Someone sold you something, or tried to, and part of you doesn't believe it. Maybe it was a vendor promising your brand would be "guaranteed" to appear in ChatGPT answers within thirty days. Maybe it was a blog post insisting that adding one small file to your site would fix your AI visibility overnight. Maybe it was a dashboard showing your "share of model" to two decimal places, presented with the confidence of a number that had actually been measured rather than estimated.
The cost of not knowing which of these claims are real is not abstract. Teams that believe the wrong ones waste budget on fixes that do nothing, sign contracts promising outcomes nobody can actually deliver, and, worse, sometimes conclude that GEO itself doesn't work when what actually failed was a specific tactic that was oversold from the start. Sorting the real from the false here protects both your budget and your ability to tell, later, whether this discipline is working for you at all.
Here is what's actually true, what's actually false, and the one belief in this list that genuinely isn't settled yet.
The Claim: An llms.txt File Fixes Your AI Visibility
This is one of the most widely repeated claims in the entire category, and it deserves to be addressed directly rather than left to linger. An llms.txt file is a proposed standard, modeled loosely on robots.txt, meant to give AI crawlers a structured summary of a site's content. It is not a mechanism that any major AI engine has confirmed it actually reads and acts on in a way that changes what gets surfaced or cited.
Adding one costs almost nothing and won't hurt you, so there's little harm in having one as a low-effort, low-priority addition. But treating it as a fix, something that meaningfully changes your visibility once it's added, is not supported by anything the major AI companies have said about how their systems actually work. If a vendor's pitch centers heavily on llms.txt implementation as a core deliverable, that's a signal the pitch is built around an easy, visible-looking task rather than the harder, less tidy work that actually moves visibility: entity clarity, genuine third-party authority, and specific, citable content.
The Claim: Schema Markup Alone Is Sufficient
Schema markup, structured data that helps machines parse what a page is about, is real, useful, and worth implementing correctly. It is not, on its own, sufficient to produce meaningful AI visibility, and any pitch implying that a schema audit and fix is the whole solution is selling you a component as though it were the entire system.
Here's the part that's true: correct schema removes friction. It helps an AI system parse your content faster and with less ambiguity. Here's the part that's false: schema markup cannot manufacture authority or specificity that isn't already present in the underlying content. A generic, unspecific page with perfect schema markup is still a generic, unspecific page. Think of schema as clearing the path for genuinely good content to be understood correctly, not as a substitute for that content existing in the first place.
The Claim: Guaranteed-Citation Services
Any service promising a guaranteed citation or a guaranteed ranking on a specific AI engine, particularly ChatGPT or Claude, by a specific date, is promising something outside its own control. These models update their core knowledge on release cycles set by the companies that build them, on timelines no outside vendor has visibility into, let alone influence over. Nobody selling a GEO service, including highly credible ones doing genuinely good work, can guarantee what a training update will or won't include.
What can honestly be offered instead: a rigorous, evidence-based process, measurable improvement on the engines that retrieve live content in real time (which respond faster and more predictably to real work), and steady groundwork that improves your odds on the slower-moving engines without pretending those odds can be guaranteed. If a contract or pitch uses the word "guaranteed" attached to a specific AI engine's output, that's the clearest single red flag in this entire list, worth walking away from regardless of how credible the rest of the pitch sounds.
Watch, too, for guarantees that sound narrower and more reasonable but function the same way. "Guaranteed improvement in your visibility score" sounds safer than "guaranteed ChatGPT citation," but if the scoring methodology behind it isn't disclosed and isn't yours to independently verify, it's possible to satisfy that guarantee on paper without anything real having changed for your actual buyers. A guarantee is only meaningful if it's tied to a metric you can check yourself, using a method you understand and could, if you wanted to, replicate independently.
The Claim: A Single Vendor Can Track Every Engine That Matters With Equal Accuracy
Vendors selling comprehensive, all-engine tracking sometimes imply equal depth and reliability across ChatGPT, Claude, Gemini, Perplexity, and every other engine you might care about. In practice, tracking accuracy and access vary meaningfully by engine, since each one has different rate limits, different API access terms, and different quirks in how consistently it returns results for the same query run twice in a row.
This doesn't make multi-engine tracking useless, and a tool covering several engines is still more useful than checking one by hand. But it's fair, and worth asking directly, to find out which engines a given tool tracks with the most confidence and which it tracks more thinly, rather than assuming uniform depth across a marketing page that lists five or six engine logos with equal visual weight.
The Claim: Anyone Can Directly Influence What a Model Learns
Some vendors imply, sometimes explicitly, that they have a direct channel to influence what a specific AI company's model learns during training, beyond simply publishing genuinely good content on the open web and hoping it gets included in a future training set like everyone else's content does. This is false for essentially every vendor making the claim. The major AI labs do not sell or grant direct training-data influence to marketing agencies or GEO vendors, and no legitimate, above-board service has a side channel into that process.
