Understand the business. Understand the customers. Then let AI bridge the gap.

The starting point

The dealer had a strong reputation and plenty of tooling: a vertical point-of-sale system running the whole business, a paid SEO platform, a Yelp ad budget, and a Google Ads account that had been paused for two years. Almost none of it was connected, measured or reviewed. Answering the owner's question meant treating the work as an operational transformation with AI at the end of it.

Understand the business: research before anything else

I started in operations, not marketing. The POS was the system of record for cash, inventory and customers, so its reliability set the ceiling on everything downstream.

The method was contextual inquiry applied to a system: a read-only walk of every module, then controlled create-verify-delete tests of each workflow that touches cash or inventory. I documented each test as it ran, which turned the assessment into the business's first operating manual for a system that had none.

The assessment surfaced fourteen data problems nobody had seen, including phantom receivables, months of uncashed checks, two cash balances that disagreed with each other, inventory carried at the wrong cost, and stale open orders. Each was logged with a root cause, an owner and a remediation path. Several became the specification for the agents.

Understand the customers: three personas, one position to defend

The same research discipline, pointed at the other side of the counter. Observation at the counter, conversations with staff and a competitor scan produced three personas: the inheritance seller, the bullion buyer and the cash-for-gold customer. They arrive with different questions and different stakes, and the business had been treating them as one.

What held for all three was one position the store could defend: "We show you the math." Every later decision, from the Yelp plan to the ads keywords to what the agents watch, traces back to those three personas, that position and the problem list.

Strategy with gates, not a wish list

The strategy was a sequence, each step gated on the one before it:

1. A 30-day Yelp plan, measured at day 30.
2. Google Business Profile and the SEO platform, run as a 20-minute weekly routine.
3. A conservative Google Ads relaunch on sell-side keywords only, with a go/no-go decision.
4. The online store, launched in waves.

Five scorecard numbers, dated go/no-go gates through December, and a decision log. Decisions were written down as they were made: no Saturday hours yet, no asking for reviews, and nothing deployed to the live site without a person approving the batch. The plan lives on one board with an automated weekday brief, so the owner can see execution without a status meeting.

The stack, wired together with no code access

The POS vendor renders the website, and there is no code access. The work was finding the levers that did exist: a real CMS behind the storefront, a tag-manager slot, and the allow-list that governs every third-party script. With those three, in one week:

• On-page and technical SEO. Brand-voice titles and descriptions on the top pages, stale sitemaps replaced, canonical-URL and tracking errors fixed.
• Attribution. Call tracking with a website number pool and a dedicated Yelp number, both feeding GA4 and Google Ads.
• Governed automation. The SEO platform's on-page engine installed but held in review-only mode. It proposes; a person approves the batch.
• Visibility monitoring. Local rank tracking, a monthly Google Maps local-ranking heatmap, and AI-search visibility monitoring against hand-written local queries.
• A new channel. A marketplace connector live with graded coins listed on eBay, business policies set, and the connector's category bug diagnosed and worked around.

The principle throughout: instrument first, then automate, and never let a tool write to the live site or the live ad account unsupervised.

Let AI bridge the gap: agents on the controls, with authority earned at gates

With the data flowing, the last step was to stop depending on someone remembering to check. Five scoped agents built on the OpenAI API, each with one job, a schedule set by when its data is reliable, and a risk tier for what it flags. The agent with the most to get wrong was deployed last.

At the start, every agent read and flagged. None moved money, published or sent. A person approved every outbound action, escalation thresholds were tiered by financial exposure, and the owner could pause any agent by turning off its schedule. The reconciler has no bank login and ran last because it earns trust last.

That was the starting authority, not the end state. Authority is earned at gates: an agent moves up a level once its flags have been right for long enough and the owner has seen what it would have done. Two have moved so far. The premium tracker now changes prices itself when competitors move, inside a guardrail set around the spot price: nothing sells below spot and nothing buys above it. The eBay helper, which began by drafting replies for approval, now rewrites listing descriptions and fixes categorizations on its own. Every agent that acts logs what it changed and why, and publishes a summary the owner reads instead of approving each change. The gates for the next level, and the governance that goes with them, are the current work.

For a cash-and-metal business, the first job of automation is to leave the books more reconciled than it found them.

Success is judged at the gates. The Yelp plan is measured at day 30, the Google Ads relaunch has a go/no-go decision, the local-ranking heatmap is read monthly, and each is reported against the five scorecard numbers.