The question a GEO audit is asking
A GEO audit asks whether a passage on your site can be retrieved and reused inside an answer somebody else composes. That is a narrower and stranger question than whether a page can rank, and it produces a different list of problems.
A ranking audit assesses the page as a destination. Does it satisfy the visitor who arrives, is it competitive against whatever holds the position now, will it attract the links that keep it there. Every one of those is about the whole page and about the visit as the outcome.
A generative audit assesses the page as a supply of material. The unit is the passage rather than the page, and the outcome is not a visit at all. Something gets extracted, recombined with material from elsewhere, and served to a person who may never arrive on your site. The audit is asking whether your passages survive that process intact, and whether your name survives with them.
That difference explains why a great many findings from a ranking audit are simply irrelevant here, and why a few findings nobody used to record now matter enormously. A page can be beautifully optimized to rank and be useless as material, because the qualities that hold a visitor and the qualities that survive extraction are not the same qualities.
The findings you can fix this week
Roughly half of what a competent GEO audit surfaces is fixable inside two weeks with no engineering time. It is worth knowing which half, so you do not schedule a quarter for work that takes an afternoon.
Access rules nobody chose
The rules permitting or refusing the crawlers behind AI answers frequently arrive attached to something else. A bot-management default, a firewall rule written years ago against scrapers, a setting inside a hosting plan. Nobody decided to be excluded and nobody is notified. Reading the rules directly is the only way to know, and correcting them takes minutes.
Check the edge as well as the site itself, because these are two separate places a request can be refused and only one of them is usually inspected.
Openings that circle before committing
Take the pages carrying your commercial weight and read only their first two sentences. Commercial writing habitually sets context, concedes complexity, and states the point somewhere in the middle. A system assembling a short answer needs the point stated cleanly and early.
Moving a conclusion you already wrote to the top of the page is rewriting rather than research, and it is the highest-return edit available on most sites. It also improves the page for people, which is unusual for this kind of work and worth saying plainly.
Claims that cannot stand alone
A sentence depending on the paragraph before it cannot be lifted. Pronouns pointing backwards, comparisons with an unstated other half, figures without their unit, a "this" that refers to something two paragraphs up. These read perfectly to a person going in order and become meaningless the instant they are extracted.
Auditing for this is mechanical and fixing it is quick. It is also the finding that surprises people most, because the writing in question is often genuinely good writing that simply assumes a reader who started at the beginning. There is more on this in our piece on how LLMs choose sources.
Answers buried inside interface
The fourth quick finding is structural rather than editorial. A statement that only exists inside a tab, an accordion, a carousel or a modal is a statement that may never be read by anything fetching the page. The text can be present in the file and still be functionally hidden, depending on how it is assembled.
This one is worth checking because it hits the pages businesses care most about. Specification tables, delivery terms, sizing guidance and pricing detail all tend to live behind interface, precisely because designers were asked to keep the page tidy. Tidy and retrievable are in tension here, and the resolution is usually to keep the summary in plain prose and let the interface hold the detail.
The test is simple enough to run on your own site. Fetch the page as a plain document rather than opening it in a browser, and read what comes back. Anything you cannot find in that text is not reliably available to anything else either.
The findings that take a quarter
The other half of the audit names things that cannot be fixed quickly. An audit pretending otherwise is selling a schedule rather than describing the work.
Depth comes first. Being retrieved for a subject depends on there being enough about that subject on your site that reaching for you is a reasonable move. A single page, however good, is thin evidence of anything. Building a connected body of work on one subject takes months, and no amount of budget compresses it much, because the pages have to be genuinely different from each other to be worth having.
The common mistake here is breadth. Covering forty subjects shallowly reads as a large site and behaves as a small one, because retrieval happens per subject and forty thin subjects is forty thin subjects. One subject covered properly beats ten covered adequately, and that is an uncomfortable finding for anybody who has just commissioned a content calendar.
Being discussed elsewhere comes second. Your category is talked about in places you do not control, and whether you appear in those conversations changes what gets said about you. There is no setting for this and no supplier can deliver it as a unit of work. There is only doing things worth writing about and being easy to reach when somebody wants to write about them.
An honest audit separates these two from the two weeks work explicitly, and says which one is binding for you. Working on depth when your actual constraint is an access rule is a quarter spent on nothing at all.
How to tell a real one from a renamed one
The vocabulary is new enough that an existing service can be re-labeled without anybody quite lying. Four questions separate them, and all four can be asked before you commit to anything.
