Winning Product Radar

Can an n8n automation mine the Meta, TikTok and Google ad libraries for products selling at $60 or more, categorise them, name the marketing angle, and rank them by real signals? Yes, with three corrections to how most people imagine it. This page is the scrape of Arie Scherson's channel, the verdict, and the build blueprint.

• Verdict
• The channel
• What each library really gives you
• Architecture
• The $60 gate
• Scoring model
• Category and angle
• Storage and the digest
• n8n workflows
• Cost and build order
• Experiment: niche sweep
• AdSpy, reverse-engineered
• Limits, legal, unverified

The channel: what Arie Scherson actually does

Arie Scherson (@aries1) describes himself as "breaking down brands that scale on Meta". The full video list was pulled with yt-dlp (the channel page is a JavaScript shell to a plain fetch). The CSV of all 1,228 videos was delivered alongside this page.

Cadence is falling: 22 uploads in January 2026, 3 in September. The two biggest recent videos are both brand breakdowns (62K and 20K views, September 2026). The breakdown is the product; the automation's output should read like one.

His stack, from 18 transcripts and the video descriptions

His numbers, verbatim where possible

• Price floor from cost. "If our cost of goods is 15, unfortunately we can't really charge anything below like $50." $27.88 landed means "$89 minimum, $89 or even $99". In practice $49.99 to $99 for $5 to $28 landed cost, $129 to $249 for $39 to $64.
• Break-even ROAS at or below 1.5. "You want to be below 1.5. You don't want to be above 1.5. It's going to be really hard to make that work on Meta." Target 1.3 to 1.5.
• Active-ad count is the trigger. 150 ads: "not bad". 420: "something I don't often see". 930, 1,200, 2,400: "you better stop scrolling".
• Traffic to revenue. One session is worth $4 to $5 in sales, or "$2 per visitor" on another day; stores "converting at 5 to 6%" on his theme, "over 4%" on a good brand.
• Test budget. $50 to $200 per product, 6 to 10 statics per launch, a $500 stop-loss.
• Qualitative gates. Solves a painful problem, a "new mechanism people haven't seen", high quality (low chargebacks, repeat buyers), demand visible in search or ads, high perceived value, small and light, bundle-able with a free gift, niches with lifetime value (jewelry, fashion, beauty, pets).

His breakdown template (the output format to copy)

The same sequence appears in every 2025-26 breakdown: (1) active-ad count in the Ad Library or GetHookd; (2) traffic chart, monthly visitors, country split, revenue estimate; (3) oldest and longest-running ads, statics first, open ad details, read headline, bullets, offer, guarantee; (4) click Shop now to the landing page or advertorial; (5) above-the-fold images and buy box; (6) offer construction: bundles, free gift, timer, BOGO, guarantee; (7) social proof counts; (8) cart and upsells; (9) best sellers; (10) comment report: angles, pain points, avatar; (11) what to source and at what price. His creative vocabulary: statics, UGC, VSL, carousel, catalog, partnership, theme-page, advertorial, quiz funnel, retargeting testimonial. His angle vocabulary: problem-aware callouts ("yeasty ears, itchy paws"), "selling the mechanism, not the object", gift angle, counterfeit warning, competitor price callout.

What each library really gives you

Every row below was either probed live with curl on 21-22 October 2026 or taken from the platform's own documentation fetched the same day. Verification status marks what an independent agent reproduced.

Architecture

Everything runs as small sub-workflows (one advertiser or one batch of 50 ads per execution) so n8n Cloud's 640 MiB on Starter never holds a whole dataset. Video bytes never pass through n8n when the Apify transcript actors are used; if the Gemini node is used instead, it is one video per execution.

