The first 30 days of customer-zero: Merkava on Merkava

Before Merkava sold to anyone else, it ran the full team on its own venture for 30 days. Here is what shipped, what broke, and what changed about the product as a result.

The first 30 days of any product is the most informative. For an AI executive team selling itself, the only honest first 30 days is running the team on the company's own venture before anyone else hires it.

This is what shipped, what didn't, and what we changed as a result.

What was in place day zero

Withmerkava.com existed as a marketing site with about 95 pages. The product (the cockpit at app.withmerkava.com) was built but had not been hired by anything. The merkava venture inside emmett (our internal handle for Merkava Core) was active but had no executive team running on it.

Day zero, we hired the full C Suite for the merkava venture: GROWTH, SALES, OPS, TECH (FINANCE auto-activates at $3K MRR).

Week one — the audit + fixes

Beacon audited withmerkava.com and found 28 gaps. High-severity: missing /llms.txt, missing WebSite/Organization schema, missing ai-content-policy meta, missing FAQPage on /pricing. Medium: incomplete OG on most pages, missing canonical on 12 pages, sitemap was 6 weeks stale.

Beacon opened 16 PRs. We merged 14, sent 2 back for narrower scope. Total founder time on the merge queue: about 90 minutes across the week.

Week two — content + comparisons

Quillsly drafted 4 comparison pages (Zapier, n8n, Lindy, Make). The drafts were close to publishable; we made voice edits on 2, accepted 2 as-is. SAM (Quillsly's distribution feature) queued the comparison pages for LinkedIn distribution but stayed in stealth mode by default.

Webster generated OG images for every page that was missing one. ~40 pages got new OG cards in one batch.

Week three — operating cadence

OPS started running the weekly digest. The first one was rough — too many entries, no priority — and we tightened the brief. By week three's digest, the format was useful: 5-7 things shipped, 1-2 things at risk, 0 things blocked.

Centerline (operations coach) caught the first drift signal: GROWTH had 3 active campaigns but no measurement on conversion. We added the measurement layer + a kill rule for non-performing campaigns.

Week four — the public surfaces

Thread 6 (public Drive output URLs) shipped in week 4 — a generic /work/:drive/:tenant/:slug render layer. Customer-zero seeds a demo artifact on every Railway boot, so the route is observably alive at app.withmerkava.com/work/quillsly/merkava/welcome-to-thread-6.

The directory submission queue was drained via cowork in a 90-minute session: 30 directories, ~22 successful submissions, ~8 blocked (captcha, account-required).

What changed about the product

Three things from this 30-day run made it back into the customer-facing product:

Trust ladder defaults. The first PR Beacon opened was for an ai-content-policy meta change that we wanted to review. The second was the same kind of change. By the fifth, we trusted it enough to auto-merge low-impact fixes. We added a "trust ladder" UI with three rungs (review-each → auto-low-impact → auto-all) so customers can opt into the same progression on their schedule.

Stealth mode default-off for the first hire. We had stealth on by default for any new venture. After the dogfood run we realized that was the wrong default — stealth should be opt-in by category (LinkedIn off until you complete OAuth, blog publishing on by default). Customers don't want their first hire to silently NOT publish.

The /work mechanic. Watching customer-zero's dogfood URL be the only thing populating /work was the moment we realized this surface needs to be productized for Cohort + Affiliation specifically — those are the Drives where public URLs are intrinsic to the operator's job-to-be-done.

What did not work

The first SAM scribe runs failed with FK constraint errors. The personas table had been seeded under the wrong venture_id. We patched it manually for week one but the bug stayed open for 48 hours longer than it should have.

LinkedIn OAuth was never completed in the 30-day window. Posts queued correctly but never published. We removed LinkedIn cadence from the success metrics for the dogfood run.

What this earned

By day 30: 192 pages on the marketing site (vs 95), 99% schema coverage, 8 published articles, 27 comparison pages, daily auto-regen via GHA, weekly audit sweeps, the referral loop wired with Stripe credits, the /work public render layer live.

That's the run. Customer one was Merkava itself, and the dogfood made the product better.