Console

Builds the broker universe: people at the appointed brokerages whose titles match the construction, energy, and surety practice list.

Search costs nothing, so the universe is free to build and only revealing an email spends a credit. 338 of the brokers on file came from here.

Named practice leaders pulled from firm leadership pages, press releases, and trade coverage, each stored with the URL it came from.

Apollo misses the senior named leaders, which are exactly the people worth a first email. 105 rows came in this way and the import is a CSV, so anyone can add to it.

It also caught two firms that no longer exist independently. Those rows are flagged, not rewritten.

Construction Dive, ENR, Insurance Journal, Risk & Insurance, Insurance Business, read on demand and classified into signal types.

Third-party intent for free. Project awards, appointments, and moves are published; nobody in this market is systematically acting on them.

Extra feeds and keyword watches, added in the Console. Describe one in a sentence and a model drafts the config.

The five feeds are my guess at what matters. An underwriter who reads one newsletter religiously should be able to add it without asking an engineer.

700,000+ project starts a year with owner, general contractor, and value attached.

The project-award signal at full strength instead of whatever the trade press happens to cover. Dodge already sells to insurers, so the data is proven; the question is price against the free version.

Only after the free trade-press signal proves it converts. Buying data before the play works is how a budget disappears.

Resolves anonymous company visits on shepherdinsurance.com to a firm, and ideally an office.

A broker reading the Excess Casualty page is first-party intent and is worth more than any signal in the trade press. Nothing captures it today.

Sends every broker-phrased prompt to every engine with a key, and stores the answer, the carriers named in order, and the citations.

It is the free baseline. Before spending anything on measurement, you need thirty days of numbers to argue with, and this produces them at the cost of the model calls.

AEO and GEO tracking across ChatGPT, Perplexity, Google AI Overviews, Google AI Mode, Gemini, Claude, Copilot, and Grok, with share of voice, citation rate, sentiment, and competitive positioning.

The cheaper alternative in the category, with a free tier. Kept on the list as the fallback if Profound's price does not survive procurement.

The category leader: visibility, share of voice, citation share, and sentiment across every major answer engine plus Google AI Overviews and AI Mode, with prompt volumes and conversation-level data.

The audit already speaks Profound's language: I used their published metric definitions for visibility, average position, share of voice, and citation share, so the day Shepherd subscribes, every stored baseline carries over with zero lost history. And it covers the one surface no model key can reach - Google AI Overviews and AI Mode, where a broker who types a question actually lands.

Buy at day 30, not day 1, once the free baseline has moved. Published pricing starts around $500 a month.

Two API keys that cost nothing and light up two more engines in the audit, including the grounded Google one.

The adapters are already written and tested. This is the cheapest coverage in the whole plan: ten minutes and no card.

Scores brokers on the four axes, classifies signals, researches the one-pager with live search, and drafts every email an underwriter signs.

Quality matters most on the two things a human reads: the brief before a call and the email that goes out under someone's name. Everything cheap and repetitive runs on the smaller model instead.

Nothing calls a model on a timer. Every model call in this build happens because a person clicked.

Approved drafts leaving from the underwriter's own mailbox, with replies threading back into the engine.

Two real options: send natively through the Gmail or Microsoft API, or push into whatever sequencer the team already pays for. The answer depends on what the underwriters use today, which is a week-one conversation, not a guess.

Deliberately unbuilt. Wiring a sender before agreeing whose mailbox it is would be the wrong order.

Wherever a submission is recorded today. Salesforce, HubSpot, or the policy system.

Every number on this site stops at 'email sent' until a submission record closes the loop. It is also the appointment feed that triggers the day-61 sequence for a brand new office.

Question one on our next call: where does a submission live, and what counts as one.

A slash command that returns the broker one-pager where the underwriters already are.

The brief is the most-used thing in this build and nobody wants to open a website before a call. Small piece of work, real adoption difference.

Paid ABM pointed at the exact company and title list Play 1 is about to email, exported straight from the audience builder.

It is the only paid channel where the targeting matches the audience: named firms, named titles, no email address required. Start small with a held-out control so the lift is measurable rather than assumed.

Kill at 90 days if the targeted offices do not beat the control.

Money behind the two clips that already earned attention organically, aimed at the same broker audience Play 2 builds.

Never boost a clip that has not earned it. Starts day 61, so there is organic data to pick from.

Turns one recorded hour with the underwriters into hooks, a storyboard, and a month of platform-native cuts.

The bottleneck on video is never the camera, it is deciding what to say and cutting it. This removes the deciding.

A contractor at roughly half a day a month, or me.

Twelve to sixteen clips a month is not a full-time job and does not need an agency. Captions and scheduling are under $100 on top.

Postgres behind everything: brokers, signals, cadences, drafts, briefs, audits, spend.

Free at this size, and it is plain Postgres, so nothing here is locked to it.

Hosting, deploys on push, and the environment where the keys live.

Right call while this is a take-home: nothing to babysit. A VPS becomes worth it when the monitors need a real scheduler rather than a serverless cron.