Metrics, 9.9 Sprint & Roadmap

North-Star & Funnel Definition

The north star: First Qualified Sourcing Action (FQSA)

Signup is not the goal. A seller who creates an account and never finds a supplier is a cost we paid for nothing, and a signup count that hides that fact. The north-star metric for this plan is:

FQSA — First Qualified Sourcing Action: one unique user completes either (a) supplier_contact_requested on at least 1 supplier or (b) results_saved with ≥ 3 distinct supplier_id values in a single saved list, where the originating search has category_l1 in a 9.9-relevant taxonomy and query_type in (text, image).

Counted once per user (the first qualifying event), reported as a daily count plus a cumulative count against the Sep 14 target. Every user who has an FQSA is an Activated Seller — the only cohort we are allowed to describe as "value delivered" in any report, case study, or investor/retro note.

Why this and not signups, searches, or "matches shown":

• Signups are a vanity metric here because the friction to sign up is near zero (free window, social login, no card). Signup rises with any clickbait hook — a giveaway, a misleading "factory price" thumbnail, a discount-code post — and correlates poorly with whether a seller actually found stock. Worse, signup can be involuntarily inflated by paid campaigns optimised to Lead on Instant Forms where the user never reaches a search result. If we manage to signups we will buy a pipeline of dead accounts.
• Searches are gameable and noisy — a bot, a curious scraper, or a person testing one junk query registers the same as a buyer. Searches measure our marketing, not our product.
• "Suppliers matched" measures our index, not the user's outcome. It is an input metric and it is our job for it to be high.
• FQSA is the smallest signal that a real procurement intent was expressed (saving a shortlist to work from, or reaching out to a factory). It is also the event that best predicts whether a seller will ever pay: whoever pays for credits is paying to contact more factories, not to see more search boxes.

Secondary metrics we report alongside FQSA but never optimise to: Activated Seller → quote_received rate, return_day7, credits_consumed per activated user, and search-failure rate. Time-to-FQSA (median minutes from signup_completed to FQSA) is the health check on onboarding — target under 10 minutes for desktop-flow users, under 6 minutes for WhatsApp-flow users [ASSUMPTION], both [VERIFY] from real data in week one.

Funnel stages and exact event names

Canonical names below are the single source of truth. Migration note: paid-media.md currently refers to a FirstSupplierSearch event — that alias must be renamed to first_search (or mapped one-to-one in GA4 with an alias report) before Sep 8, or the dashboard and the ad platform will disagree and retro in W4 will be fiction.

session → signup_completed → first_search → results_saved → supplier_contact_requested → return_day7

Stage-to-stage conversion targets

Every figure here is a placeholder hypothesis written before we have data. They are the numbers we intend to falsify, not to defend.

Chained-multiplication is the honest way to read this: even with every stage hitting target, roughly 1 in 400 cold sessions produces an FQSA. With USD 60/day and a blended cost per signup of USD 2.50–4.00 [ASSUMPTION] (see paid-media.md), paid buys about 15–24 signups per day, i.e. 3–5 paid FQSA/day at plan-midpoint funnel rates, and plausibly zero for several days running. That is not a reason to skip paid — it is the reason the Sep 8–14 dashboard must be FQSA-first, and the reason no one is allowed to call paid "working" on the strength of signup counts alone.

UTM & Event Taxonomy (must exist before Sep 8)

Hard gate: if the UTM spec, the 10 events, and the attribution dashboard are not live on Fri Sep 5, paid launch slips from Sep 8 to Sep 10. Spending against an untagged funnel is unrecoverable data loss; a two-day delay is not.

UTM convention

Format, all lowercase, underscores only, no spaces, no trailing slashes, no uppercase anywhere:

utm_source    = where the click came FROM (the platform)      e.g. meta, google, tiktok
utm_medium    = HOW it came (the delivery mechanism)          e.g. paid_social, organic_social, email
utm_campaign  = our internal flight/plan name                 e.g. w3_paid_meta_99credits_sea-en
utm_content   = the creative or message variant               e.g. carousel_pricemoq_v2, ama_reminder_h18
utm_term      = PAID SEARCH ONLY: the keyword, or `not_provided` when the platform withholds it

Rules:

