Lot Service Sales Outreach

The Problem Today

Redline knows exactly what a neglected lot looks like — it just has no system that points that knowledge at prospective customers.

Redline's whole business is fixing the thing this project hunts for: cars on a dealer lot with missing photos or a badly framed first photo. Every Merchandising Specialist visit exists because a lot's online inventory wasn't shot, or wasn't shot well. Redline is the expert on this exact pain — and a non-Redline dealership with a rising missing-photo percentage is, almost by definition, a dealership that needs Redline.

The trouble is that finding those dealerships is currently nobody's full-time job, and it would be a miserable one if it were. To do it by hand you'd have to: find the dealers near each Redline store, pull up each one's online inventory, count how many cars have no photos, eyeball the first photos, do that again next week to see if it's getting worse, figure out who to contact, and keep a running list of who's worth chasing. Across the footprint that's thousands of lots — it never happens consistently, so sales outreach runs on gut feel and referrals instead of evidence.

This page is deliberately split in two. Layer A — the intelligence engine — is high value, defensible, and worth building well. Layer B — automated cold email, contact-scraping, and purchased lists — carries real legal and deliverability exposure, and we are not going to pretend otherwise. The plan builds A first and treats B as a gated, reviewed, low-volume process — or hands hot leads to a human to contact.

How It Works

Two layers. The first is an analytics product. The second is a regulated sending process that must never run ahead of the first.

Layer A — the intelligence engine (the part with real, durable value)

This layer never sends anything to anyone outside Redline. It produces one thing: a trustworthy, ranked picture of which nearby dealers have a merchandising problem and how bad it's getting. That output is valuable on its own — it's a sales-targeting dashboard for Ryan and Lisa even if no automated email is ever sent.

• Geofence first, so we only ever look at reachable prospects. We take the address of every Redline store, and only consider dealerships within a 45-minute drive (real drive time via a distance-matrix API, not as-the-crow-flies miles). A dealer 45 minutes away by road is a serviceable prospect; one 12 miles away across a bay is not. This radius also keeps the scraping footprint small and purposeful rather than a blind sweep of the whole country.
• Measure the missing-photo % from public inventory. For each in-radius dealer we read their public online inventory listings and compute, per week, the share of vehicles with no photos (and, where we can, the share whose first photo has the same framing problems the First Photo vision model already detects — off-angle, cropped, obstructed). This reuses logic Redline already understands cold.
• Track the trend, not just the snapshot. A single bad week is noise. We store a weekly time series per dealer and fire a prospect signal on either of the doc's two triggers: missing-photo % crosses above 30%, OR it rises three weeks in a row (a lot that's actively getting worse — often a sign they just lost their photographer or vendor, i.e. peak buying intent).
• Rank into a running priority list. The system maintains a persistent, scored priority list: severity (how high the %), trajectory (rising how fast), proximity (drive time to the nearest Redline store), and dealer size (more VINs = bigger deal). This list is the product. It survives week to week, ages stale leads down, and promotes worsening ones up.
• Detect local clusters → suggest in-person visits. When 5+ dealers within an hour of a center point all show problems for 3 consecutive weeks, that's not a cold-email situation — that's a region worth a Redline Sales rep's windshield time. The system flags the cluster and suggests an in-person visit route rather than a blast.
• Weekly report to Ryan & Lisa. Every week, a digest for the Monday sales meeting: new prospects that tripped a trigger, movers on the priority list, suggested in-person clusters, and — once Layer B is live — outreach activity and email-open rates. This is the deliverable they actually asked for.

Layer B — outreach & data acquisition (the part with legal landmines)

This is where it stops being an analytics project and becomes regulated commercial email, web-contact harvesting, and possibly buying third-party email lists. None of that is free of risk, and the plan treats it accordingly. See the Risks section for the full, candid breakdown — but the short version of how we'd do it responsibly:

• Contact resolution from "Meet The Staff" pages. Per the doc, we cross-reference each prospect dealer against their public Meet-The-Staff page to attach a real contact name + email. Dealers with no such page go to a queue. We treat this as best-effort enrichment, not a guarantee — and we respect robots.txt and each site's terms.
• The "email purchases" queue is a yellow flag, not a green light. The doc proposes buying email lists for dealers we can't enrich. Purchased lists are the single highest-risk input in this whole project — for deliverability, for consent, and for data quality. We build the queue but recommend it be reviewed by a human and, ideally, that purchased contacts be contacted by a person, not a machine. We do not recommend feeding bought lists into an automated blaster.
• Sending, if we send at all, is slow, warmed-up, compliant, and reviewed. Accurate headers and "from", a real physical postal address, a working one-click unsubscribe, opt-outs honored fast and forever (a permanent suppression list), a dedicated warmed-up sending domain/IP, and low daily volume. Cold-blasting from Redline's main domain would torch its email reputation — a cost that lands on the entire company's email, not just this project.
• Default posture: route hot leads to humans. The strongly recommended default is that Layer A surfaces the prospect and a Redline salesperson sends a real, personal email or makes a call. That's better outreach and far lower risk than mass auto-mailing. Automated sending is an option to switch on deliberately, compliantly, and at low volume — not the centerpiece.

