What is the Jen Test™?
The Jen Test™ is a scoring rubric for BS claims, fake authority, and inauthentic marketing. It is the productization of the discipline this practice runs on: verify first, write second.
It exists because of one week in July 2026. I was checking marketing copy and hit three separate verification failures in the space of a few days — a statistic that returned nothing on any search engine I tried, a citation whose linked source had nothing to do with the claim attached to it, and links that were either dead or pointed at a page contradicting the quote pulled from it. That week is where I started saying that AI is a confident liar. I spent 13 years in pharmaceutical research, where an unverified figure is a different category of problem than a style preference. The habit came with me.
The three things this test does not do:
- It does not detect AI writing.
- It does not predict rankings.
- It does not grade style.
Copy written by a machine that is true, sourced, specific, and in voice will clear it. Copy written by a person that is none of those will not.
It applies to anything going out under a name someone will be held to — blog posts, vendor drafts, client reports, audit findings, competitor claims, and any AI output before someone repeats it.
Why publish the test you'd be graded by?
Because a rubric its author will not submit to is the thing it exists to catch.
The Jen Test™ scores every author's copy, mine included. That was ruled deliberately, and it has teeth — further down this page you will find it run on a line of my own marketing, which did not clear. My own copy gets a scored entry before it publishes, the same as anyone else's.
The five gates. Any one of them is a failure.
Gates are binary and independent. They are never weighed against strengths elsewhere. A brilliant post with a fabricated citation is a failed post.
G1Fabricated, unverifiable, or misattributed source
Fails if any of these is true of any citation in the piece:
- The source does not exist, or cannot be found by searching its title and publisher.
- The linked page is dead, or redirects to something unrelated.
- The linked source does not contain the claim it is cited for, or contradicts it.
- The citation is a roundup, an agency blog, or a “statistics” listicle citing someone else, rather than the primary source.
- “Studies show,” “research suggests,” or “experts agree” with no study, publisher, or date named.
G1 extended — assertion without attribution is also a breach.
A claim of fact must do one of three things:
- Carry a source. An inline link, or an entry in the end-of-post citations list. What matters is that the claim can be traced; a source at the bottom of a post counts.
- Be marked as the author's own. “In every practice I've audited…” or “My view is…” needs no citation, because the attribution is the author. Use it honestly — a borrowed claim relabelled as personal observation is a worse breach than an unsourced one.
- Qualify as common knowledge. The sky is blue. Patients dislike waiting rooms.
The common-knowledge test. All four have to hold:
- A lay reader already holds it. Industry common knowledge is not common knowledge. “Google weighs page speed” feels obvious to anyone in SEO and is exactly the sort of claim that needs a source. Judge it against the practice owner reading the post, never against the person writing it.
- It contains no number, no direction of change, and no comparison. Any “more,” “rising,” “increasingly,” “most,” or “fewer” is a trend claim, and trend claims are never common knowledge.
- It makes no causal claim. X raises Y. X leads to Y.
- It says nothing about what a third party does. Search engines, an AI model, patients as a population. The behaviour of others is an empirical claim about the world.
The one-line version: would a skeptical practice owner say “says who?” If yes, it needs a source, a self-attribution, or cutting. That question is the gate.
Why the extension exists: a draft I scored asserted three claims as established fact with no source and no “studies show” phrasing at all — evading the gate's original wording while having the identical effect on a reader, and arguably a worse one, because there was no attribution phrase to notice and question. A gate that catches only the announced form of a failure catches careless writers and misses fluent ones — and fluent is what I keep being handed.
G2Invented or precision-faked statistic
Fails on any number asserted rather than sourced:
- A specific figure (“73% of patients…”) with no citation.
- A real figure carried past what its source supports — a national number applied to a metro, a 2019 number stated as current, a survey of one profession applied to another.
- False precision. A decimal place the underlying method cannot produce.
- A single measurement presented as a rate or a baseline. One run is one run.
Directional truths only. “Most practices have no idea whether AI recommends them” is fine. “68% of practices have no idea whether AI recommends them” is a breach unless it is real and sourced.
G3A self-claim that fails the one-click disprove test
Any claim about the business, the client, or the author has to survive one click by a skeptical reader. Credentials, years of experience, client counts, results, awards, “trusted by,” “as seen in.”
The worked example is mine. In July 2026 I corrected my own credential line down to the true figure, because the larger number overstated the pharmaceutical portion of a longer career and would not have survived a LinkedIn check. A month later I cut the hedging word in front of it too. Say the number, or say no number at all.
The zero-client honesty rule. Until case studies exist, proof comes from work done on the business's own properties, labeled as such. Invented or implied client outcomes breach G3.
G4Named IP misused, paraphrased, or borrowed
Fails if the piece:
- Uses a named framework it has no right to use.
- Paraphrases a framework or layer name. Named IP is used verbatim or not at all. “The Authority Stack,” “the four-part framework,” and “Jennie's test” are all breaches. The boundary is verbatim or near-verbatim naming and nothing wider, because a gate that catches its author's own correct usage is not a gate.