1. Never set UTM on internal links. Login → search, in-app share, and own-domain email links must be UTM-free or self-attributing, otherwise a user's own second visit looks like a new "referral" and we invent a channel that does not exist.
2. One utm_campaign per budget line. If two campaigns share a campaign name we cannot separate them retroactively — that is the single most expensive tagging error.
3. Link builder is mandatory: all links are generated from the sheet growth-kit/links.csv, never typed. If it is not in the sheet, it does not get posted.
4. WhatsApp/Telegram/Zalo click-to-chat links carry utm_medium=ctc and the deep link must survive the app hop (deferred deep link or a short-link service we control with 90-day attribution retention) [VERIFY] — if the link does not survive, community channels attribute to Direct and we will under-count our best channel.
5. Direct/untagged traffic is a real bucket we must keep under 30% of sessions [ASSUMPTION]; anything above that means the tagging discipline has already failed and we should stop scaling rather than keep guessing.

Allowed utm_source values (fixed list; adding a value is a one-line PR, not a freelance decision): meta, ig, fb, tiktok, google, yt, fbgroup, telegram, whatsapp, zalo, reddit, quora, shopeeforum, seo, email, referral, offline, partner, koc, direct

Allowed utm_medium values: paid_social, paid_search, organic_social, community, ctc, forum, organic, email, lifecycle_email, push, referral, qr, booth, partner, spark_ads

The 12 pre-named campaigns (no new names without a swap-out)

The 10 events (GA4 + custom, with parameters)

return_day7 is computed nightly from authenticated sessions and cohort date; it is deliberately not a client event, because retention computed from logs cannot be silently broken by a dropped SDK call.

The 6 audiences to build (by Sep 5)

1. au_signed_no_search_48h — signed up, zero qualifying search in 48h. Feeds onboarding email/WhatsApp nudge and tells us if the hook oversold.
2. au_searched_no_save — searched, never saved. Feeds results-quality follow-up; the primary suspect list for supplier-data complaints.
3. au_saved_no_contact — shortlisted but timid. Feeds negotiation-template + trust content; highest-intent non-FQSA pool.
4. au_fqsa_99 — Activated Sellers during the 9.9 window. The only legal seed for w3_paid_meta_lookalike_id_th, the case-study pool, and the 11.11 waitlist.
5. au_free_window_expiring_d3 — days remaining ≤ 3 in the free window. Feeds w3_lifecycle_freeexpiring_d3.
6. au_churned_after_first_value — had an FQSA, no authenticated session in 10 days. Win-back list, and the honest test of whether 9.9 value stuck.

Lookalike audiences must not be pushed at scale until the seed has ≥ 200 unique members [ASSUMPTION] — below that, LAL quality is noise and we will blame the creative.

Naming lint rule

Automated check runs on links.csv before any post or ad goes live; it must reject:

• any uppercase letter, any space or hyphen-in-source, any non-[a-z0-9_-] character;
• any utm_campaign not present in the 12-name list, or any name containing test, tmp, draft, copy, final, v2final, asdf, or a personal name;
• any duplicate campaign name used across two channels or two objectives;
• any missing utm_content on a paid link (we cannot learn which creative worked);
• any link posted with a bare ? and no parameters at all.

Test campaigns are the classic leak. They enter GA4 as permanent sessions, distort channel averages, and in W4 someone will quote one in a retro. Use the w3_*_test_* namespace nowhere; run tag validation in GA4 DebugView plus a staging host with the analytics tag disabled.

Channel Attribution Model

Default reporting model: first-touch on campaign for cost efficiency, last-non-direct-touch for credit, plus a simple assisted-touch view once a week. We do not have enough volume for algorithmic data-driven attribution to mean anything — at 15–24 signups/day, any model more complex than "who started it and who closed it" produces confident nonsense.

Attribution reality check

The attribution_survey_submitted field ("How did you hear about us?") is worth collecting at the post-first-save moment rather than post-signup: at signup, a seller is primed to name the brand they admire; after they've gotten value, they remember more accurately what actually got them there. It buys us three things no pixel can: awareness of dark-social (WhatsApp forwards, TikTok-to-chat shares, word-of-mouth at a wholesale market), a sanity check on the Direct bucket, and a signal for which offline effort is real when links are unclickable by design.

