How to Verify a GEO Citation Study Before You Believe It

Open any "Generative Engine Optimization" (GEO) or "Answer Engine Optimization" (AEO) vendor's blog and you'll find the same shape of content: a proprietary study, a precise-sounding statistic, and a service that fixes the problem the statistic describes. "852 pages analyzed." "78% cross-engine citation rate." "6x more citations at DA 40+."

Some of these numbers might be accurate. The point of this post isn't to declare them false — it's to give you a method for checking, in about ten minutes, before you build a strategy or write a claim on top of one. We ran this method against four real examples while researching our own roadmap. Here's what we found, and how you can reproduce it.

Step 1: Is the "independent study" actually independent?

Res, an AI-visibility agency, publishes a piece titled "Six Structural Features Separate AI-Cited B2B Articles from Invisible Ones", citing three outside studies that supposedly confirm its findings. One of them is credited to "Kumar and Palkhouski, 2025."

Search for that paper and you'll find it on arXiv: "AI Answer Engine Citation Behavior: An Empirical Analysis of the GEO-16 Framework". Look at the author affiliation line printed on the paper itself: alongside UC Berkeley, both authors list "Wrodium Research."

Search "Wrodium" and you land on its Crunchbase profile: an AI-search-visibility and content-maintenance platform. One author's personal site describes him as "Co-founder & CTO of Wrodium," building "the picks-and-shovels for the AI search gold rush."

So the "independent study" backing Res's claims was written by the founders of a competing GEO vendor, about the exact mechanism their own product sells. That's not proof the data is wrong. It's proof you shouldn't count it as independent corroboration.

How to check it yourself: search the author's name plus the company name they list in the paper's own affiliation line. Two minutes, every time.

Step 2: Is it actually peer-reviewed?

The same paper's "Data availability" section states the dataset will be released "upon acceptance to a peer-reviewed venue." At the time of writing, that acceptance hadn't happened. arXiv is a preprint repository — anyone can post there. It is not peer review.

We found at least one third-party blog citing this paper as "peer-reviewed research." It isn't, by the paper's own admission.

How to check it yourself: search the paper title plus "peer review" or "accepted." If the only venue you find is arXiv, and the paper doesn't name a journal or conference that accepted it, treat it as an unreviewed preprint — regardless of how it's cited elsewhere.

Step 3: Is the second "independent" source independent either?

We went looking for a genuinely unaffiliated confirmation and were pointed to a second paper, "What Drives Citations in Production Large Language Models?" — which explicitly critiques Kumar & Palkhouski for lacking domain-level controls, and argues most of the standard AEO checklist (FAQ blocks, schema, Core Web Vitals) washes out once you control for brand identity.

That sounds like the independent check we needed — until you look at the authors' affiliation: Discovered Labs, whose own homepage describes it as "the first AEO/GEO agency that helps B2B companies dominate AI search results." The paper's own findings are also published as marketing content on the agency's site.

Two competing GEO vendors, two competing frameworks, two self-published studies that each happen to support the product each company sells. Neither is proof the other is wrong. Neither is proof either is right, on its own.

How to check it yourself: don't stop at one paper. If a second source seems to confirm or contradict the first, run the same affiliation check on it before treating it as a tiebreaker.

Step 4: Are the "sources" citing each other?

A third example, from an agency called Synapse Edge, backs its Perplexity SEO guide with statistics attributed to "Otterly.AI," "LeadWalnut," and "Harbor SEO." Look each of those up: all three are themselves GEO/AEO vendors selling comparable services. The article ends with a link to Synapse Edge's own paid audit tool.

This is what's sometimes called citation laundering: vendor A's number gets cited by vendor B as an outside source, which gets cited by vendor C, until the number looks like industry consensus — when it's actually the same small cluster of companies citing one another.

How to check it yourself: for every "source" a vendor cites, ask what that source sells. If the answer is "the same category of service," the citation isn't external corroboration — it's the same market talking to itself.

Step 5: Is there a source with a different incentive?

The cleanest counterweight we found in this whole exercise wasn't a GEO vendor at all. It's Google's own Search Central documentation on AI Overviews and AI Mode:

"There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary."

Google isn't selling a GEO audit. Its incentive is different — it wants your site crawlable so its own systems work — but that's a materially different incentive than "sell you a subscription to fix a problem we just quantified." That doesn't make Google's page infallible, and it only speaks for Google's own AI features, not for ChatGPT, Claude, or Perplexity's separate retrieval pipelines. But it's the one source in this whole exercise with no service to sell you on the back of the claim.

How to check it yourself: for any GEO claim, ask whether a platform's own developer documentation says anything on the topic. It usually has a different, and often more boring, answer.

A disclosure, because the method demands it

We build Agentabile, a scoring tool for how readable a site is to AI agents. We have a commercial interest in this market too. This post exists partly to build trust in that work — pointing out how a competing category oversells its research is, among other things, good marketing for a company that doesn't want to.

We think the honest way to handle that isn't to pretend the interest doesn't exist. It's to say so directly, and to build the piece so you don't have to take our word for anything — every link above is public, and you can run the same five checks yourself in the time it takes to read this sentence twice.

A case where this stopped being about content marketing

Everything above concerns commercial overclaiming — annoying, occasionally misleading, but low-stakes. For a case where a documented influence campaign, filed under the U.S. Foreign Agents Registration Act, allegedly used similar techniques to shape what AI models say about a live geopolitical conflict, see the underlying FARA filings here: Piro filing, document trove.