What you can trust

Every loop ships with four things

If a loop on this site is missing any of these, it is a bug. Tell us.

1. A goal prompt. What the agent is told to do, in full. No hidden system prompt, no paid tier that unlocks the real one.
2. A verifier prompt. A second agent that grades the first one's work and can fail it. You run it in a fresh session so it cannot see the maker's reasoning.
3. A stopping condition. Something checkable, so the loop ends on a state rather than on a vibe. "Every page has a before and after score" is a stopping condition. "Make it better" is not.
4. A receipt from a real run. What it changed, what it cost, and what it did not prove. A loop with no receipt does not appear on this site at all.

How to read a receipt

The field that matters most is the one most sites omit: attribution limits, which says what the run does not prove. If a receipt claims a result without naming what could otherwise explain it, do not trust the result.

Every number on a receipt names its source, its collection date, and its comparison window. Nothing is edited in place: a correction appends a dated line with the old value, the new value, and the reason. If a number here ever changes silently, that is a bug worth telling us about.

What the lab will not claim

No citation counts, on any engine, for either property. No engine is configured, so there is no measurement, and the scoreboard says "not yet measured" rather than showing a zero. Zero is a measurement; none was made.

No traffic, ranking or pipeline promises. No causal language: receipts say "observed after" or "consistent with", never "because of". No paid placement, no paid verdicts, no affiliate links, and no product shown that cannot be delivered.

Who is behind it

VibeMarketers is an open lab by SuperMarketers. VibeMarketers tests the patterns. SuperMarketers turns them into an AI-native marketing operation built around your company. If it explains, tests, or packages a reusable loop, it lives here. If it diagnoses a company or sells implementation, it lives there. The lab never diagnoses a company; that is the SuperMarketers assessment.

Gen is the operator and the byline. The agents are the team, and every human touch is counted in the log. Everything below is the governance that keeps the above honest: the charter the lab pre-registered, the gates, and the confounders it has disclosed against itself.

Experiment 001, pre-registered

One dated experiment is the master charter. Every loop's receipts are sub-experiments under it. The charter below is fixed; changing it starts a new experiment.

The day 90 verdict rests on Perplexity alone unless a second programmatic engine is added. Perplexity is the easiest engine to get cited on and the least commercially relevant, and the charter says so.

The three gates

Three gates, nothing else. Each is one PR approval or one Beehiiv send, and each approval is one line in the human-touch log.

1. Publishing a new loop page.
2. Sending The Lab Report.
3. Any change to supermarketers.ai.

Proposed gate service level: a loop that finishes and waits more than two days for an approval counts as a system error on the scoreboard.

What a touch is

A touch is one approval, one edit to a drafted artifact, or one manual intervention in a run. Reading is not a touch. Writing issue 1 by hand is counted as touches at one per section. The cap for the 90 days is 40 touches. Pre-launch design work, the setup before day 0, is excluded from the cap, and the log says which entries are which.

What the runtime may do

• Run any loop in the registry on the lab's own properties.
• Update loop pages, receipts, scoreboard, log, sitemap, and llms.txt from structured records.
• Open a PR.
• Draft a Lab Report.
• Run a requested free run within budget.
• Pause itself and write an exception.

What the runtime may not do

• Invent a loop that is not in the registry.
• Change the query set, the charter, or the gates.
• Merge to main.
• Send email.
• Post to Gen's accounts.
• Contact anyone.
• Spend above budget.
• Create accounts, change secrets, DNS, or billing.
• Write to any domain it does not own.
• Publish a number without a source.

Flags, budget, exceptions

Six flags, all off by default, and one command pauses everything. A budget controller sets per-run, daily, and monthly caps with a warning threshold and a hard stop, and every run records its cost by loop. A failed checker, a missing source, a budget warning, or a run with no valid work produces an exception file in the repo and publishes nothing. The lab never fabricates work to satisfy the weekly cadence. A quiet week says so on the scoreboard.

Maker and checker

Every loop runs as a maker and a checker in separate agents. The checker verifies against real command output, the voice scanner, and the evidence rules. Before launch the checker must fail one deliberately bad run, and it did; the failure is in the repo's loop state file.

Evidence rules

Every number on a receipt names its source, collection date, and comparison window. Every sentence that makes a claim is one of: observation, calculation, interpretation, hypothesis, result. Causal language is banned unless the charter allows it; receipts say "observed after" or "consistent with." Mock data appears only in tests and fixtures, and is labeled. A loop page with no receipt does not publish.

Corrections policy

A receipt is never edited in place. A correction appends a dated line with the previous value, the new value, and the reason. The scoreboard and the log follow the same rule. There are no silent corrections.

Confounders, disclosed

• Engine index changes can move citations with no intervention from the lab.
• Gen's LinkedIn posting drives traffic and may drive citations. Those posts are counted as distribution touches, separate from the gate touches.
• The query set is written by the same people who write the pages. Writing pages to the fixed queries is the intervention, by design, so the set is frozen and versioned and every receipt that cites a citation count carries this line.
• The proof is self-referential. Getting a niche site cited for queries its own authors wrote is easier than getting a SaaS cited for a competitive category. supermarketers.ai runs as a second property so the gap can be seen.

The free run service level

A free run is queued, not immediate. The lab commits to running a queued request within the next scheduled window and emailing the receipt, a gate of 48 hours from the request. Requests are rate limited by domain and by IP, and capped by the monthly budget. When the budget is reached the queue holds and the requester is told so. Free runs are closed until the Worker passes its abuse tests.

Vocabulary

The words the lab uses for its own parts, moved here from the homepage so the reader who wants to argue with the method finds them in one place.

Capture

Reading the company into the Brain, so every later loop cites positioning instead of inventing it.

Build

Turning what the Brain knows into pages, answers, and drafts.

Ship

Putting the work where a reader or an engine can reach it.

Improve

Measuring what moved, rewriting what slipped, and running it again.

The brain

One directory that is the source of truth every loop reads: positioning, ICP, buyer questions, claims and proof, competitive intel, content history, voice.

The voice

The rules a draft has to pass before it publishes, enforced by a scanner, not a preference.

The gates

The points where a human has to approve before anything leaves the lab. A PR approval and a send approval.

The ledger

Every human touch, one line each, dated. It is how the 40 touch cap is counted and it is public.

What the lab is not for

Agencies, tool shoppers, engineers. No paid placement, no paid verdicts, no affiliate links, no product shown that cannot be delivered. The lab never diagnoses a company; that is the SuperMarketers assessment.