Its limits are equally real and must be stated in every report that uses it. Self-reported source is inflated for prestige channels (people say "I found you on Google" when they typed our name after a group post), deflated for paid (many users genuinely do not know an ad was an ad), biased toward the most recent memorable thing, and only answered by 15–25% of users [ASSUMPTION], who are not a random sample. So survey data may name a channel we are missing and may re-rank our hypotheses, but it must never be used to reallocate budget on its own. Rule of practice: if survey volume for a channel exceeds 10% of responses while tracked FQSA for that channel is zero, treat it as a tagging bug to investigate, not as an undiscovered gold mine.

Daily & Weekly Review Rituals

Daily standup: 12 minutes, same 9 columns, filled by 09:30 SGT

Five alert thresholds that force an intervention (not a discussion)

1. search_no_results > 15% for 2 consecutive days → pause whichever campaign is driving the missing queries (most likely the TikTok image hook), same day. Add the top 10 failed queries to the index backlog and tell the sellers we asked for them.
2. Blended cost per signup > USD 6 for 3 consecutive days [ASSUMPTION] → cut the worst campaign to zero and reallocate to best CPL; if no campaign is under $4 by Sep 11, cut total spend to $30/day rather than fund a bad bet for a week.
3. Signup→first_search < 40% on any single day → all content shipping freezes; the 1.5 people spend that day on onboarding (empty-state copy, the first-query suggestion chips, image-search guidance). Nothing else matters while that number is broken.
4. Untracked sessions (no utm_*, no referrer) > 30% → stop scaling spend until tagging is fixed; publish no channel-level claim in the meantime.
5. Sep 8–9 only: oldest support item > 6 business hours → halt every scheduled post and AMA extra; everyone answers the inbox. During the 9.9 window, a seller whose supplier question goes unanswered for a day has already gone back to their old supplier.

Weekly review: 60 minutes, Fri 15:00 SGT

1. 0–5 — North-star line, last 3 weeks, against target. No commentary yet.
2. 5–15 — Funnel diff stage by stage vs last week. Question: where did the leak move, not did we grow.
3. 15–25 — Campaign P&L: spend ÷ FQSA per campaign. Decide kill / keep / scale for each of the 12 names, out loud, in writing.
4. 25–35 — Qualitative: five verbatim search_no_results queries and five au_searched_no_save users traced by hand. Numbers tell us what, this tells us why.
5. 35–45 — One channel experiment for next week: hypothesis, owner, success number, date. Exactly one, given capacity.
6. 45–55 — Content carry-over: which assets retire, which get localised (ID/TH first), which become the 11.11 spine.
7. 55–60 — Write the one-page memo and commit next week's numbers. The memo is the raw material for W4 retro, so retro is never a research project.

Three automated, three manual

Automate (before Sep 8, no exceptions): (1) daily funnel-by-campaign CSV into one sheet with FQSA as the top row; (2) paid spend vs tracked FQSA vs platform-reported conversions per campaign, with the delta shown — persistent delta means broken conversion API; (3) the search_no_results query log, deduped, in category order. Keep manual, deliberately: (4) verbatim failure reading (a human decides which queries are real demand vs junk); (5) community/group sentiment and admin-relationship health (no tool detects "we are about to be removed"); (6) KOC/partner reconciliation against invoices and disclosures.

9.9 Sprint: Hour-by-Hour Playbook (Sep 7–9)

All times SGT. Owner A = growth/content (1.0 FTE), Owner B = product/sourcing + support (0.5 FTE). Sep 7 = Mon, Sep 8 = Tue (ads live, AMA evening), Sep 9 = Wed (day-of).

Countdown guardrail

Never fake urgency. Every countdown must be computed from the real expiry timestamp (Sep 14, free-usage end), not from a marketing date, and must be removed the moment it hits zero. If we extend the free window, we announce the extension as a change with a new real date — we never roll a clock back without saying so. The reason is not ethics-as-decoration: this product's entire value proposition is trust in numbers we pull from a factory floor. A fake countdown is the fastest way to teach sellers that our numbers are marketing too.

What you may claim (the social-proof gate)

Write the Sep 9 14:00 post only from our own event log, under these rules:

Rule of thumb: below 500 tracked sellers with a qualifying search, we compete on specificity and honesty, never on scale. If the gate cannot be met, the 14:00 slot publishes the questions-and-gaps post instead.