- Publishes a definition that is not settled. Naming a framework nobody has agreed a definition for, and defining it on the fly, is a breach. The correct move is to name it, mark it as in build, and say plainly that the definition is pending. Publishing an unsettled definition is the specific habit this test exists to catch.
G5Claims about how AI models decide
Never describe a model's internal process, order of operations, or ranking method. No “the model first checks X, then weighs Y.” No “ChatGPT ranks practices by.” No invented mechanism.
Describe what a model can and cannot associate your practice with, based on what is publicly readable about you. Anything past that is fake authority about a system nobody outside the lab can observe, which is the exact failure this test is named for.
This one applies to my own older material too. A line I wrote in early August — “the 3 Cs are what a model decides” — breaches it, and has been reworded rather than reused.
What gets scored once the gates clear?
Five dimensions, 0, 1, or 2 each. Ten points available. These are the editorial layer — the gates decide whether a piece is honest, and the dimensions decide whether it is any good.
| Dimension | 0 | 1 | 2 | |
|---|---|---|---|---|
| D1 | Specificity to the reader | Generic copy that would read the same for anyone | On-category, but not specific to this buyer | Names procedures, economics, or objections only this reader has |
| D2 | Voice | Banned language present. Sounds like every agency blog | Clean but flat. No point of view | Sounds like a person with an opinion who has done the work |
| D3 | A real claim | Restates the obvious. “Reviews matter” | A correct but familiar point | A specific, defensible position a reader could act on or argue with |
| D4 | Evidence quality | Unsourced assertion throughout | Sourced but thin — one source, or all secondary | Primary sources, named and dated, load-bearing to the argument |
| D5 | Structural honesty | Headline promises what the body doesn't deliver | Delivers, but slowly | Answer up top, structure matches the promise, no padding |
Three verdicts:
- PASS — 8–10 with no dimension at zero. Publishable as received, or after edits under ten minutes that change no claim.
- CONDITIONAL — 5–7, or 8 and above with any single zero. No gate breached, but the scored dimensions fall short.
- FAIL — any gate breached, or 4 and below regardless of how the points are distributed.
A zero blocks a pass and does not rescue a piece from the fail band. Short-form copy — a CTA, a social post, an email subject line — is scored on gates only; there is no structure to be honest about in two lines, and a total out of ten would mean nothing.
Record the verdict before editing anything. “Passed on the first try” means the piece as delivered, before any edit, question, or send-back. Once a draft is fixed, the first-try verdict is gone, and the round-trip is the cost you were trying to measure.
What does it look like run on my own copy?
Here is the most recent entry, in full, from my own scoring log. It is my copy, it did not clear, and it published only after the finding was addressed.
The line, as I wrote it:
The AI Visibility Audit shows you exactly where you stand across the assistants your patients actually use and what's holding you back.
Gates: G2, G3, G4, and G5 clear. G1 is flagged.
The finding. That phrase presupposes a claim about a population — that your patients use AI assistants to find a dentist. Under the common-knowledge test above, that is a claim about what a third party does, and a skeptical practice owner can answer it with “says who? Mine don't.” That claim is the entire premise of my business, which is exactly why it belongs somewhere it can carry a source, instead of riding along as a presupposition inside a call to action.
Verdict: CONDITIONAL — clears on one line change.
What it became:
The AI Visibility Audit shows you exactly where you stand in AI search — and what's holding you back.
An earlier version of that same line carried a specific count of AI platforms. I cut it, because it was a number I could not commit to honouring precisely at delivery, and a number I cannot stand behind is a G2 exposure on my own storefront.
That is what the test looks like when it lands on the person who wrote it. Two findings, one line of copy, published after the fix rather than before.
How do I run this on the people I'm paying?
You do not need the full rubric to start. Four of the five gates translate directly into questions you can ask an agency, and none of them is hostile — they ask for the record, which anyone doing the work already has.
Copy this, change the name, send it.
Subject: Checking what's being measured
Hi [name],
Before we go further I want to make sure I understand what's being measured. Four things:
- For any claim about how my practice shows up in AI search — what was the exact question asked, in the words a patient would use?
- Was it run more than once, and in more than one assistant? I'd like to know how much the answer moves between runs.
- For any statistic in a report or a proposal — can you point me to the original source, with a date?
- Can you write down where my practice stands today, dated, so we have a starting point to compare against later?
I'd just like the record so I can track it over time.
Thanks,
[your name]
Questions 1 and 3 are G1. Question 2 is G2 — one run is one run. Question 4 is the baseline, and it is the one most likely to go unanswered.
Separately: anyone who claims to know how the models rank, or offers to feed them “training signals,” has breached G5 before you've asked anything. For the plain-language version of what does and doesn't work here, can you recommend ways to increase ChatGPT recommendations for dentists ethically? is answered on the FAQ.
Where this comes from
If your marketing person never asks “wait, is that actually true?” — you're paying for slop with your name on it.
The Jen Test™ is one of the named frameworks behind this practice. The others are on the Frameworks page. The one everything on the visibility side sits on top of is The AI Authority Stack™, and the condition it exists to resolve is The Invisible Practice Problem.