468 sessions were recorded between February 25 and March 13, 2026.1 Of these, 190 (41%) were button-click-only bounces — zero product input, not evaluatable as conversations. All analysis uses the 278-engaged denominator.
We selected 8 sessions to maximize diagnostic coverage — not to represent the average, but to expose the range of failure modes and the rare successes.
| # | ID | Product | Stage | Msgs | Why Selected |
|---|---|---|---|---|---|
| 1 | f645035c | Custom Leather Backpack | COMPLETE | 22 | Happy path — only session with confirmed SR in sample |
| 2 | 98ced3ab | Wallet + Multitool | S4 | 34 | Highest engagement dropout. Passive close pattern. |
| 3 | 32a76c73 | Coffee Bags + Glass Bottles | S2 | 40 | Longest S2 session. Multi-product. Contact form gap. |
| 4 | 962896d6 | Custom Logo Tumbler | S2 | 10 | Typical quick mid-funnel exit |
| 5 | b0dcf410 | Ceramic Perfume Bottle | S1 | 22 | Best value delivery turn in entire dataset |
| 6 | 2db27f98 | 2K Computer Monitor | S3 | 28 | MOQ feasibility loop — user changed qty 6 times |
| 7 | 4d340dad | Market Advice (Chinese) | S1 | 16 | “What should I source?” dead-end. Full Mandarin. |
| 8 | fa0892e5 | Blind Box | S1 | 8 | Product misunderstanding, no recovery |
Full production transcripts, lightly annotated. Highlighted lines mark key moments referenced in the problem catalog (Part III).
42 problems identified across 8 sessions, grouped into 7 categories. Each includes the problem, where it occurs, an exact quote, and an estimated frequency across the full 468-session dataset where possible.
The bot reaches decision points and puts the burden on the user instead of driving forward.
| ID | Problem | Session(s) | Quote | Est. Freq |
|---|---|---|---|---|
| P-01 | Passive close after buying signal. User says “Ok” after pricing — a buying signal — and the bot retreats to “let me know.” | 98ced3ab | “If you have any other questions or need further assistance, just let me know!” | ~15 S4 users lost this way2 |
| P-02 | “Just let me know” as final message. User explicitly asks “What’s next step?” and the bot responds with another passive prompt instead of driving to contact capture. | 98ced3ab | “If you’re ready to proceed, we can finalize your request. Just let me know if you’d like to continue!” | Appears in 5/8 sessions as a pattern |
| P-03 | No forward momentum after destination update. User provides destination — a clear intent signal — and the bot stalls with “let me know.” | 962896d6 | “Let me know if you want to proceed or adjust any details!” | Common at S2 boundary |
| P-04 | Contact form deferral with no fallback. Bot says “fill in your details below” but the frontend card may not render. No conversational fallback to collect name/email directly. | 32a76c73 | “The frontend will show a contact form card for you to fill in your details. Please check there!” | 2/8 sessions show user confusion at this step |
| P-05 | Contradictory finalization. Bot says “I don’t need any specific info from you right now” then “fill out the contact form.” User left, confused about what to do. | 32a76c73 | “I don’t need any specific info from you right now. Just fill out the contact form that appears next!” | 1/8 but reveals a structural gap |
| P-06 | No contact details captured at S4. Across all Stage 4 sessions, the bot never directly asks for name, email, or phone. It defers entirely to the UI form. | 98ced3ab, 32a76c73 | (pattern across sessions) | 15 S4 users with zero contact capture2 |
| P-07 | No recap before finalization. The bot moves to “fill in your details” without summarizing what the user is committing to (product, qty, price, destination). | f645035c, 98ced3ab, 32a76c73 | (pattern — finalize flow skips recap) | Consistent across all sessions reaching S3+ |
The bot’s text delivers almost no sourcing expertise. 78.7% of turns score D2=0 under Eric’s rubric.4
| ID | Problem | Session(s) | Quote | Est. Freq |
|---|---|---|---|---|
| P-08 | “Pick your options above” as navigation prompt. Bot’s text is a UI pointer, not a conversation. No sourcing context accompanies the handoff. | f645035c, 962896d6, b0dcf410, 32a76c73 | “Got it! Pick your options above, or add a reference image.” | Appears in ~70% of turns across all sessions |