Offer Design for 9.9 → 11.11 Bridge

Four variants; all are the same underlying product with different framing and different friction. Choose by audience segment, not by which one we like.

Variant 1 — Free-credit window (default, mass)

Fits: first-time Shopee/TikTok Shop sellers, price-sensitive, low trust budget. Mechanic: unlimited qualifying searches + 30 credits through Sep 14; saved lists persist regardless. Honest wording: "Free until 14 Sept. After that, sourcing searches cost credits, and we'll show you the price before you spend anything. Your saved supplier lists are yours either way." Why honest: persistence-of-work is the real reason to sign up; "free" alone is a race to be cheapest and we will lose it to anyone with more money. Risk: attracts window-shoppers who never intend to buy stock. Mitigate by measuring FQSA (not signups) as the quality check on the free offer.

Variant 2 — "We-keep-looking" promise (not a guarantee)

Fits: sellers with an urgent specific SKU; wholesale/offline sellers with a concrete order to fill this week. Mechanic: if a seller's search returns no supplier that fits their price/MOQ, a human re-sources it within 24 business hours and reports back, with what was found and what wasn't. Honest wording: "If our results don't fit your price or MOQ, tell us — we'll manually hunt for 24 hours and show you what we found. If there's nothing, we'll say there's nothing." Do not phrase this as a "first-stock guarantee", "satisfaction guaranteed", or any money-back claim — we cannot control factory stock, lead times, shipping, or customs, and any guarantee-flavoured wording creates a liability we cannot honour [VERIFY] with counsel before using the word "guarantee" in any market. Why do it anyway: the manual hunt is also our best product research and our cheapest index expansion, and it converts exactly the sellers who would otherwise churn at search_no_results. Capacity limit: cap at 15 manual hunts per week during the sprint, and say so. [ASSUMPTION] — validate against B's 0.5 FTE.

Variant 3 — 9.9 sourcing cohort / bootcamp bundle

Fits: sellers who want process, not tool — the majority of the FB-group cold-start audience; strongest in ID, VN, TH. Mechanic: 4-day cohort (Sep 8–11): day 1 category selection, day 2 image-search + shortlisting, day 3 MOQ/negotiation with supplier questions, day 4 landed-cost and shipping reality. Includes the Sep 4 asset pack, credit grant, and a shared ask-in-chat channel. Capped at 60 seats, honestly, because we can only answer 60 people's questions well. Honest wording: "Small cohort, 60 seats, 4 days before 9.9. We teach sourcing and we answer your questions. We don't promise sales — nobody can." Why it matters: cohorts produce the testimonials we're allowed to use, the case studies for W4, and the KOCs for October. Conversion to FQSA inside a cohort should be measured against the 0.24–1.1% cold baseline above [ASSUMPTION].

Variant 4 — Partner bundle

Fits: sellers already using a logistics agent, payment gateway, product-photography or listing-tool partner; SG/MY segments with higher willingness to pay. Mechanic: partner's audience gets a credit grant + our category pack; partner gets co-branded content and a disclosed referral arrangement. Honest wording: "This bundle comes with [partner] because they handle shipping and we handle sourcing. We're paid a referral fee if you take it — that's why it's discounted. Reviews and pricing are ours." Disclosure is mandatory in every market, and per-partner terms go in one sheet before any bundle is posted [VERIFY]. Why: cheapest FQSA available if the partner's audience is genuinely sellers; most damaging if it's consumers.

W4 bridge into 11.11 pre-heat (Sep 15–21)

Carries over: the FQSA definition and event taxonomy (never change a metric mid-flight — changing it in W4 to flatter retro numbers is prohibited); the campaign naming system with the w4_* and 1111 prefix; au_fqsa_99 as the seed for every paid audience and for the 11.11 waitlist; the top-3 winning categories from search_performed by category_l1; cohort format and its curriculum; the KOC/partner list with per-creator FQSA; the localised ID/TH page cluster. Retires: the "free forever-ish" framing (replaced by priced credits, shown before payment — monetisation begins with honesty, not a surprise invoice); one-shot countdown mechanics (replaced by a calendar-anchored lead-time story, which is genuinely urgent: air-freight stock for 11.11 needs to be ordered roughly 3–5 weeks ahead [ASSUMPTION], sea freight 6–8 weeks [ASSUMPTION], both to be verified with real partner lead times); the 9.9-specific asset pack (superseded by an 11.11 selection calendar published weekly from Oct); paid scaling discipline from Sep 8–14 (frozen at whatever the Sep 14 cost-per-FQSA justified, ramp only per the unlock gates in paid-media.md). The single narrative line for October: "9.9 was the warm-up; 11.11 is the real order." Sell lead-time certainty, not discounts.