What is true, and is likely the origin of this myth in a distorted form, is that some AI companies have data partnerships with specific large platforms, licensing arrangements with publishers or content platforms for training data access. That is a business-to-business licensing relationship between a platform and an AI company, not a service any GEO vendor can offer or replicate on your behalf. If a pitch implies your brand can buy its way into a model's training data directly, that's not how it currently works for anyone outside those specific, named licensing deals.
The Claim: Precise Share-of-Model Measurement
Several tools in this category report a "share of model" metric, framed with the precision of a real, measured statistic, sometimes down to a specific percentage. It's worth understanding what this number actually represents before trusting it as ground truth.
No AI company publishes the actual internal weighting of how frequently your brand versus a competitor influences a given model's outputs across the full range of possible queries. Any share-of-model number you see from a third-party tool is necessarily a sampled estimate, built from a specific set of tracked prompts, checked at specific intervals, against a small fraction of the full space of things people actually ask these systems. That's a genuinely useful directional signal. It is not the precise, complete measurement the framing sometimes implies.
The honest version of this metric matters and is worth tracking. The dishonest version is presenting a sampled estimate with false precision, as though it were a verified, complete count rather than an informed approximation built on a limited prompt set (see Building Your Prompt Set, since the quality of the underlying sample is what makes this number more or less trustworthy in the first place).
The Claim: More Content Always Means More Visibility
Volume gets treated as a strategy on its own more often than almost anything else in this list, and it's worth separating what's true from what's false here specifically, because the false version is expensive to act on.
What's true: some volume of content is necessary. A single page, however well-crafted, can't cover every angle of a category, and having no content at all on a given topic is a real gap. What's false: that publishing more content, on its own, reliably increases visibility regardless of the content's specificity or quality. Ten generic posts do not outperform two genuinely specific, citable ones, and teams that chase a publishing quota instead of a quality bar often end up with a large archive that performs no better, sometimes worse, than a much smaller, sharper one, because a bloated archive of thin content can dilute a site's overall signal rather than strengthen it.
The honest framing: content volume matters up to the point where you've adequately covered your core buyer questions with genuinely specific, well-supported pages. Past that point, additional volume returns less than the same effort spent retrofitting or deepening what already exists (see Retrofitting Content You Already Have for the specific method).
What's Genuinely Unresolved
Not everything in this category sorts cleanly into true or false, and it's worth naming the piece that genuinely doesn't, rather than forcing a false confidence onto a question the industry itself hasn't settled.
How much direct weight community platforms like Reddit carry in a given AI engine's real-time retrieval versus its training data, and how that weight varies engine to engine, is a genuinely open question. There's reasonable evidence that some engines, ChatGPT in particular, draw heavily on this kind of content, partly due to known data partnerships and partly due to the conversational, question-shaped nature of the content matching how these systems are queried. But the precise mechanism, how much, under what conditions, and how consistently across different query types, isn't something any outside party can state with real confidence right now, including this platform. Treat community platform presence as a reasonably well-supported bet worth making, not a mechanism anyone can currently describe with precision.
Frequently Asked Questions
Is llms.txt worth implementing at all? Yes, as a low-cost, low-priority addition, since it costs almost nothing and won't cause harm. No, as a meaningful visibility strategy on its own. Treat it as a small item on a longer list, not a fix.
If schema markup isn't sufficient, is it still worth doing? Yes, definitely. It removes real friction and is a prerequisite for other work to register correctly. The myth isn't that schema is useless, it's that schema alone is enough, which it isn't.
How do I evaluate a vendor's claims without falling for an oversold pitch? Ask specifically what's guaranteed, and if anything about a specific AI engine's output is guaranteed by a specific date, treat that as a serious warning sign. Ask how a "share of model" or similar metric is calculated, and expect a real, specific answer about the underlying prompt set and sampling method, not a vague reference to proprietary methodology.
Does this mean GEO vendors and tools are all overselling? No. Plenty of vendors and tools in this category do genuinely rigorous, honest work and are transparent about what is and isn't measurable or guaranteed. The claims debunked here are specific, common overreaches, not an indictment of the entire category. Evaluating a specific vendor against this list is a reasonable, fair way to separate the honest ones from the overselling ones.
What's the single fastest way to spot an oversold pitch? Listen for absolutes attached to things nobody controls: guaranteed rankings on a named engine, a promise to influence model training directly, or a precision-looking metric with no explanation of how it's actually calculated. Honest vendors tend to talk in ranges and probabilities. Overselling ones tend to talk in guarantees.