The $60 gate: ad to price

1. Landing URL. Meta actors return link_url; TikTok detail enrichment returns landing_page. Facebook's l.php shim returns 400 without its hash, so decode the u parameter locally. Strip utm_*, fbclid, ttclid, gclid. Follow 301s (Shopify Markets redirect, for example gymshark.com to a regional checkout host). Linktree pages carry their links in __NEXT_DATA__.
2. Store detection. The old x-shopid header is gone. Use powered-by: Shopify, the link: </.well-known/ucp> header, Shopify.shop and Shopify.currency globals in the HTML, or cdn.shopify.com/s/files/1/ paths (headless stores still carry these).
3. Price. If the landing path is /products/<handle>, fetch /products/<handle>.js: price (cents, the default variant), price_min, price_max, compare_at_price, selling_plan_groups. If the landing page is a collection, advertorial or quiz, fetch /products.json (250 a page) and match the ad's headline or image to a product with an LLM, falling back to the catalogue's best-seller collection. Currency comes from /meta.json (products.json has none); convert with a daily FX table in Postgres.
4. Non-Shopify. JSON-LD Product.offers.price and og:price:amount (present on some stores, absent on custom themes); WooCommerce's public /wp-json/wc/store/v1/products (prices in minor units with currency); otherwise the fetched HTML (via a scraping API such as Zyte or Firecrawl when the store blocks plain fetches) through an extraction prompt in the Anthropic node. Jina's free reader is licensed for non-commercial use only, so it is not in the chain. If nothing yields a price, the product is held with price_status = unknown and surfaces in a weekly "needs a look" list rather than being ranked.
5. The rule. The advertised offer must be at or above $60 in USD: take the larger of the default-variant price and the price of the bundle or variant the ad's landing page sells (price_max of the advertised product). A $59 single sold through a $107 duo offer passes; a $59.99 single with no bundle is gated out. Store price_min, price_max and the advertised price separately. Expected coverage on day one: roughly two thirds of Shopify landers automatically, the rest via fallbacks.

Scoring model

Weights sum to 100. Each signal is scored on its own scale and normalised (points = weight × score / max). Signals with no data (a product not on TikTok Shop, say) are nulled and the total is rescaled to the weight that was available, so a Meta-only product is not punished for TikTok silence. Then the multiplicative penalties apply. All time windows are computed from the daily snapshot table, never from a single read.

On "engagement". Public like, comment and share counts exist only on TikTok's Top Ads. No Meta ad library exposes engagement, so for Meta the model uses what the library does show (impression band, collation count, longevity, ad volume) as the engagement proxy. If the optional WinningHunter API is subscribed, its per-ad Meta like, comment and share counts slot into signal 6 alongside the TikTok metrics; without it, Meta engagement is out of scope and the report says so.

Penalties and exclusions

• Saturation (multiplicative). Distinct advertiser pages running the same canonical product in the last 30 days: ≤ 5 → ×1.00 · 6-14 → ×0.90 · 15-40 → ×0.75 · 41-80 → ×0.55 · 81+ → ×0.40. A further ×0.80 when the same supplier photos appear on 5+ stores. Same product is decided by image perceptual hash first, then title embedding.
• Brand size (multiplicative). Page likes ≥ 1M, or catalogue ≥ 1,000 products, or ≥ 40% of the page's ads are catalog/DPA templates, or the ads sell no product and no offer → ×0.50 and tag "brand, not product". 100K-1M likes with a catalogue over 300 → ×0.80. This is what keeps Gymshark, Temu and SHEIN off the list.
• Retargeting and zombie ads. Any ad active 30+ days with an impression band under 10K or the "Low impression count" badge is excluded from signals 1-3 and tagged; it is still mined for objection-handling copy.
• Exclusions. Restricted categories (cure claims, weapons, adult, tobacco and vape, gambling, counterfeits) are dropped; seasonal products (annual Trends spikes) are flagged and their longevity capped at 45 days.

Ranking. Tiers: 70+ launch-grade, 55-69 watch, under 55 skip. The digest shows new entrants to launch-grade, products whose score rose 8+ points in 7 days, and the top five of the watch tier. Ties break on momentum, then on price headroom.

Worked example

Category and marketing angle

The classification spec was delivered as ad_classifier_v1.json: system prompt, user-message template, the JSON schema (validated against Claude's structured-output rules: no min/max, additionalProperties:false everywhere, zero unions), and the taxonomies. The summary:

