Generative engine optimization services promise to make AI search engines mention your brand. Some of that work is real and worth paying for. A large part of what is currently sold under the name is a repackaged content retainer with a new cover sheet. This page is how to tell the difference, written by someone who sells the service and would rather be judged on the method than the label.
If you want the underlying discipline rather than the buying guide, start at AI search visibility.
What the work actually consists of
Strip the vocabulary away and a legitimate engagement is four things.
A baseline. A fixed sheet of buyer questions, run cold against each engine before anything is changed, with mentions and citations recorded per engine. Without this there is no way to demonstrate a result later, and any provider who skips it has chosen not to be measurable. The procedure is public: how to track brand mentions in AI search.
An access audit. Can each engine’s crawler fetch your pages. A yes or no answer per vendor, per agent, that takes under an hour and resolves a meaningful share of zero scores outright.
A structural pass. Rewriting the pages you already have so the answers are liftable: question-shaped headings, the answer in the first two sentences, one idea per paragraph, self-contained sentences. This is editing, not production, and it is where most of the gain sits.
A specificity pass. Finding the pages that say nothing a model could not have generated, and putting something real in them. Numbers, dates, limits, results from actual work. This is the part that cannot be outsourced to somebody who does not know your business, and it is the part most often quietly dropped.
Then re-measure on the same sheet. Everything else is optional.
The detail of what moves and in what order is in how to improve brand visibility in AI search engines.
What it cannot do
Nobody controls what an answer engine says. There is no submission endpoint, no placement to buy, no ranking factor to pull. The systems are non-deterministic, and two identical questions a day apart can produce different sources with nothing having changed.
That has three consequences a proposal should reflect. There can be no guaranteed placement. There can be no promised timeline for a mention in a named engine. And a single-run change, up or down, is not a result, which is why the honest cadence is monthly and the honest judgment is on three runs, not one.
A provider who is comfortable saying all of that is a better sign than one with a confident number.
How to read a proposal
Five questions, and the answers tell you most of what you need.
What is my number now, and how did you get it? If they cannot answer before the engagement starts, they will not be able to prove anything at the end of it.
Which engines are you measuring, separately? A single blended visibility score hides the case where one engine carries everything and the others are zero. Those need different fixes.
What are you actually changing? “Content optimization” is not an answer. “We rewrite the top forty pages so the answer sits under a question-shaped heading in the first two sentences” is.
What happens if the number does not move? A real answer distinguishes the four failure modes: not fetchable, fetchable but unquotable, quoted but not credited, absent from the category. Each has a different next step. A vague answer means the diagnosis is not part of the offer.
Who writes the specifics? The specificity pass needs facts only your business has. If the plan does not include getting those out of you, the output will be the generic text that does not get quoted.
The tell-tale signs
A guarantee. Covered above. Nobody has this to sell.
Volume as the plan. More posts is the reflex answer and it actively works against you here, because it gives an engine more text that is not worth repeating.
llms.txt as the centrepiece. Worth an hour, worth having, not a strategy. What changed in v2 has the detail.
A traffic report labelled AI Overviews. Search Console’s generative report counts impressions where links to your site were shown in a generative feature. It carries no clicks and no per-prompt breakdown, so anything more granular than that is modelled or scraped, not measured. The report and its limits are worth reading before you accept one.
No mention of indexing. For the Google surfaces, retrieval runs on the ordinary index. If your pages are not in it, that is the bottleneck and no amount of GEO touches it. Why pages do not show on Google is where that starts, and the paid version is indexing.
Whether you need to buy it at all
For a site with a small number of pages and someone in-house who writes, the honest answer is often no. The baseline is a spreadsheet and an afternoon. The access check is under an hour. The structural pass is editing, and the person who knows your subject will do it better than a contractor learning it.
Buying makes sense when the sheet is long enough that running it monthly stops happening, when nobody in-house has the time for a forty-page editing pass, or when the number came back low and you want the diagnosis rather than another round of guessing.
That last one is the part worth paying for. Everything before it is written down on this site for free, on purpose.
If you do want it run for you, it is AI search optimization. That page describes the method and claims no outcomes, because this site’s own baseline was zero when the cluster was planned and it would be a strange thing to sell results I have not yet shown.
Quick answers
What is the difference between GEO, AEO and AI SEO?
Mostly branding. Generative engine optimization, answer engine optimization and AI SEO are three names for the same work: making a site reachable, quotable and attributable to systems that answer questions instead of listing links. The terms were coined by different people at roughly the same time and the field has not settled on one. Judge a proposal on what it says it will do and how it says it will measure it, never on which acronym it chose.
How do FAQs help with AEO?
They help because of their shape, not their name. An FAQ block forces the exact structure an answer engine can use: a heading phrased as the question somebody actually asks, and a short self-contained answer directly under it with no preamble. That is why FAQ sections get lifted so often. The failure mode is padding a page with questions nobody asks, answered generically, to look thorough. Those add nothing, because there is no specific fact in them to quote. Five real questions answered concretely beat twenty invented ones.
What should a GEO engagement actually deliver?
At minimum: a documented baseline taken before any work, the prompt sheet it was taken with, a named list of the access and structural problems found, the changes made, and a re-measurement on the same sheet afterwards. If a proposal does not include a before number and the method used to get it, there is no way for either side to know whether the work did anything, which is convenient for exactly one of the two parties.
How much should generative engine optimization cost?
There is no market rate yet, which is itself worth knowing when you read a quote. What you can do is judge the price against the work described. A baseline plus an editing pass over existing pages is a defined piece of work with a defined size. An open-ended monthly retainer with no measurement attached is not, and the absence of a number to report against is usually why it is priced that way.
Can my existing SEO person do this work?
In most cases yes, and it is often the better option because they already know the site. The parts that are new are the measurement discipline and the passage-level editing habit, and both are learnable in an afternoon from a guide like this one. What is genuinely specialist is the diagnosis: reading which of the failure modes you have when the number comes back low. Hire for that if you need it, not for the acronym.