How to Spot a GEO Vendor Who Is Selling Nothing

Generative engine optimisation is young enough that nobody has a long track record in it, which is precisely the condition in which confident sales claims flourish. Some of what is being sold is careful work. Some of it is a landing page, a dashboard and a set of promises that cannot be kept by anyone, including the people making them. The good news is that the difference is detectable without any technical background, because the unkeepable promises describe machinery that does not exist.

We have an obvious interest here: we sell a measurement tool in this market. So the test below is written to apply to us as well, and where it does, it is stated.

Claims that cannot be true

Guaranteed citations, guaranteed rankings, guaranteed visibility

Nobody controls what a language model says about a brand. The answer is assembled at the moment of asking, from a model whose weights the vendor cannot touch and a retrieval step whose ranking they cannot see. Add that generation is probabilistic, so the same question asked twice can produce different names in different order. A guarantee of appearing in an answer is a guarantee over machinery the seller does not operate.

The softer version of the same claim is a promised percentage improvement in some proprietary visibility score. It sounds more measured. It is the same promise about the same uncontrolled system, wrapped in a metric the vendor defines and computes themselves.

A direct line to the model providers

Presented as a partnership, a private integration, or preferential inclusion in a knowledge base. Paying an intermediary to insert your brand into a model's training data or into an assistant's index is not a service the model providers offer. Anyone claiming otherwise should be asked, in writing, to name the programme and the counterparty.

Submission to AI engines

A resurrection of the search engine submission business from the early web, which was already obsolete when it was widespread. There is no queue to join and no form that grants inclusion. Where mechanisms exist for signalling to automated readers, they are open, documented and free, and they are hints rather than guarantees.

We will stop the AI from saying wrong things about you

A model producing a wrong statement about your company is not a bug with a support ticket. Publishing correct, retrievable information gives retrieval something better to find, which is genuinely worth doing and is the only honest version of this offer. It is a probability shifted, not a switch flipped, and it applies unevenly across assistants. Anyone selling reputational control over generated answers is selling a lever that does not exist.

Statistics with no source you can open

The single most reliable tell in this category. Sales material in generative engine optimisation is saturated with confident figures about how buying behaviour has shifted, how much traffic has moved, how many decisions now start with an assistant. Ask for the underlying study. What frequently comes back is a link to another article that cites another article, or a study conducted by the vendor on data nobody else can inspect. A figure whose source you cannot open is a rhetorical device, not evidence.

Case studies you cannot verify

An anonymised client in your exact sector, with a striking result, no verifiable name, no methodology and no way to check what else changed in the same period. Even in good faith, an uncontrolled before-and-after in a market where the assistants themselves change monthly cannot support the causal claim being made from it.

What a serious offer looks like instead

• The mechanism is explained, not gestured at. How retrieval selects passages, why a model's absorbed memory lags, why answers vary between runs. A vendor who cannot explain why something works usually cannot tell you when it will not.
• The limits are stated before you ask. Which assistants are covered and which are not. How often measurements are taken. What the tool cannot see, and what it will never be able to promise.
• Measurement is separated from influence. These are different businesses with different honesty requirements. A vendor who blurs them is usually charging for the first while implying the second.
• The raw output is available. Not only a composite score, but the actual answers, the prompts that produced them, the sources cited, exportable so you can check the score against reality or run your own analysis.
• The methodology is inspectable. Which prompts, which languages, which platforms, how often, how a score is computed. If a number cannot be reproduced from a stated method, it is decoration.
• Nothing depends on believing them. Every important claim should be checkable by you, in an afternoon, with a browser.

Questions to ask on the call

Five that tend to end the conversation quickly when there is nothing behind it. Which assistants do you query, and how often. Can I see the unedited answers behind this score. What exactly does your service do to my site or my content, and what does it do somewhere else. What result are you contractually promising, and what happens if the assistants change next month. And, most usefully: what can you not do.

A vendor who answers the last question with a real limitation is worth more of your time than one who treats it as an objection to be handled.

Applying it to us

Fairness requires running the test on ourselves in public. PSentry measures whether and how a brand appears in answers from ChatGPT, Claude, Gemini and Perplexity, per language and per market. That is the whole product. It does not optimise anything, it does not edit your site, it does not communicate with the model providers on your behalf, and it does not promise that your visibility will improve, because that promise would be about machinery we do not control.

The limits are equally concrete. Scanning is scheduled rather than continuous, so it is a series of snapshots and not a live feed. It covers those four assistants, not the overview features inside search engines or the assistants embedded in office software. Answers vary between runs, so a single result is weak evidence and a trend across runs is the thing worth reading. And a score is a summary: the exportable underlying answers are what you should check it against.

None of that is modesty for its own sake. A market this new punishes buyers who cannot tell measurement from magic, and the correction, when it arrives, tends to hit everyone selling in the category.

Frequently Asked Questions

Is generative engine optimisation a scam?

The underlying work is real: making your facts retrievable, publishing in the languages you sell in, keeping addresses stable, stating specifications as text. What is not real is control over what a model outputs. The category contains both, which is why the claims matter more than the label.

Can anyone guarantee a citation in an AI answer?

No. Not an agency, not a tool vendor, not a consultant with a track record in search. The system that decides is neither owned nor observable by any of them.

What should a monitoring report actually contain?

The prompts used, the platforms and languages covered, the dates, the unedited answers, the sources cited, and an explanation of how any composite score was computed. If you cannot reconstruct the score from what is in the report, ask why.

Is it worth hiring an agency for this?

It can be, for the publishing work: restructuring technical documentation, writing in your export languages, fixing the pages that are not machine-readable. Judge them on that work, which is verifiable, rather than on a promised position in an answer, which is not.

How do I verify a vendor's claim myself?

Take their headline claim, ask the same question of the assistants directly, in the languages you sell in, over several days. If your own results contradict theirs, ask them to reproduce it in front of you. A methodology that only works in the vendor's own dashboard is not a methodology.