Pipeline per ad

1. Video to text. Claude takes no video. Either the Apify transcript actors (templates 19027 for Meta, 19153 for TikTok) or the n8n Gemini node's Analyze Video (~100 tokens per second of video, about $0.003 to $0.006 per 30-second ad) produce transcript, on-screen text and a visual description in one call. No separate OCR step.
2. Classify. Anthropic Chat Model node (Haiku 4.5 at about $0.006 per ad, Sonnet 5 at about $0.012 when a dispute needs a tie-break) with the Structured Output Parser and the schema from the spec. Input: ad copy fields, transcript, visual description, thumbnail, and the context line SOURCE / PAGE / FIRST_SEEN / ACTIVE_DAYS / ADS_FROM_THIS_PAGE. Output fields: evidence, is_physical_product_ad, product_name, product_description, category_l1, category_l2, price_seen, primary_angle, secondary_angle, hook_type, format, offer_mechanics, target_customer, claims_made, emotional_driver, confidence, needs_review (all required). Non-product ads (apps, services, lead-gen, brand awareness) come back with is_physical_product_ad = false, category_l1 = other and needs_review = true; needs_review is what feeds the weekly "needs a look" list. Set the model explicitly: the Anthropic node defaults to claude-sonnet-5. Put the few-shot examples inside the cached system block so it clears the minimum cacheable length, and confirm caching from cache_read_input_tokens in the node output.
3. Aggregate to the product. The product's angle is the dominant primary_angle weighted by each ad's collation_count and active days; secondary angles are listed with their share. This is what the digest prints.
4. Evaluate before trusting. Hand-label 100 ads (stratified by source and vertical, two labellers, 30 hostile cases), freeze model + prompt + schema, report per-field agreement. Below about 90% on clear-cut cases, the field is not ready for ranking.

-- classification prompt, condensed (full text in ad_classifier_v1.json)
SYSTEM: You are a performance-creative analyst for DTC e-commerce brands. You classify ONE paid-social ad
at a time from the evidence supplied (ad copy fields, transcript, on-screen text, keyframe descriptions,
thumbnail). Pick category_l1 and category_l2 ONLY from the enums. primary_angle is the argument the whole
ad makes; hook_type is the first 3 seconds only. Non-product ads (apps, services, lead-gen, brand awareness
without a product): is_physical_product_ad=false, category_l1='other', needs_review=true. Never infer a price
that is not shown; price_seen is only what the ad or its text states.
USER: AD_ID {{ad_id}} SOURCE {{meta|tiktok|google}} PAGE {{page_name}} FIRST_SEEN {{start}} ACTIVE_DAYS {{days}}
ADS_FROM_THIS_PAGE {{n}} AD_COPY primary_text=… headline=… cta=… TRANSCRIPT … ON_SCREEN_TEXT … VISUAL …

Storage and the digest

create table advertisers (id bigserial primary key, source text, page_id text, page_name text,
  page_likes int, domain text, country_set text[], first_seen date, last_seen date, unique(source,page_id));
create table ads (id bigserial primary key, source text, ad_key text, advertiser_id bigint references advertisers,
  landing_url text, resolved_url text, body text, headline text, cta text, media_url text, format text,
  started_at date, first_seen date, last_seen date, is_catalog bool default false, unique(source,ad_key));
create table ad_snapshots (ad_id bigint references ads, day date, is_active bool, collation_count int,
  active_hours int, impression_band text, eu_reach int, likes int, comments int, shares int, ctr numeric,
  primary key(ad_id,day));
create table stores (id bigserial primary key, domain text unique, platform text, myshopify_domain text,
  currency text, product_count int, shop_id bigint, review_count int, review_source text, checked_at timestamptz);
create table products (id bigserial primary key, store_id bigint references stores, handle text, title text,
  canonical_key text, image_phash text, price_usd numeric, price_min_usd numeric, price_max_usd numeric,
  has_subscription bool, has_bundle bool, published_at date, product_type text, tags text[],
  price_status text check (price_status in ('ok','below_gate','unknown')), unique(store_id,handle));
create table ad_products (ad_id bigint references ads, product_id bigint references products, primary key(ad_id,product_id));
create table ad_classifications (ad_id bigint primary key references ads, is_physical_product_ad bool,
  product_name text, product_description text, evidence text, category_l1 text, category_l2 text,
  primary_angle text, secondary_angle text, hook_type text, format text, offer_mechanics text[],
  target_customer text, claims text[], emotional_driver text, confidence numeric, needs_review bool,
  model text, classified_at timestamptz);
-- canonical products: products.canonical_key groups the same item sold by many stores (see W4b);
-- store title embeddings in a pgvector column (title_embedding vector(1536)) on the products table.
create table enrichment (product_id bigint references products, day date, google_ads_low int, google_ads_high int,
  google_has_video bool, tiktok_top_ads bool, tiktok_ctr numeric, tiktok_engagement int, fastmoss_gmv_7d numeric,
  fastmoss_creators int, trends_slope numeric, primary key(product_id,day));
create table product_scores (product_id bigint references products, day date, s1 int, s2 int, s3 int, s4 int,
  s5 int, s6 int, s7 int, s8 int, s9 int, s10 int, s11 int, s12 int, available_weight int, raw numeric,
  saturation_mult numeric, brand_mult numeric, score numeric, tier text, distinct_advertisers int, flags text[],
  primary key(product_id,day));
create table dead_letter (id bigserial primary key, workflow text, item jsonb, error text, at timestamptz default now());

