# RESEARCH BRIEF — Zaza Carz Inc (San Leandro, CA): inventory, economics, platform & operations, analysed to the bone

> **MODE: DEEP RESEARCH · MAXIMUM REASONING EFFORT · EXHAUSTIVE.** Think step by step and at length before writing. Use your **code-interpreter / Python** to load every CSV and compute; use **web browsing** for comps, recalls and vendor pricing; use **vision** on every hosted image. Do not sample, do not summarise a subset as if it were the whole, do not skip cars. Work through **all 94 vehicles individually**. Plan first (list the sub-tasks), execute each, then run a **chain-of-verification pass**: re-derive every headline number from the files, re-check every estimate against its comps, and list what you could not verify. Then a **self-critique pass**: what would a sceptical dealer-principal, a wholesale buyer and a lender each attack in this report — fix it. Only then write the executive summary. Prefer tables. Be specific, quantitative and sourced. Length is not a constraint; completeness and evidence are.

You are a senior automotive-retail analyst + used-car appraiser + dealer-operations consultant + web-platform forensic analyst, working with full autonomy and maximum depth. You have web access and vision. Your job is to turn the attached capture of a real used-car dealership's website (taken 2026-09-17) into the most complete, evidence-graded research report that can be produced from public data. Every number you output must be either **(a) observed** (cite the file + field/row), **(b) derived** (show the arithmetic), or **(c) estimated** (state the method, the comparable sources you used, and a confidence). Never blur the three.

## 1. What you have been given (17 flat files + hosted images)

| File | Contents | Use it for |
|---|---|---|
| `01_vehicles_master.csv` | 94 vehicles: VIN, year/make/model/trim, asking price, mileage, colours, drivetrain, fuel, MPG, CARFAX badge, feature/photo counts, description flags, **NHTSA decode** (HP, displacement, plant, body class, GVWR, airbags…), hosted photo URLs | the appraisal table, pricing analysis |
| `02_descriptions.md` | every listing description verbatim | disclosures, tone, "cash deal only", template detection |
| `03_features.csv` | 11,756 equipment lines (id, feature) | trim/equipment verification, option-value adjustments |
| `04_photos.csv` | 643 photo URLs (hosted + original CDN) | per-photo inspection if needed |
| `05_nhtsa_vin_decode_full.csv` | full NHTSA vPIC decode | engineering facts, recalls context (pull NHTSA recalls by make/model/year yourself) |
| `06_site_backend_vendors_report.md` | how the website works: hosting/ASN, platform (Carsforsale.com), every endpoint, DataDome/reCAPTCHA, Google tags, Drivee AI (Pirelly), CARFAX, CDN, TLS, cookies, data model, diagrams | platform/vendor/cost analysis, sales-flow reconstruction |
| `07_endpoint_probes.md` | 103 live probes of every filter/sort/AJAX/JSON endpoint with results | technical evidence |
| `08_dealer_profile_and_context.md` | address, hours, phones, About text, lead channels, **operator assumption: rent = USD 14,600 / month**, Terms clauses | dealer profile, opex model |
| `09_platform_vendors_traffic.json` | structured vendor/stack facts, GTM/GA ids, DataDome key, Drivee AI JWT claims, every third-party host contacted | vendor identification and cost research |
| `10_taxonomy_models_categories.json` | Carsforsale taxonomy ids, models JSON (with live counts), all 96 category pages | inventory structure checks |
| `11_inventory_summary.md` | human summary tables | quick orientation |
| `12_data_dictionary.md` | meaning of every column | definitions |
| `13_terms_and_conditions.txt` | the site's Terms (Carsforsale standard) | legal/compliance section |
| `14_computed_stats.json` | totals, distributions, group stats (by make/year/body/drivetrain/fuel/plant) | headline numbers (re-derive, don't just copy) |
| `15_image_index.md` | URLs of 8 overview contact sheets + 94 per-car condition sheets + screenshots | **look at these** |
| `16_method_playbook_and_pipeline.md` | how the data was captured (real Chrome, 97 requests, 0 blocks; verification steps) | provenance section |
| `17_features_frequency.csv` | frequency of each of 1,038 distinct features | equipment norms |

**Images (Cloudflare-hosted, public):** `https://zaza-dealer-data.pages.dev/sheets/overview-01.jpg … overview-08.jpg` (12 cars each) and `https://zaza-dealer-data.pages.dev/sheets/car-<id>.jpg` (every photo of one car, labelled with title/price/mileage/VIN). Full-size photos: `https://zaza-dealer-data.pages.dev/photos/<id>/01.jpg …`. Site screenshots: `https://zaza-dealer-data.pages.dev/site/*.png`. **You must open and inspect every one of the 94 condition sheets** — condition grading without looking is not acceptable.