Master Roadmap: Sep 2 → Nov 11

Do not start before Sep 22: payments/wallet integration, market-specific product customisation beyond language, a mobile app, custom supplier-model retraining, affiliate platform back-end, new countries beyond ID/TH/VN/PH/MY/SG, and any hire beyond the current 1.5 people. Every one of these is a legitimate future need and a fatal distraction during the free window.

Capacity Reality & Cut List

Planning assumption: 1.5 people — A = growth/content/community (1.0), B = product/sourcing/support (0.5). At that headcount, roughly 60 hours of real work per week exists after email, payments, and life. The roadmap above asks for about 100. So the cut list is the plan, not a caveat.

The one highest-leverage 2-hour block per week: - Sep 2–7: A builds the dashboard and runs the link lint until it actually works. Nothing else in this file has any value if Sep 8 is unmeasured. - Sep 8–14: B clears the search_no_results backlog (top 20 queries indexed by hand) — this directly raises the north star on traffic we've already paid for. - Sep 15–21: A writes the retro memo and the two case studies while memory is fresh. Deferred, this never happens, and October repeats September's mistakes. - Sep 22–30: B interviews 5 Activated Sellers by WhatsApp about what blocked them from ordering. Cheapest product insight available, forever. - Each October week: A publishes the week's 11.11 lead-time content, anchored on a confirmed freight cut-off. This is the only thing that makes our urgency honest rather than recycled.

Risk Register

If Only 3 Things Happen Next Week

If everything else in this file is dropped, these three are what make Sep 8 survivable and W4 retro possible.

1. Get attribution live before any spend (A, Sep 2–5). Why: Sep 8 without campaign-level attribution is 7 days of non-repeatable experiments. By the time we look in W4 we will know how much we spent and nothing about what earned the sellers, so October's budget decisions become vibes. This is the one dependency nothing else in this file can work around. Minimum viable version: the 12 campaign names registered in one sheet; 6 events instrumented rather than 10 (signup, search, no-results, saved, contact, paywall); one Google Sheet updated each morning by hand from GA4 exports — no dashboard build, no warehouse, no BI tool. Plus the claim: on Sep 5 you can look at any signup and state which campaign produced it.

2. Fix the top search failures so traffic doesn't land on a dead end (B, Sep 3–6). Why: the funnel target that most worries us is signup→first_search, and the fastest way to destroy it is paying for searches that return nothing. Failure also poisons the community channels — one "I tried it and there was nothing for my category" in a 40,000-member group costs more than the whole USD 420. Minimum viable version: export every search_no_results from Sep 2–6; hand-index the top 20 queries (usually 5–6 categories) by finding real 1688 suppliers manually and pasting them into the index; ship an honest empty-state that logs the request and promises a reply. Then post publicly: "we couldn't find X last week; we added it." That post is more trust-building than any ad.

3. Plant the warm channels and hold 20 real seller conversations (A+B, Sep 4–7). Why: community, WhatsApp/Telegram and the Sep 4 asset pack are the only channels where a 1.5-person team can produce FQSA at near-zero cash cost, and they are the only channel that keeps working when paid gets banned. Twenty conversations also generate the copy for every offer in this file — the objections sellers raise in DM are the exact words the Sep 7–9 posts should use. Minimum viable version: one asset pack shipped Sep 4 in the 4 groups we can serve honestly; 20 DM/chat replies logged with the objection written verbatim into a sheet; one cohort thread open with registrations; consent captured in writing before any broadcast send (risk register above). "20 conversations" means 20 real exchanges, not 20 reactions.

Everything else — TikTok, offline QR, partners, SEO expansion, the second AMA — is a good idea queued behind these three, and is the correct thing to lose if the week goes badly.