Idempotency: every write is an INSERT … ON CONFLICT DO UPDATE through the Postgres node ("Insert or Update"); the Supabase node has no upsert, so point the Postgres node at the Supabase connection string. Keys are (source, ad_key) for ads and (store, handle) for products.

What lands in Slack or your inbox

n8n workflows

Hosting. n8n Cloud Starter (€20, 2,500 executions, 5 concurrent, 640 MiB) runs W1-W8 only if sub-workflow executions stay under the plan's execution count; at 1,200 TikTok rows plus a few thousand Meta ads a day the per-item sub-workflows exceed 2,500 executions in days. Realistic options: n8n Cloud Pro (€50, 10,000 executions, 20 concurrent) with W4 and W5 batching 25 items per execution, or self-hosted Docker on a $12-24 VPS with Postgres and Redis in queue mode, which also removes the memory ceiling and lets the Code node use packages. The Cloud execution-time cap is contested (pricing page says 5 min on Starter; n8n staff said in July 2026 that no hard cap is enforced), so confirm with support before relying on long waits.

Cost and build order

First week

1. Day 1, the spine. Supabase schema above; n8n with the Apify node installed; one actor run with a single Ad Library search URL (q=heated brush, country US, active status) and count 300; W4 on the result. Done when 20+ products sit in products with a USD price and a price_status. This also answers the open question about impression bands: look at the actor's raw fields.
2. Day 2, time. Daily schedule, the snapshot table, page tracking for every advertiser found; signals 1-5 computable after 48 hours of snapshots.
3. Day 3, meaning. W5 on ad copy and thumbnail only (no video yet). Category and angle in the table; spot-check 30 by hand.
4. Day 4, the other two libraries. W2 for TikTok Top Ads joined on landing domain; W3 SearchSuggestions for every advertiser.
5. Day 5, the score. W7 SQL with penalties; the first ranked list. Read the top 20 and the bottom 20.
6. Day 6, video and delivery. Transcript actor or Gemini node; W8 digest into Slack in the card format above.
7. Day 7, the hostile suite. Gymshark and Temu must score as "brand, not product". A 60-day ad with the low-impression badge must be excluded. A $59.99 single with no advertised bundle must be gated; a $59 single sold through a $107 duo must pass. Twelve dropshippers of one ring must collapse to one product with a ×0.75 saturation. A catalog ad must not count as a hook. If any of these passes without a fight, the check is wrong.

Experiment: the garden, handy and household sweep

Run on 25 October 2026 at the owner's request: 45 product keywords (lawn mower, pergola, tool kit, stud finder, mosquito trap, cookware set and 39 more), 250 ads each, and a "sniper" rule that prices the creatives Meta shows as 10+ ads use this creative and text first. Discovery took about 90 minutes of actor time and roughly $7.60 of Apify; the price gate then ran locally. Everything below comes from the stored data, and the classification of the 38 heavily tested concepts was done by six classifier agents and one synthesis agent using the taxonomy spec.

What the sniper rule caught, by advertiser type

24 of the 38 concepts belong to the requested niche; the other 14 leaked in because Meta's keyword search matches body text (a drama app under "sprinkler", a denture cleaner under "raised garden bed").