Ask what this audit measures that a ranking audit does not. If the answer is nothing specific, you are buying a ranking audit with a new cover on it. A good answer names passages, retrieval and a baseline.
Ask whether it includes a baseline, how many questions it uses, and how many times each is asked. A single pass over a handful of questions is not a baseline. The same question asked twice can return different names, so one reading cannot tell you what is normal.
Ask which crawlers were tested, by name, and what rule governs each. A general statement that the site is crawlable does not answer this, because search crawlers and answer crawlers are different clients and a site can be open to one and closed to the other.
Ask what the audit will tell you that you could not have worked out yourself. A supplier who cannot answer this has scoped a document rather than a diagnosis. It is a fair question and a good supplier will have a crisp answer to it, usually about the baseline or about depth.
The honest trade-off
The strongest argument against buying one is that the two weeks half is genuinely self-serviceable. Reading your own access rules and your own opening sentences needs no tooling and no expertise beyond knowing what to look for, and this page has now told you what to look for.
The strongest argument for buying one is the baseline and the outside read on depth. A defensible baseline is repetitive work that only means something if it is done consistently, and consistency is the first thing an internal team drops when a quarter gets busy. Judging whether your own subject coverage is thin is close to impossible from inside a business that has been writing about itself for years, because everything looks covered when you remember writing it.
Buy the parts you cannot see from where you sit. Do the rest yourself, and do it first, because it changes what you are buying.
What a GEO audit cannot see
Three things sit outside what any audit of your site can establish, and a document that claims them anyway is overreaching.
It cannot tell you why a specific answer named somebody else. An assistant composing a reply is drawing on material from many places and weighing it in ways nobody outside can inspect. You can see that a rival was named. You cannot see the reason, and any confident account of the reason is reconstruction rather than observation.
It cannot promise a position. There is no position to hold. A name is either in the answer or it is not, and the same question asked twice can come back differently. A supplier offering a ranking commitment here has borrowed a promise from a different discipline where it made more sense.
It cannot see what happens between the answer and the sale. A person who reads your name in a generated answer and then searches for you directly arrives looking like direct traffic. The assistant that caused the visit leaves no trace in your analytics. This is why baselines matter so much here. Being named is measured by asking, because the usual measurement route does not exist.
Naming these limits is a good sign in a supplier rather than a weakness. An audit that is candid about its blind spots has thought about where its evidence comes from, which is more than can be said for one that reports everything with equal confidence.
How to read the report when it lands
Read it in a deliberate order rather than front to back, because the ordering in the document usually reflects what was easiest to detect rather than what is most consequential.
Start with access. If anything in the report says a crawler is refused, nothing else in the document matters until that is corrected, and any recommendation about content is premature. This is a five-minute read and it can invalidate the other forty pages.
Go to the baseline second. Look at how many questions it used, how many times each was asked, and how much the answers varied between runs. High variation is not a fault in the measurement. It is a property of the thing being measured, and a report presenting a single confident percentage from a single pass has hidden that from you.
Read the page-level findings third, and read them looking for pages you already knew were important rather than for the longest list. Ten pages fixed properly beats sixty pages noted.
One more habit is worth building. Note which findings the report could only have produced by asking an assistant, and which it could have produced by reading your site alone. The second kind is useful and cheap. The first kind is the part you are actually paying an outsider for, and if there is very little of it in the document, you have bought a site review with a new title on the cover.
Read the depth assessment last and read it skeptically. It is the least verifiable section and the one most likely to conclude that you need more content, which is also the recommendation most likely to result in a further engagement. That does not make it wrong. It makes it the section to ask hardest about, and the question to ask is which single subject, not how many pieces.
Bottom line
A GEO audit earns its cost when it separates the two weeks work from the quarter work and tells you which one is binding. That single ranking is the deliverable. Everything else in the document is evidence for it.
Run the access check and the opening-sentence check yourself before briefing anybody. They cost a day, they resolve the most common cause of total invisibility, and they turn the supplier conversation from a search into a scoping.
Then insist on a baseline before any work starts. It is the cheapest item in the engagement and the only one that cannot be produced retrospectively, which is precisely why it is the one that goes missing.
And treat the first report as a starting position rather than a verdict. The useful version of this work is a reading you take again in three months against the same questions, so the document is worth less than the habit it should leave behind.