| P-09 | Pricing “is above” with zero numbers in text. The bot shows pricing in a UI card but the text contains no dollar amounts, no MOQ context, no comparison. | f645035c, 962896d6, 32a76c73, 2db27f98 | “Estimated pricing above is based on market data. Change destination or quantity to refine.” | Standard response in ~80% of pricing turns |
| P-10 | Only 2/89 turns deliver specific value. Session b0dcf410 Turn 7 ($2.62-$5.14 DDP, Alibaba links) is the exception that proves the rule. The bot can surface pricing in text — it just doesn’t by default. | b0dcf410 | “Sourcy DDP Range: $2.62 - $5.14 per unit | Alibaba FOB Range: $1.50 - $4.00 per unit” | 2.2% of turns |
| P-11 | Generic one-liners instead of category expertise. Bot inserts filler like “Leather backpacks balance style and durability” or “Higher refresh rates enhance gaming” — neither specific nor actionable. | f645035c, 2db27f98 | “Higher refresh rates enhance gaming experience significantly.” (user wanted professional monitors) | ~19% of turns have D2=1 (generic) |
| P-12 | No material tradeoff guidance. Leather backpack session: no full-grain vs top-grain. Tumbler session: no 304 vs 201 steel. Wallet session: no leather grades or stitching methods. | f645035c, 962896d6, 98ced3ab | (absence — zero material expertise across sessions) | 0/8 sessions contain material-specific advice |
| P-13 | No price anchoring. A sourcing expert would open with “Custom backpacks run $8-25/unit depending on material” to anchor expectations early. The bot never provides a price range before the user commits to customization. | All | (absence) | 0/8 sessions contain early price anchoring |
When users ask trust questions, the bot gives corporate filler. Mean Trust Score across sessions: 3.9/10.5
| ID | Problem | Session(s) | Quote | Est. Freq |
|---|---|---|---|---|
| P-14 | Corporate trust response. User asks “How can I trust Sourcy?” and gets PR copy with no specifics — no supplier count, no order stats, no named references. | 98ced3ab | “Sourcy leverages real market data, expert sourcing insights, and a network of vetted suppliers…” | Trust questions appear in ~5% of sessions |
| P-15 | Empty social proof. “What are ur top customers?” gets a generic category list instead of any real signals (e.g., “50+ brands in consumer goods” or “$33M GMV”). | 98ced3ab | “Sourcy serves a diverse range of clients, including e-commerce brands, retailers, and manufacturers…” | Same pattern whenever social proof requested |
| P-16 | No QC/sample mentions. Across 8 sessions, quality control process, sample availability, and inspection protocols are never proactively mentioned. | All | (absence — QC mentioned in 1/8 sessions, passively) | 1/8 sessions. Eugene’s V3 TR-01 confirms systematic gap.5 |
| P-17 | No DDP transparency by default. Only session b0dcf410 shows landed cost breakdown. All other pricing turns hide the DDP/FOB split behind the UI. | b0dcf410 (only positive) | “Sourcy DDP Range: $2.62 - $5.14” | 1/8 sessions show DDP transparency |
| ID | Problem | Session(s) | Quote | Est. Freq |
|---|---|---|---|---|
| P-18 | Sent user to competitor platforms. User asks for certified manufacturers. A sourcing platform’s bot tells them to search Alibaba, Global Sources, Made-in-China. | 4d340dad | “我无法直接提供具体厂家，但可以建议你在 Alibaba, Global Sources, Made-in-China” | Eugene identified 22 users in this “what should I source” pattern3 |
| P-19a | No routing to Sourcy’s team. When the bot can’t help (no specific product, or needs human sourcing), it should route to Sourcy’s sourcing team. It never does. | 4d340dad, fa0892e5 | (absence — no handoff mechanism to human team) | All dead-end sessions lack routing |
| ID | Problem | Session(s) | Quote | Est. Freq |
|---|---|---|---|---|
| P-19 | Product category misidentification. “Blind box” (mystery collectible toy category) interpreted as a packaging product (cardboard/acrylic box). | fa0892e5 | “Material: Cardboard is popular for cost-effectiveness, while acrylic offers a premium feel.” | Eugene flagged this pattern in his analysis3 |
