A few weeks ago I read a systematic review of 9,249 SEO, AEO and GEO documents. Its headline finding was that 602 of 769 SEO tactics with a measured outcome had no controlled test behind them. Its second finding was that AI visibility had among the thinnest evidence of any topic in the field.
I sell AI visibility. So the obvious next move was to find out how much of what I had published would survive that standard.
Not enough of it.
What I found
Across six pages I found 23 distinct claims I could not back with a source — timelines I had no data for, audit-frequency figures I had not earned, causal language where I only had a mechanism, three unsupported superlatives, and one statement that was simply false.
The timelines
The single most common problem. Versions of "most brands see measurable citation improvements within 30 to 90 days" appeared across several posts, including inside the FAQ markup that AI engines read directly. It sounds like the kind of thing an experienced practitioner would know. I don't know it. Nobody does — there is no controlled public evidence establishing how long any GEO change takes to move citation, and anyone quoting you a firm window is guessing.
The audit-frequency claims
"In 80% of GEO audits the content is fine." "In the large majority of audits I run." These are the ones I'm least comfortable with, because they borrow the authority of a dataset that doesn't exist. My observations from production work are real. The percentages attached to them were not.
The causal language
FAQPage schema "dramatically increases" citation probability. Author bylines make a page "materially more citable." Every one of those is a reasonable hypothesis from how retrieval systems work. Not one of them has been measured. The fix was not to delete the recommendations — it was to say what they actually are.
The false one
A post said several major AI engines had "confirmed they use" llms.txt to prioritize crawling. That isn't established, and Google has said publicly it doesn't use it. It was a confident sentence that nobody checked, and it was wrong.
Some of those lines were drafted with AI and approved by me. The prose was good. That was the problem — fluent writing is exactly what makes an unsupported claim look like expertise.
Beyond the blog posts
Three more things turned up that were arguably worse than the copy.
A cold email template claimed access I couldn't have. It opened by saying I'd pulled a prospect's server logs to see how AI crawlers behaved on their site. You can't see another company's server logs from outside. A technical buyer would have stopped reading at the first line. The rewrite uses only what's genuinely observable from outside — robots.txt, pages rendered without JavaScript, externally measured response times — and says plainly that the log-level picture requires their data.
Five pages had FAQ schema that didn't match the page. The structured data and the visible FAQ had drifted apart — different wording, and in one case questions in the markup that didn't appear on the page at all. That's non-compliant with Google's structured data guidelines. It would also have quietly broken the test I describe below, because you can't isolate the effect of schema when schema and page are saying different things.
A check I wrote caught itself. The script that verifies the test treatment flagged every page as changed when none had been — a whitespace difference, not a content one. It failed safe and refused to run, which is what it was built to do. But it was a small, live demonstration of something the review makes a point of: every check needs testing against real output, including the checks.
What I changed
| Before | After |
|---|---|
| "Most brands see results within 30–90 days" | No timeline claimed. Freeze a prompt set before changing anything and measure your own site. |
| "In 80% of GEO audits…" | Stated as an observation from my own work, without an invented percentage |
| "Dramatically increases citation probability" | Stated as a hypothesis, with the reasoning behind it |
| No indication of evidence strength | Each post opens with a short note on what kind of evidence it rests on |
| Stats quoted without context | Every number kept with its source and denominator; key posts carry a "how to cite this" block |
| Recommending tactics nobody had tested | Testing one, publicly, with the plan published first |
That last row is the one that matters. The honest response to "there is almost no controlled evidence for GEO" isn't just softer language. It's producing some.
Test 01 is a pre-registered, randomized test of whether FAQPage schema affects AI citation — fourteen pages on sites I own, split into seven matched pairs, with the schema randomly removed from one page in each pair. The design, the random seed, the 42 prompts, the decision rule and the scoring code were all published before a single data point was collected. The result gets published in December whatever it says, including if it undercuts advice I've been giving.
It's a small test. Its own pre-registration says it can find a large effect and can't reliably find a small one. That's stated up front on purpose, so a null result can't be dressed up afterward as anything other than what it is.
If you publish in this space
Run the same audit on your own work. It takes an afternoon. Search your published content for fixed timelines, percentages about your clients or audits, causal verbs attached to tactics, and superlatives. For each hit, ask whether you could point someone to the source. If you can't, either find one or change the sentence.
It's uncomfortable. It's also the only version of this job that holds up when a client asks where a number came from.
Frequently asked questions
Scan for specific linguistic patterns rather than reading for tone: fixed timelines with no source, frequency claims about audits or clients, causal verbs like increases or drives attached to tactics, superlatives like highest-leverage, and factual assertions about what platforms do. Then check each hit against a source. Pattern scanning finds the candidates; a person has to decide which ones are real problems.
Yes. Google's structured data guidelines require markup to reflect content that is visible to users. If the questions or answers in your FAQPage JSON-LD differ from what is on the page, the markup is non-compliant, and any test of whether the schema affects AI citation is also compromised, because schema and page are no longer saying the same thing.
Using AI to draft is not the problem. Approving a confident draft without checking each claim against a source is. AI drafting tends to produce plausible, fluent assertions — timelines, percentages, causal claims — that read like expertise. The review step has to check the claims, not the prose.
Sources
- Ćorluka, K. (2026). The SEO Meta Study. Ovisy. ovisy.hr/blog/seo-meta-study
- Google Search Central. General structured data guidelines.
- Summers, J. C. (2026). Test 01 pre-registration. sablesearch.com/research/faq-schema-test/