1. Headline finding

The 10+ versions rule across 45 keywords and 10,174 ads surfaced 38 unique concepts, but 12 of them (216 versions) are one mosquito-zapper advertorial run from ten zero-like personal-name pages into achelesslife.com, and another 14 are not in the niche at all. Only four of the 24 in-niche concepts resolved a price, and just one cleared the $60 gate: the Solar Eagle Garden Light at $70.92 (vvtashen.com, 14 versions, 91 days). The heavy-testing signal is real but it mostly finds advertorial rings and launch bursts; the $60+ product list came from the separate price sweep (ezarctools.com saw $74.97, ironridgetools.com lawn sweeper $349.99), not from the sniper.

2. What the sniper rule actually surfaces

3. The contenders

1. Solar Eagle Garden Light (vvtashen.com, 14 versions, 91 days, 45 likes, $70.92 ok). Identity/pride angle, UGC hook "This eagle is absolutely stunning in person!", no offer. Risk: a 45-like store carrying a patriotic novelty that may be a Q4 gift spike.
2. Solar dummy security-camera light (enjoyaitlife.com, 18 versions, 55 days, 12,470 likes, price unknown). Problem-agitate "Worried about trespassers roaming around your yard after dark?" plus "2,649 sold in last 24 hours". Risk: price unresolved; slug -3-2-2 says duplicate imports, likely sub-gate.
3. Glow wireless doorbell BOGO (vionb.com, 10 versions, 48 days, 2,253 likes, price unknown). Sharpest concept in the set: us-vs-them "Glow wants an outlet", BOGO plus 60-day guarantee, renter-targeted lander. Risk: commodity doorbell, margin lives entirely in the BOGO.
4. Laser magnetic stud finder (warmsoftnook.com 25 versions in 2 days, 378 likes; vibenestgoods.com 10 versions, 19 days, 682 likes; identical "My original invention" copy, both unpriced). Two sellers bulk-launching the same item is a sourcing signal. Risk: launch test, nothing proven; sub-$30 category.
5. Toolsons MultiTool Set ($36.96, 31 and 27 versions, 147 and 165 days, 4,973 likes). Longest-lived legitimate store in the niche; demo-driven feature copy "One Kit. Infinite Builds." Risk: below gate, pure commodity.
6. Japanese-teeth pruning saw (ezarctools.com, $74.97, 8 ads, 49 total versions, 111 days, 12,839 likes). From the price sweep, max 8 versions so the sniper missed it. Risk: unclassified, no copy analysed yet.
7. Heavy-Duty Push Lawn Sweeper (ironridgetools.com, $349.99, 2 advertisers, 21 versions, 55 days). Two sellers at a high ticket. Risk: freight and returns at $350; seasonal.

4. The traps

• The achelesslife.com ring: Zuleira Carrington, Julie Oliver, Andrea Ortega, Brittany Williamson, Ashley McCoy, Kayla Bishop, Jessica Howell, Rebecca Fuller, Michelle Soto, Lauren Meyer; all 0 likes, "Learn more", landers /zapper/ and /bugs-killer/. A sister ring (Callie Kelly 14 likes, Rebecca Sanders 9 likes, mynovapod.com) sells denture cleaner and leaked via "back garden". Rule: collapse concepts by landing domain, not page; a 0-like personal-name page plus a non-store domain is one advertorial, scored once and flagged policy_risk medium/high.
• Big brands: Harbor Freight, OdoBan, HomeBuys, Work Sharp ({{product.name}}). Rule: catalog/dynamic exclusion already removed 2,593 ads; add page_likes above ~100k or empty landing_url as brand media.
• Leaked non-niche (14 of 38): matches on body words ("workbench" for Rivanta, "sprinkler" for HighLevel, "vacuum-insulation" for GLEA, "Electric Adjustable Table" for HomeBuys). Rule: in_requested_niche=false drops them; require the keyword in headline, product title or slug, not body.
• Launch bursts: Filamniceent.NI03 at 25 versions in 2 days. Rule: versions per day above ~5 with days under 7 is "watch", not "proven".