## 2. Deliverable — one long report, this exact structure

### A. Executive summary (1 page)
Total asking value, estimated total wholesale value, estimated total retail-market value, implied gross if sold at ask, the dealer's positioning in one sentence, the five most important findings, and the three biggest uncertainties.

### B. Per-vehicle appraisal — all 94, one row each (table + a 2–4 line note per car)
Columns: id · vehicle · ask · miles · **condition grade (photo-based, 1–5 with rationale: paint, panels, wheels, glass, interior wear, warning signs, tyres if visible, dealer plate frames/temporary tags, evidence of reconditioning)** · notable equipment (from `03_features.csv`) · CARFAX signal · **est. wholesale / trade-in value (range)** · **est. retail market value in the SF Bay Area (range)** · ask vs market (cheap / fair / expensive, with %) · implied front-end gross at ask · days-to-sell risk · red flags (mileage vs age, sold-while-listed, missing photos, description contradictions, model-year vs VIN, known model problems e.g. BMW N20/N63, Nissan CVT, Hyundai Theta II engines, Ford PowerShift, Takata recalls — check NHTSA recalls per VIN pattern).
Method for values: use public comps you can actually retrieve (KBB, Edmunds, CarGurus, Cars.com, Autotrader listings for the same year/model/trim within ~100 mi of 94577; Manheim/MMR or Black Book if you can access them; otherwise build wholesale as retail-minus-typical-margin and say so). State the date of each comp. Show your adjustment logic for mileage and condition.

### C. Inventory economics
* Total asking value vs estimated wholesale value → implied aggregate gross, average gross per unit, distribution (which cars carry the margin).
* Price-setting behaviour: the $x,900 rule, how tightly the lot prices around $8,900, price per mile, price vs model year curve, whether German premium cars are priced at a premium relative to their wholesale.
* Stock profile vs the Bay Area market: is this cheap or expensive stock for the area? Which segments (older German luxury at 120–180k miles, hybrids, Nissan CVT cars) carry the most risk? What would a wholesale buyer pay for the entire lot as a package?
* Sold units: what the 4 sold cars tell you about what moves.

### D. Dealer operating model (build a monthly P&L with ranges)
Given rent = **USD 14,600/mo** (operator assumption), estimate: turnover (units/month) from lot size and typical independent-dealer turn rates for value stock; average front-end gross (from B/C); reconditioning per unit (smog in CA, safety, detail); staffing (a lot this size — how many people; commission structure typical for cash lots); DMV/BAR licence, bond, and CA dealer costs; insurance (garage liability, inventory); marketing (Carsforsale.com plan, any third-party listing syndication — check whether the inventory appears on carsforsale.com itself, CarGurus, Autotrader, Craigslist, Facebook Marketplace, OfferUp; search the VINs); Drivee AI subscription; CARFAX dealer program; utilities/software/phone; floor-plan interest (cash-only suggests none — argue it); sales tax/doc fee revenue (CA doc fee cap); back-end products (likely absent on a cash lot). Conclude with a break-even units/month and a plausible profitability range with explicit assumptions. Sanity-check the rent against San Leandro industrial/auto-lot rents.

### E. Platform, vendors and what they likely cost
For every vendor in `06`/`09`: what it is, what tier the dealer is probably on, evidence, and **researched list prices** (Carsforsale.com dealer website + inventory management plans; DataDome — is it CFS-wide or per dealer; reCAPTCHA Enterprise pricing; GTM/GA free; Drivee AI / Pirelly 360 pricing; CARFAX dealer advantage program; Google Maps embed; domain/TLS; Cloudflare). Produce a "monthly digital stack cost" range and say which costs are borne by Carsforsale vs the dealer.

### F. Sales & operations flow — reconstruct end-to-end
From the site's lead endpoints, forms, consent text ("I consent to be contacted by Carsforsale.com and the dealer…"), Drivee AI capabilities ("pricing, availability, trade-ins, book a test drive, switch to text, request a call"), CARFAX, the 7-day hours, cash-only policy, appraisal/trade-in forms: describe the likely funnel (traffic sources → lead → CRM → contact → visit → cash sale → paperwork), which systems touch each step (CFS CRM/lead routing, Drivee AI, phone, DMV e-filing via CFS or a DMV BPA provider, CARFAX), where the dealer likely sources cars (auction vs trade-in vs consignment — evidence: "We want your vehicle" banner, appraisal form, plate frames, one-owner share), how photos are taken (same backdrop, plate frames, sequence — count the standard shot list), and how long a car takes from acquisition to listing.

