A Marketing Health Check You Can Run on Any Company

Nobody runs a marketing health check on themselves, because there is nothing honest to run. Ask an owner whether their marketing is working and you get a number somebody else produced. Impressions. Posts published. Blogs written. Leads, sometimes, counted by the agency that is paid per lead.

Ask a buyer diligencing that same company and it is worse. The marketing is described as a strength, in a deck, by the seller, and there is no instrument on the table to test it.

Activity is easy to count, which is exactly why it gets reported. Results are hard to count, and invisibility is hardest of all, because nothing shows up in a dashboard to tell you the customer never found you.

#Problem

Three problems are stacked here.

The first is that the scoreboard is kept by the people being graded. The agency writes the report. The in-house team picks the dashboard. Neither is lying, usually. They are reporting the metrics that make the work legible, and those are metrics of effort.

The second is that nobody outside the company can check anything. Analytics is behind a login. The ad account is behind the agency. The buyer gets a data room and a statement that digital is performing well. Every conventional audit begins with access, and access is the one thing an operator auditing their own agency, an agency pitching a prospect, and a buyer before signing an NDA all do not have.

The third is newer. Answer engines now sit between the customer and the business. A site can be perfectly readable to a human and invisible to the thing summarising the category. Nothing in a monthly report captures whether a model can find you, parse you, attribute you, and cite you, because the agency reporting on you does not measure it either.

The trap: you keep paying for a service nobody has an instrument to test, and the first honest reading arrives when revenue drops.

#Solution

marketing-healthcheck is a free, MIT-licensed Claude skill that grades a company’s digital marketing from the outside.

What it is not: a crawler that dumps four hundred warnings, a rank-tracker subscription, or anything that needs a login. It never touches an analytics property, an ad account, or a page behind authentication.

What it is: an instrument that judges a business on signals anyone can observe from outside. What ranks. What the profile looks like. What the site publishes. What the page loads. What tags fire. What DNS says. Twenty checks in six groups, scored 2 strong, 1 partial, 0 weak, null not run, rolled into one letter grade with the evidence attached.

The six groups map to six questions a customer’s experience actually turns on. Found is whether anyone can reach you: local pack rank at each location, organic visibility, how much of it is just people typing your name, whether an answer engine can read you, and whether your own pages are even indexed. Trusted is the review corpus and the profile. Saying is whether anything gets published, whether the pages are distinguishable from each other, and whether a human name is attached. Working is speed, mobile, HTTPS and email authentication. Measured is whether analytics is installed and configured, whether conversions fire, and whether the ad tags match the advertising the business says it does. Owned is the uncomfortable one: who controls the domain, the DNS and the tags, and which vendors are authorised to act as you.

That last group is why buyers care. A marketing programme can score well on every visible check and still be a liability if the agency owns the domain, the analytics property and the ad account, because none of it transfers at close.

The same run reads as an AI-readiness check on the marketing surface, because the signals an answer engine needs are the signals being tested: structured data, authorship, page uniqueness, llms.txt, indexation.

#Value

The agency report

Activity, counted by the party being evaluated.

A crawler audit

Four hundred warnings, no verdict, no priority.

A rank tracker

One group of the six, on a subscription.

marketing-healthcheck

One grade, one fix first, evidence and a re-runnable command behind every cell. Free for twelve of the twenty checks.

Three artifacts come out, in this order: a simple view with the grade, the one-line verdict and the fix to make first, then the matrix of every check against every subject, then a method note saying what was run, what was not, and what the result cannot prove.

That last one matters more than it sounds. Public signal shows visibility, hygiene and delivery. It does not show conversion rate, lead quality, close rate, cost per acquisition, or whether the business is profitable. The report says so before anyone has to ask.

#How it works

You say the domain. Claude does the rest, inside Claude Code or the Skills UI, the same way the second-brain skill turns a folder of documents into something you can question.

1. 01
   Scope it. Domain, category, and every physical location with its own address. A business with six locations is six tests, not one.
2. 02
   Run the free tier. Twelve checks from HTTP and DNS: email authentication, conversion tracking, indexation, page uniqueness, publishing cadence, tag and vendor inventory. Zero cost, minutes per subject.
3. 03
   Decide on paid work. Rank grids and review corpora are the only checks that cost money. The skill prints the plan and the estimate and refuses to spend without a budget flag and a yes.
4. 04
   Score and report. Rubrics from checks.json produce scorecard.md and scorecard.json. Simple view first, matrix second, method note last.

Three rules do most of the work of keeping the result honest.

Test at the business’s own coordinates. Local rank changes over a few hundred metres. Run the category term from the city centre when the business sits ten minutes out and a good account grades terribly. An unfair audit is worthless the moment the other side checks it, which is the whole point of handing them the command.

Never score a check you did not run. A null is honest. A 0 is a claim, and every claim gets argued with. The generated verdict even ships marked [DRAFT - rewrite], because a mechanical count of weak and strong checks reads the same for every company with the same distribution, and a sentence that could describe anyone describes no one.

Know the arithmetic before you spend. Rank grids bill per grid point per keyword, and they dominate the bill by an order of magnitude. A 5x5 grid across six locations for three keywords is 450 points. Going to 7x7 costs 96 percent more and buys resolution that almost never changes the verdict.

Everything reads public pages the way a browser does, throttled, identifying itself with a contact address you set. No login, no paywall, no challenge. Every response is cached under out/raw/, so a disputed cell can be checked against the bytes it was scored from without refetching anything.

#Why it works

The instrument is useful precisely because the other side can pick it up.

An agency can argue with a feeling. It cannot argue with a DMARC record that does not exist, a conversion event that never fires, or ten location pages that are the same page with the town swapped. Those are uniform failures, they are checkable in thirty seconds, and the report leads with them for that reason. Mixed results, where half the markets rank and half do not, have real explanations and belong further down, after the unarguable part has landed.

Scoring per location rather than per company is the other half of it. A six-location business judged on its best market is a brochure, and judged on its worst is a hatchet job. Score every market, then roll up, and the pattern does the arguing for you: a gap present in every location is a systemic failure nobody defends, while a gap in three of six is a real question about those three markets. Averaging the two together loses both findings.

A score you can re-run is the only kind that survives the meeting where it gets argued about.

The grade is not the product. The evidence under it is.