5. Angle and hook patterns

Among in-niche heavy testers, trend_native_viral dominates by count (12 mosquito creatives, all bold_claim hooks) and feature_benefit_rational by store count (Toolsons, stud finders, Enjoyaitlife, Arrowmax, Harbor Freight). Long-copy confession_story_open appears only outside the niche (denture, Avera, Rivanta). Offer mechanics are almost absent: limited_stock in the ring, one BOGO (Glow), one deposit lock (ARANIX). Verbatim hooks:

• "THIS DEVICE IS HARD TO FIND RIGHT NOW" (7 pages, scarcity arm) vs "AMERICANS ARE RUSHING TO BUY THIS DEVICE" (bandwagon arm, same body)
• "Every doorbell sold today wants three things from you: your Wi-Fi password, an account, and a monthly fee. Glow wants an outlet."
• "🔥2,649 sold in last 24 hours🔥"

6. Keyword lessons

Heavy: mosquito trap (26 sniped, max 35, one ring), electric screwdriver (23, but yielded Arrowmax and HomeBuys), wet dry vacuum (22, both leaks), stud finder (10, max 25), tool kit (10, max 31), solar garden lights (13, two real stores). Greenhouse and chainsaw sniped 13 each yet survived into the 38 as nothing or Cheeky Soap. Twenty-three seeds produced zero: dehumidifier, closet organizer, cookware set, smart lock, weeding tool, paint sprayer, steam cleaner, pizza oven, deck tiles, laser measure, robot lawn mower, bird feeder, patio furniture, cordless drill, garden shed, pressure washer, pergola, garden hose, hedge trimmer, self watering planter, telescoping ladder, pool cleaner robot, window cleaning robot. Broad category words find brands; product nouns find testers.

Next ten:

1. mosquito killer lamp — the ring never names the product; search the product, find the stores behind it.
2. solar security camera light — Enjoyaitlife's winner, caught only by accident under "solar garden lights".
3. wireless doorbell — Glow's category surfaced under "socket set".
4. solar garden statue — Vvtashen's $70.92 eagle suggests a patriotic-decor band above the gate.
5. lawn sweeper — two advertisers at $349.99 already.
6. fruit picker — pinkdiyprint.com $67.86, 8 ads in 13 days.
7. pruning saw — ezarctools $74.97, 111 days.
8. mini chainsaw — "chainsaw" had 102 advertisers but the broad word surfaced soap.
9. spin scrubber — spin mop hit max 17 versions; the cleaning-gadget cousin was never seeded.
10. ultrasonic pest repeller — pest is the strongest emotion in the data and mosquito season ends in October; indoor pest is the next quarter's version.

7. What the pipeline could not see

214 landing pages had no readable price; 16 of 24 in-niche concepts are unpriced, including the entire ring, both stud finders, Enjoyaitlife and Glow, so the $60 gate judged 4 concepts. "Learn more" advertorials hide the product name and the store. No video transcripts exist, so the mosquito ring's 216 versions are indistinguishable beyond two headline swaps. The single most valuable build: follow each advertorial's outbound order link to the checkout domain and run the products.js price resolver there, which unlocks price and product identity for 14 of the 38 concepts in one step.

Products at $60 or more found in the sweep, ranked by total creative versions

71 products cleared the gate, 43 fell below it, and 327 landing pages carried no readable price (advertorials, quiz funnels, non-Shopify stores). The top rows:

Yield per keyword

Broad category words find brands; product nouns find testers. Sorted by heavily tested creatives found.

Downloads: the classified creatives and the sweep shortlist are attached in the chat where this experiment was run; the queries that produce these tables live in ~/n8n/radar/experiment.sql.

AdSpy, reverse-engineered

Mapped on 25 October 2026 inside the trial account, with three ad views spent out of 940. AdSpy is an Angular app over a private REST API at api.adspy.com/api/ (endpoints token, user, ad, search, tech, Country, AffNetwork, label, logx) plus a SignalR hub for live updates. The search is one GET: /api/ad?searches[0][type]=texts&searches[0][value]=heated brush&page=N with the filter keys below added as query parameters. There is no public API, no export button, no CSV, no alerts; the only CSV code in the bundle is the PrimeNG table library's and nothing calls it. The trial truncates every advertiser name ("Shava…(subscribe to see name)") and every ad text, and meters ad views (user.viewsLeft, 940 left); a 417 from the API sends you to the subscription page. The paid plan is $149 a month, Facebook and Instagram only.

