Building a 'Karachi-Scale' Lead Gen Bot

Why Standard Lead Gen Fails in Pakistan

Most agencies in Karachi are still running the same playbook from 2018: find a business on Google Maps, copy-paste their email from the website, and blast a generic pitch. The result? Sub-1% reply rates, bruised egos, and wasted hours. I built the Karachi-Scale Lead Gen Bot specifically because I was tired of watching talented people grind away at work a machine should be doing — and doing better.

The core insight is simple: a prospect who receives a pitch that contains three specific facts about their business — their PageSpeed score, a gap in their competitor's SEO, and a mention of a recent local news story — is 11 times more likely to respond than one who gets a generic email starting with "I hope this message finds you well." The bot doesn't just find leads. It understands them before the first word of outreach is written.

The 11-Source Enrichment Stack

The enrichment pipeline is the heart of the system. Before a single pitch is generated, the bot runs each prospect through eleven independent data sources. Here's exactly what each layer contributes:

Google PageSpeed Insights (PSI): Pulls Core Web Vitals scores. A mobile score below 50 is an immediate pain point to open with. We've found that Karachi restaurant websites average a 31/100 mobile score — a gift for any opener.
Hunter.io: Verifies email deliverability and finds the correct decision-maker address. Pitching info@ is amateur hour.
Wappalyzer: Reveals the tech stack. If a business is running a WordPress site from 2019 with no CRM, that's a diagnosed problem, not just an assumption.
WHOIS Lookup: Domain age, registrar, and expiry. A 6-year-old domain expiring in 30 days tells a story about neglect that the pitch can address directly.
Open Graph Meta Scrape: Checks social preview images and descriptions. 70% of Karachi businesses have broken OG tags, meaning their WhatsApp shared links look broken to every customer.
Social Presence Scan: Checks Instagram, Facebook, LinkedIn, and Google Business profile completeness. Missing platforms are explicit opportunities.
Abstract API: Email validation and company data enrichment at scale, without burning Hunter credits on invalid addresses.
Trustpilot & Google Reviews: Sentiment analysis of the top 20 reviews. A pattern of "slow delivery" complaints is a direct hook for a logistics automation pitch.
TripAdvisor (hospitality verticals): Competitive ranking and review trend analysis for restaurants, hotels, and tourism operators.
Yelp: For businesses with any US or expat customer base — crucial for Karachi's DHA and Clifton F&B sector.
Unsplash / Visual Quality Check: Assesses whether the business is using generic stock photos. Visual poverty is a legitimate growth problem.

Once all 11 sources have returned their data, a Gemini 2.5 Flash agent synthesizes the enrichment into a structured JSON brief. This brief is what the AI pitch writer actually reads — not raw scraped data.

The 3-Tier AI Logic Chain

Cheap models hallucinate. Expensive models are slow. The solution is a tiered chain where each model does only the work it's best suited for.

Tier 1 — Gemini 2.5 Flash (Triage): Takes the raw enrichment brief and scores the lead on a 0-100 opportunity scale. Leads below 40 are dropped. This alone eliminates 30-40% of the pipeline before a single premium token is spent. The Flash model also categorizes the business type (Restaurant, Clinic, Retail, Real Estate, Education, Barber/Salon, Other) which gates which pitch template architecture gets used downstream.

Tier 2 — Gemini 2.5 Pro (Strategy): Receives the enrichment brief + category classification and generates three distinct pitch angles. Each angle focuses on a different pain point uncovered in the data. The Pro model is instructed to write in a "Diagnostic Gift" format — the opener reveals a specific, real problem the prospect didn't know we'd identified. This is not flattery; it's forensics.

Tier 3 — Claude Sonnet (Closer): The highest-fidelity writer in the chain. Takes the single best pitch angle (selected by Pro) and writes the final email with precise tone calibration. For Pakistani B2B targets, the tone is authoritative but direct. For expat-facing businesses, the register shifts to Western professional. Claude is also responsible for the subject line — often the only thing that determines whether the email is opened at all.

You can explore how the full outreach pipeline is structured by visiting the Karachi Agency page, or test a live SEO diagnostic at the SEO Audit Tool to see the kind of data the bot collects on every prospect.

Database Architecture and Deduplication

One of the most painful lessons from early runs: pitching the same business twice. It destroys credibility instantly. The bot maintains a SQLite database via OutreachDB — every pitched business is logged with timestamp, pitch tier used, email sent, and reply status. Before any enrichment is run, the bot checks this table. If a business has been contacted in the last 90 days, it's skipped. If it was contacted and replied positively, it's flagged for the human follow-up queue rather than automated re-engagement.

The database also tracks enrichment cache. PSI scores are cached for 7 days. WHOIS data for 30. This prevents redundant API calls and keeps costs predictable — a critical concern when running at 500+ leads per week.

Real Numbers From Production Runs

In a recent 200-lead batch targeting Karachi F&B businesses with no functional website, the results were:

Average enrichment time: 14 seconds per lead
Leads dropped after Tier 1 triage: 38%
Emails sent: 124
Open rate: 61% (subject line engineering works)
Reply rate: 9.7%
Discovery calls booked: 8
Conversion to paid project: 3 (from that single batch)

At a PKR 150,000 average project value, that's PKR 450,000 in pipeline from one automated run that took no human labor beyond initial setup. If you want to build a system like this for your agency, the AI Freelancers Course covers the full architecture, or you can generate a cold email right now to see the output quality firsthand.