| P-20 | No recovery after user correction. User explained what a blind box actually is. Bot acknowledged with “Got it!” then repeated “choose from the options above” — same mismatched options. | fa0892e5 | “Got it! For your blind boxes, please choose from the options above.” | Unclear — need more correction-recovery examples |
| P-21 | Insight doesn’t match stated use case. User specified “Sleek & Professional” monitors. Bot offered gaming tip: “Higher refresh rates enhance gaming experience.” | 2db27f98 | “Higher refresh rates enhance gaming experience significantly.” | Contextual mismatch in ~15% of insight turns |
| P-22 | Restricted product not flagged. Perfume is a flammable product category requiring MSDS, hazmat shipping classification, and special packaging. Bot proceeded without any safety flag. | b0dcf410 | (absence — no restricted-product check triggered) | Unknown — need to audit all product categories |
| P-23 | Brand reference ignored. User mentioned “chako lab design tumbler” — a specific brand reference. Bot didn’t acknowledge or use it to personalize the response. | 962896d6 | “Got it! Pick your options above.” (no brand acknowledgment) | Likely common when users paste brand names |
5 of 8 sessions contain at least one turn with format issues.
| ID | Problem | Session(s) | Quote | Est. Freq |
|---|---|---|---|---|
| P-24 | Three-question dump. Bot fires 3 numbered questions in a single message instead of one at a time. | 98ced3ab | “1. What’s the primary wallet material? 2. What’s the desired surface texture? 3. What’s the overall shape/style?” | Occurs when user describes a new product without chip UI |
| P-25 | Markdown image wall. 5 concept images dumped with bold labels and markdown image syntax in a single message. | 98ced3ab, 32a76c73, b0dcf410 | “1. **Monochrome Focus** ![image] 2. **Elegance Defined** ![image]…” | Every visual concept turn (100% of visual responses) |
| P-26 | 12-question interrogation. Chinese session: bot fired questions across 3 product categories simultaneously — 12 questions in one message. | 4d340dad | [3 category headers, each with 3-4 numbered questions] | Triggered by “all of them” requests |
| P-27 | Bold/numbered list in conversational context. Bot uses structured formatting (bold headers, numbered lists) that breaks conversational tone — reads like a report, not a chat. | 4d340dad, b0dcf410, 32a76c73 | (pattern across sessions with detailed responses) | 5/8 sessions have at least one occurrence |
| P-28 | Internal architecture exposed. “The frontend will show a contact form card” — user-facing text leaks implementation details. | 32a76c73 | “The frontend will show a contact form card for you to fill in your details.” | Appears at finalization in sessions with form rendering issues |
| ID | Problem | Session(s) | Quote | Est. Freq |
|---|---|---|---|---|
| P-29 | MOQ loop with no escape. User changed quantity 6 times. Bot responded with the same feasibility template each time. Never asked why the user kept changing. | 2db27f98 | “Feasibility is red again. Would you like to increase to 100 or 300?” (for the 3rd time) | Unknown — feasibility sessions need audit |
| P-30 | No stall detection. Session 2: two consecutive empty “let me know” messages. Session 3: user typed “?” twice. Bot doesn’t detect that the conversation is dying. | 98ced3ab, 32a76c73 | “?” (user) → same template response (bot) | Appears in sessions where user is confused |
| P-31 | Product switch without reconciliation. User started with multitool, added wallet. No cross-product intelligence (“since you’re ordering both, you might get a packaging discount”). | 98ced3ab, 32a76c73, b0dcf410 | (absence — multi-product sessions treated as independent) | 3/8 sessions involve product switches |
| P-32 | Abandoned product not followed up. Session 5: user added lipstick, then focused on perfume. Lipstick never mentioned again. Session 3: glass bottles abandoned for coffee bags. | b0dcf410, 32a76c73 | (absence — no “what about the lipstick you mentioned?”) | 2/3 multi-product sessions |