What it searches and filters

What one ad exposes

Card: advertiser, created on, last seen on, ad text, headline, display domain, CTA, total likes with the reaction breakdown (like, love, haha, wow, sad, angry), comments, shares, gender icon, country, media. Detail page: demographics (gender, countries, age groups), a likes over time chart, landing-page screenshot, the final URL with every UTM and campaign ID intact, a BuiltWith-style technology table (analytics, payment, CMS, email, hosting), "View Best Comments", and labels. The ad object in the bundle carries createdOn, countries, genders, minAge, maxAge, adType, offers, mainAttachment, subAttachments, comments, linkToAd.

What to take from it, and what not to

• Its unique data is engagement. Reaction counts, comments, shares, daily-likes velocity and the likes-over-time curve per Meta ad exist nowhere in the public libraries. Everything else (dates, countries, CTA, landing URL, run time) the pipeline already gets from the Ad Library actors, and AdSpy has no TikTok, no Google, no prices.
• Do not automate against it. The API is private and view-metered, the trial truncates the two fields that matter, and a scripted session is the fastest way to lose the account. The pipeline is built on sources you own; AdSpy is a lens on top.
• Use it two ways. (1) As the weekly human check on the shortlist: open each launch-grade product's advertiser, read daily likes, the reaction mix, the best comments, and the likes curve, and enter three numbers into the sheet (daily likes, total engagement, comment sentiment). (2) As a discovery recipe that replicates the pipeline's filters by hand when you want a sanity check: Technology = Shopify, Button = Shop Now, Created Between = last 60 days, Seen Between = last 7 days, Daily likes ≥ 10, Media = Video, sort by Longest running, then by Likes. If the same numbers are wanted automatically, WinningHunter's API ($79 a month) returns per-ad like, comment and share counts and plugs into signal 6.
• Two filters worth copying into the pipeline. A commerce-intent filter (store technology detected and a buy-type CTA) applied before classification cuts non-product ads early. A comment-intent signal (share of comments with purchase intent versus complaints, from the public comments on the ad's post) is a cheap, honest engagement-quality score; it joins signal 6 as a tie-breaker.

Limits, legal, unverified

What the pipeline cannot do

• It cannot see Meta spend or impressions for US-only ads beyond the UI's bands and badges; EU reach is the only numeric spend proxy, and only for ads that delivered in the EU.
• It cannot see Meta engagement (likes, comments, shares) from any ad library; WinningHunter's API is the only tested source, and ShopiLab/SmartReply-style comment analysis needs a Facebook page's comments, which this design does not scrape.
• It cannot read TikTok's Top Products or Keyword Insights tabs (login-gated, no actor found), nor US ads in TikTok's Commercial Content Library (EU-only by design).
• It cannot discover brands through Google; Google only confirms brands found elsewhere.
• Review velocity (reviews per week) for third-party stores has no known source; review counts come from Yotpo and Okendo only when the store uses them.
• Traffic estimates like Arie's monthly-visitor charts need Similarweb's paid API (business tier) or a tool like Pluto Hunt; the free DigitalRank endpoint gives a global rank only.

Legal and operational posture

Still unverified (and what resolves each)

• Impression bands on every US ad in the Ad Library UI. Arie's 2026 recordings sort by impressions; a May 2026 vendor guide says commercial impressions are not exposed. Resolved by the day-1 actor run.
• Shopify's Web Bot Auth 429 for datacenter crawlers (reported since late July 2026). Resolved by running W4 from n8n Cloud once; if 429s appear, add the proxy credential.
• n8n Cloud execution-time cap (pricing page vs staff statements). Resolved by one email to n8n support.
• Google SearchCreatives request format (returned an empty object with a text query; the domain form was not tested within the probe cap). Irrelevant if SerpApi is used.
• Kalodata Open API tier (self-serve vs enterprise-only at $570+/month). FastMoss is the documented alternative, so this only matters if Kalodata's data is preferred.
• The Meta official API's ~200 calls/hour limit has no primary source; it only matters for the optional EU-reach supplement.
• Independent re-verification covered 10 of the 24 claims planned (15 agent verdicts; five claims were checked by both a refute lens and a reproduce lens); the rest were cut by session limits. None was refuted; four were refined (the EU-only API scope, the Google unverified-advertiser rule, the TikTok gate wording, the Shopify taxonomy's move to release assets). The scouts' own live probes cover the TikTok, Shopify, Google and Meta-API claims above.