We Found Fabricated Clinical Citations on a Medspa's Website — Here's How AI-Generated Content Goes Wrong

The pattern is almost always the same

A medspa's marketing team, or the agency they hired, sits down to build out or refresh treatment content. There are twenty to thirty procedure pages to write. Each one needs a description, a bulleted list of benefits, and — if the practice wants to look serious — a nod to clinical evidence. So the writer opens a language model and asks for a treatment page on, say, microneedling with citations to peer-reviewed studies.

The model produces a fluent page. There's an introduction, an explanation of the mechanism, a bulleted list of benefits, and a paragraph at the bottom that reads something like "Clinical studies have demonstrated efficacy for the treatment of atrophic acne scars (Fernández-Sánchez et al., 2019, DOI: 10.1111/jcd.12934)." The DOI is formatted correctly. The author name looks plausible. The journal abbreviation is real.

The writer copies the page into the CMS. Nobody checks the citation. Six months later the practice owner is wondering why organic traffic is flat and consultation bookings are down.

What we actually found

On a recent audit of a Phoenix-area medspa's site — seven treatment pages, roughly 5,000 words of clinical content — we found eleven citations that failed verification. The breakdown:

• Seven non-existent DOIs. Formatted correctly. Resolved to nothing when queried against Crossref. The papers do not exist.
• Three misattributed studies. Real papers, real DOIs — but the finding described on the page was not what the paper concluded. In one case the citation supported a claim opposite to the paper's actual results.
• One PubMed ID that resolved to a cardiology paper on beta-blocker adherence, cited on a page about hyaluronic acid filler longevity.

None of these had been flagged by anyone — not by the copywriter, the agency, the practice's marketing lead, or the four different SEO auditors who had previously looked at the site. Every one of them had done link-checks and technical audits. Nobody had opened a DOI resolver.

Why the model does this

Language models are trained to produce statistically plausible text. A citation-shaped string is text. When a model is asked to cite a study, it generates something that looks like a citation — the right author-density, the right DOI format, the right journal name — because that is what the surrounding context calls for. Whether the citation refers to a real paper is a separate question the model has no reliable mechanism to answer.

This is not a bug that a better prompt fixes. It's a structural property of how these systems generate text. Some newer models are better than others at grounding claims to real sources when explicitly wired to a search or retrieval layer. Almost nobody producing marketing content is doing that. The output looks correct. It reads confidently. It ships.

Why nobody catches it

Four reasons, in order of frequency:

1. The person writing isn't clinical. A marketing writer or agency copywriter reads "Fernández-Sánchez et al., 2019" the same way they read a footnote in a magazine — as decorative authority. They aren't in the habit of resolving DOIs.

2. The person reviewing isn't verifying. If the practice owner is a physician, they might catch a factual error inside the page. They almost never audit the citations, because auditing citations was never part of the review process for marketing copy.

3. Traditional SEO audits don't look here. An SEO audit checks title tags, meta descriptions, internal links, page speed, and schema. It does not verify medical citations. That's not what SEO auditors are trained to do.

4. Google doesn't tell you. There is no crawler warning, no Search Console message, no visible penalty that reads "we noticed you cited a paper that doesn't exist." The damage happens quietly through E-E-A-T signals, downgraded rankings on YMYL content, and — increasingly — reduced inclusion in AI-generated overviews that scrutinize source quality more aggressively than classical ranking did.

Why this matters beyond SEO

Fabricated clinical citations aren't just a search problem. They're a credibility problem with regulators, with state medical boards, and with plaintiffs' attorneys in the event of an adverse outcome. A page that claims a treatment is "supported by peer-reviewed literature (DOI: [nonexistent])" is a page that will read very badly in a courtroom.

The FTC's guidance on health and wellness marketing is unambiguous: claims must be substantiated. A fabricated citation is not substantiation — it's the appearance of substantiation, which is worse than none at all, because it implies the practice believed it had evidence when it did not. State medical boards have taken action against practitioners for misleading online marketing before, and the introduction of AI-generated content is not going to make regulators more lenient.

How to check your own site

You don't need an agency to run a first pass. For any citation on your site with a DOI, paste the DOI into doi.org and see if it resolves. For any citation with a PubMed ID, plug it into PubMed and confirm the paper exists and that its abstract supports what your page says it supports. It takes about 30 seconds per citation. If you have twenty-five citations across your site, that's twelve minutes of work — and one uncomfortable morning if the numbers come back bad.

If it turns out you have a problem, the remediation isn't complicated: remove the fabricated citations, replace them with real ones (or with hedged claims that don't require them), and add a review step to your content workflow that verifies every citation before publish. The harder question is how many other pages on your site were generated the same way and never audited.

The wider signal

Fabricated citations are the most visible symptom of a deeper problem: AI-generated content is being shipped into regulated verticals without the review layer those verticals require. Medspa content is squarely inside Google's YMYL ("your money or your life") category. So is financial services content. So is legal content. So is anything to do with medication, treatment, diagnosis, or clinical claims.

The businesses winning in these verticals are the ones treating AI as a first-draft tool with a real reviewer at the other end — not as an autonomous publisher. The businesses losing are the ones that let a generative model become the last person to touch the page before it went live.