| P-33 | Explorer users hit a dead end. Users who want recommendations (“what should I source?”) have no path forward. Bot funnels them into customization before they’ve committed to a product. | 4d340dad | “你对哪种母婴产品感兴趣？” (into customization flow prematurely) | 22 users in this pattern across 4683 |
| P-34 | Confusing error message for Alibaba match failure. “Couldn’t find matching listings on Alibaba for 2000 units” implies the quantity is wrong. The issue is product matching, not quantity. | 32a76c73 | “We couldn’t find matching listings on Alibaba for 2000 units.” | Unknown — depends on Alibaba match rate |
| P-35 | Stage classification unreliable. Session b0dcf410 classified as Stage 1 despite reaching pricing, feasibility, and near-finalization. Multi-product or non-linear sessions lose stage tracking. | b0dcf410 | (data issue — metadata says S1, behavior shows S3+) | Affects accuracy of funnel metrics |
| ID | Problem | Session(s) | Description | Impact |
|---|---|---|---|---|
| P-36 | No qualification questions. Bot never asks about use case (personal? resale? corporate gift?), budget range, or buying timeline context beyond the structured inputs. | All | D3 (qualification) scores 0 in 86% of turns. Signal obtained only once across 89 turns. | Can’t segment leads by intent |
| P-37 | No conversational contact capture. When the UI form doesn’t render, the bot has no fallback to ask “What’s your name and email?” directly in chat. | 32a76c73, 98ced3ab | Entirely dependent on frontend card rendering | Hard-blocks conversion when UI breaks |
| P-38 | No catalog/browse mode. Users without a specific product (“what should I source?”) need a different flow — trending products, category recommendations, guided discovery. | 4d340dad | 22 users hit this pattern3 | 100% loss rate for explorer users |
| P-39 | No Sourcy-vs-DIY comparison. Bot never explains why using Sourcy is better than the user doing it themselves on Alibaba — even when users ask trust questions. | 98ced3ab, 4d340dad | (absence) | Trust Score TR mean: 3.9/105 |
| P-40 | No urgency or scarcity signals. At decision points, the bot could mention supplier lead times, price volatility, or limited MOQ windows to drive action. It never does. | All | (absence) | No conversion urgency at any stage |
| P-41 | No shipping/customs preview. Only session b0dcf410 mentions “Sourcy will handle customs duties.” All other sessions give pricing without any shipping context. | b0dcf410 (only positive) | “Shipping to the Philippines may include customs duties, but Sourcy will handle that for you.” | 1/8 sessions address logistics |
| P-42 | No human handoff mechanism. When the bot reaches a dead end (can’t help, user frustrated, complex request), there is no way to escalate to a human sourcing agent. | 4d340dad, fa0892e5 | (absence — conversation just ends) | 100% loss rate for dead-end sessions |
Which problems are killing which stage transitions? Mapping the 42 problems to the funnel stages where they cause users to leave.
Users who typed a product but never reached customization/pricing.
| Problem | Mechanism | Evidence |
|---|---|---|
| P-08 “Pick options above” | First substantive bot response after product input is a UI pointer with zero sourcing context. No hook to keep the user engaged. | 4/8 sessions open with this exact response after product clarification |
| P-13 No price anchoring | User gets no indication of what things cost until deep into the flow. No early value to justify continued engagement. | 0/8 sessions provide a price range before customization |
| P-33 Explorer dead end | Users without a specific product (“what should I source?”) have no path forward. Bot pushes customization before commitment. | 22 users (8% of engaged) in this pattern. 100% loss rate.3 |
| P-19 Product misidentification | Bot interprets the product wrong and doesn’t recover when corrected. User gives up. | Session fa0892e5: 8 messages, user left after correction was ignored |
The biggest dropoff. Users who reached pricing/visuals but never got to feasibility or refinement.
