60–70 requests/hour total for the VPS, one scraper at a time.
• Transcription via API only; local whisper batches are banned (they saturated the VPS for days).
• Media > 2 MB to B2; JSON/text stays local. Look for archived files with mr-archive find before re-downloading.
• Reuse finished studies (competitor-case-studies.md) before paying for new data.