Data flow

           ┌──────────── LAYER A · INTELLIGENCE (safe, high value) ────────────┐
[ Redline store addresses ]
       │ geocode
       ▼
[ 45-min drive-time geofence ] ──► list of in-radius dealer domains
       │
       ▼
[ Inventory scraper ] ──► per-dealer VIN list + photo counts + first-photo images
       │                        (respect robots.txt / terms; throttled)
       ▼
[ Analyzer ]  missing-photo %  +  first-photo framing (reuse vision model)
       │  write weekly snapshot
       ▼
[ Time series / CRM-DB ] ──► trigger engine:  % > 30   OR   rising 3 wks
       │                                       cluster: 5+ dealers / 3 wks / 1 hr
       ▼
[ PRIORITY LIST ]  scored, ranked, persistent ──► [ Weekly report → Ryan & Lisa ]
       │                                              (+ in-person visit suggestions)
       └───────────────────────────────────────────────────────────────────────┘
                          │  hot prospect
   ════════════════════════ COMPLIANCE GATE ════════════════════════
                          ▼
           ┌──────────── LAYER B · OUTREACH (regulated, gated) ───────────┐
[ Meet-The-Staff enrichment ] ──► name + email   ── or ──► [ purchase queue ⚠ human review ]
       │
       ▼
   DEFAULT: hand lead to a human salesperson  ──►  personal email / call
       │
   OPTIONAL (deliberately enabled): automated send
       ▼
[ Email platform ]  warmed domain · CAN-SPAM headers · physical address
       │             one-click unsubscribe · permanent suppression list
       ▼
[ Open / engagement tracking ] ──► back into weekly report
       └──────────────────────────────────────────────────────────────┘

The AI/automation, honestly

There's no magic here, and it's worth being plain about what each piece actually is:

• Scraping is conventional, careful web crawling of public inventory pages — the hard part is that every dealer site is built differently, so it's an ongoing maintenance job, not a one-time script (see Cost to Own).
• First-photo grading reuses the same vision-LLM rubric from the First Photo project — applied to other dealers' photos instead of Redline's own.
• The trigger/trend logic is plain arithmetic on a time series (thresholds, three-week slopes, clustering within a radius). No ML needed and none pretended.
• An LLM is genuinely useful for drafting a personalized outreach note from a dealer's specific numbers — but a human should review before anything goes out, especially early.

Architecture & Stack

Scheduled scraping + analysis on AWS, a CRM/DB of record for the priority list, and a deliberately separate, gated email subsystem.

Why this shape

The work is fundamentally weekly and batch, not real-time, so it's a scheduled pipeline (EventBridge → Step Functions) rather than a standing service. Scraping runs on Fargate where a headless browser is needed and Lambda where simple HTTP fetches suffice, all throttled and polite. A relational store (Postgres) is the right home for the CRM-DB because the priority list, the weekly time series, dealer↔Redline-store distances, and contact records are all relational and queried as a set.

The email subsystem is drawn as a separate box behind a gate on purpose. It should be operationally and architecturally distinct from the intelligence pipeline so that (a) Layer A can ship and deliver value with Layer B switched entirely off, and (b) sending uses a dedicated, warmed domain isolated from Redline's transactional/internal email. For sending we'd lean toward an established email platform with built-in suppression, unsubscribe handling, and open tracking rather than rolling our own on raw SES — the compliance machinery is worth not reinventing.

Build Plan

Build the whole intelligence engine and prove its value before writing a single line of the sending layer. The compliance review is a hard gate, not a formality.

Data & Access Needed

Most inputs are public or already inside Redline. The items that gate the project are legal sign-off and a sending domain.

Risks & Open Questions

Layer A's risk is engineering (scraping is fragile). Layer B's risk is legal and reputational, and it is real. We are not going to soft-pedal it.

Cost to Own

The build is the easy part. The durable costs are scraper maintenance and — if Layer B runs — the ongoing compliance and reputation discipline.

The honest version: two things need ongoing human attention, and pretending otherwise would set the wrong expectation.

• Scraper maintenance. Expect to fix broken parsers regularly — this is steady, low-grade engineering work for as long as the system runs, and it's the line item most likely to be underestimated. A scraper that silently miscounts is worse than one that's obviously down.
• Compliance & reputation discipline (only if Layer B runs). Someone must own the suppression list, watch complaint/bounce rates, keep the unsubscribe flow working, and re-check the legal posture as laws and list sources change. This is governance, not code — and it doesn't end.