| Problem | Mechanism | Evidence |
|---|---|---|
| P-09 Pricing “is above” | User sees a price card in UI. The text says nothing about what it means, how it compares, or what to do next. No interpretation, no hook. | Standard response in ~80% of pricing turns across all sessions |
| P-03 No forward momentum | After destination or pricing update, bot says “let me know” instead of asking a specific next question. | Session 962896d6: user provided Jakarta destination, got passive close, left |
| P-11 Generic one-liners | Bot’s material insights are filler (“durable and stylish”) not hooks. Nothing to make the user think “this bot knows sourcing.” | ~19% of turns. Never product-specific enough to be useful. |
| P-25 Markdown image wall | Visual concepts dumped as raw markdown. On some interfaces, this renders as broken image links or a wall of text. | 100% of visual responses use this format |
| P-29 MOQ loop | Users who explore different quantities get stuck in a repeat cycle with no escape or guidance. | Session 2db27f98: 6 quantity changes, same response, user left |
Users who reached feasibility but didn’t get to finalization.
| Problem | Mechanism | Evidence |
|---|---|---|
| P-04/P-05 Contact form gap | Bot says “fill in your details below” but the form doesn’t render. Bot has no fallback. User asks “where?” and gets “the frontend will show a contact form card.” | Session 32a76c73: user asked “where?” and “what info do you need?” — got contradictory answers |
| P-37 No conversational contact capture | When the UI form fails, the bot has no way to collect name/email/phone in chat. | Hard-blocks conversion in any session where form doesn’t render |
Users who reached finalization and left without submitting. The most painful dropoff — these were ready to convert.
| Problem | Mechanism | Evidence |
|---|---|---|
| P-01/P-02 Passive close | Bot says “just let me know if you’d like to continue” instead of recapping the deal and asking for contact details. | Session 98ced3ab: user asked “What’s next step?” → bot: “Just let me know!” → user left |
| P-06 No contact capture | The bot never directly asks for name/email at S4. Relies entirely on the UI form appearing. | 15 S4 users, zero contact details captured in non-SR sessions2 |
| P-07 No deal recap | User is asked to “fill in details” but the bot never says what they’re committing to: product, qty, price, destination, timeline. | No recap observed in any S3+ session |
| P-14/P-15 Trust failure | User at S4 asks “How can I trust Sourcy?” and gets corporate filler. This is the moment to close — and the bot fumbles it. | Session 98ced3ab: trust question mid-finalization, got PR copy |
| Fix | Problems Addressed | Users Affected | Effort |
|---|---|---|---|
| Active close-driving at S4 | P-01, P-02, P-06, P-07 | ~15 users (recoverable) | Prompt change |
| Pricing interpretation in text | P-09, P-10, P-17 | ~129 S2 users (partial) | Prompt change |
| Explorer routing | P-33, P-18, P-38 | ~22 users | Prompt + flow change |
| Conversational contact fallback | P-04, P-05, P-37 | S3-S4 when form fails | Prompt change |
| Trust content defaults | P-14, P-15, P-16, P-39 | S4 trust questioners | Prompt change |
| UI bug fixes (V2) | P-25, P-04, P-28 | S2-S4 display issues | Engineering (Ziya) |
| Correction recovery | P-19, P-20, P-21 | Product mismatch sessions | Model/architecture |
The tool pipeline works — users who persist through the full flow reach SR submission. The problems are in the text layer, the close-driving, and the edge cases.
Of 42 problems: 7 are close-driving (prompt-level fix, highest SR impact), 6 are value delivery (prompt-level fix), 4 are trust gaps (content fix), 5 are format violations (prompt constraint fix), 5 are product understanding (model/knowledge fix), 7 are flow issues (prompt + architecture), and 7 are missing capabilities (new features needed).
The single highest-leverage change: replace every “let me know” with a specific next-step question. This is a prompt change that could recover a meaningful portion of the 15 Stage 4 users who reached the finish line and left.