### G. Website & technology assessment
Quality of the digital presence versus competitors; SEO (sitemap, JSON-LD, canonical); performance; accessibility; conversion elements; data exposure (VINs in HTML, exposed API keys — assess risk, not exploit); bot defence adequacy; what a modern dealer stack would add. Rate it.

### H. Risks, red flags & open questions
Title/branding, odometer plausibility, model-specific failure risks, compliance (CA dealer regs, Buyers Guide/As-Is, smog, doc fees), the site's Terms vs data use, and what to verify in person or via DMV.

### I. Data provenance & confidence
Summarise how the data was captured (from `16`), what was verified (VIN check digits 94/94, NHTSA 94/94, page totals, category closure) and the confidence you assign to each section.

### J. Appendices
Full appraisal table as CSV-in-markdown; comp sources with URLs and dates; assumptions register; glossary.

## 3. Effort & method directives (non-negotiable)
* **Plan → execute → verify → critique → write.** Show the plan at the top of your working; keep it updated.
* **Compute, don't eyeball**: every total, average, distribution and ratio must come from code over `01_vehicles_master.csv` / `03_features.csv` / `14_computed_stats.json` (re-derive `14` — treat it as a cross-check, not a source).
* **Every car, every sheet**: 94 condition sheets opened and described; 8 overview sheets opened. Keep a running log "sheet car-<id>: opened ✔, observations: …".
* **Comps**: for each of the 94 cars find ≥2 retrievable retail comps (same year/model/trim, ±25k miles, ≤100 mi of 94577) and one wholesale/trade-in reference (KBB trade-in / Edmunds TMV / Black Book / MMR if reachable); record URL + date + price + miles. Where unreachable, say so and use the stated proxy method.
* **Recalls & known failures**: query NHTSA recalls per year/make/model and known-issue literature for the specific engines/transmissions in `05` (e.g. N20/N47/N63, Theta II, Nissan JF011E/JF015E CVT, PowerShift DPS6, ZF 8HP, Takata inflators); flag per car.
* **Syndication check**: search 10–15 VINs on carsforsale.com, CarGurus, Autotrader, Cars.com, Craigslist SF Bay, Facebook Marketplace, OfferUp; report where the dealer advertises and at what prices (price parity?).
* **Vendor pricing**: retrieve current list prices/tiers for Carsforsale.com dealer sites, DataDome, reCAPTCHA Enterprise, Drivee AI / Pirelly 360, CARFAX Dealer program, Google Maps Embed; cite pages + dates.
* **Assumptions register**: every estimate's assumptions in one table (value, source, sensitivity).
* **Confidence**: H/M/L on every section and every per-car value; state what evidence would raise it.
* **Contradictions**: dealer text vs NHTSA, listing vs sold status, description vs photos — resolve explicitly.
* **No hedging filler**: replace "may vary" with a range and a reason.

## 4. Working rules
1. Read all 17 files fully before writing. Load the CSVs into a table and compute; do not eyeball totals.
2. Inspect all 8 overview sheets and all 94 car sheets. Quote what you see ("front bumper scuff, left", "curb rash on 3 wheels", "aftermarket head unit") — specifics, not adjectives.
3. Prefer retrievable comps; cite URL + date. If a source is paywalled, say so and use the best public proxy.
4. Keep observed / derived / estimated labels on every figure; give ranges and a confidence (H/M/L).
5. Where the data contradicts itself (e.g. dealer engine string vs NHTSA displacement), say which to trust and why.
6. Do not contact the dealer, submit forms, or attempt to access anything beyond public pages and the hosted mirror.
7. Length: as long as it needs to be. The per-vehicle section alone should be ~94 × (table row + short note). Use tables generously; no filler.
8. Finish with a one-page "if you were buying this lot / competing with this lot / lending to this dealer" verdict.
9. Before delivering, run the checklist: 94 rows in Section B? every row has condition grade + wholesale range + retail range + verdict? every headline number re-derived? every comp dated? assumptions register complete? contradictions resolved? Only deliver when every box is ticked.

Start with the executive summary only after everything else is done; write it last, paste it first.