Guides · 2026-08-28 · By VSNARY | Emmanuel Orta · 0 views

GEO, AEO, LLM SEO and the agentic web: what the terms mean and which parts can be measured

Three names for one shift, sold as three disciplines. The distinctions are mostly marketing; the useful split is between the parts of the work that produce a checkable number and the parts that do not.

Generative engine optimization (GEO), answer engine optimization (AEO) and LLM SEO are three names for the same shift: a growing share of questions about your business are now answered without anyone visiting your site. Practitioners will offer distinctions — AEO leaning toward featured snippets and direct answers, GEO toward generated prose, LLM SEO toward the models specifically. In practice the work overlaps almost completely, and which term someone uses tells you more about when they entered the field than about what they do.

The distinction worth drawing is different, and it cuts across all three labels.

The split that matters #

Both halves are legitimate. The problem is that the second is routinely presented with the confidence of the first. A share-of-voice figure is a statistic over prompts somebody guessed at, produced by querying an API that is not the consumer product anyone actually uses, on a day the model happened to be in a particular state. Change the prompt list and the number changes. That is not a flaw in the arithmetic — it is what the arithmetic was always describing.

How it differs from SEO, honestly #

Less than the branding suggests, and more than the sceptics allow.

What carries over: being crawlable, being fast, being genuinely useful, having other people reference you. None of that stopped mattering.

What is genuinely new: there is no position to hold. A ranked list has slots; an answer has a paragraph and maybe a citation. You are either used or you are not, and the binary is harsher than page two.

What is new and mostly unmeasured: the cost side. Every SEO tool measures what a page says. Almost none measure what a machine is handed when it asks for that page. The heaviest page measured here delivered 476,540 bytes carrying 2.3% visible text, and no validator had an opinion about it.

The measurable work, in order #

Whatever the label on the invoice, this is the part that produces numbers you can check.

Reachable. Does the fetch succeed for that specific crawler? ANSWER_ENGINE_REFUSED fires on 5.8% of scans, and 4.5% carry an AI opt-out — much of it inherited rather than chosen, given that the eight largest opt-out groups on the web sit within 12% of each other.

Readable. Is what came back parseable? A cached challenge page at HTTP 200 reads as allow-everything. STALE_CACHE_SERVED fires on 17.4% of scans, the most common defect in the dataset.

Quotable. Is there enough text to use? 27.6% of scans return a page that is overwhelmingly code, and 31.7% publish no llms.txt at all.

Resolvable. Can you be told apart from similar entities, and are the specific facts declared? Organization is detected on roughly 354,000 sites web-wide; OpeningHoursSpecification on roughly 23,000.

What the people selling it have done to their own sites #

Thirty-three SEO and AI-visibility vendors were run through a supply-side scanner — the same battery they would point at a customer, on the same day, none told in advance.

Several were serving a cached machine file that no longer matched their own origin, meaning a crawler and a human get told two different things by a company that sells judgement about exactly that. The point is not that they are bad at their jobs — it is that these checks are not specialist requirements, and the people closest to the advice mostly have not implemented it.

The number no single audit can produce #

Whatever the label, most of this is sold as an audit: a report on a day. The same sites, measured every day, tell a different story from any one of those days — and the gap is computed below when this page loads, not written into it.

A one-off engagement at any budget can only ever produce the first figure. The second requires the same sites to have been watched before anyone asked the question — which is a difference in kind, not in thoroughness.

The test to apply to anyone selling any of the three #

Ask what would have to be true for their number to be wrong, and how you would find out. If there is no answer, it is not a measurement — it is a label. A percentage you cannot check is a claim about something nobody measured.

Then ask which surface was queried, which model in which mode, and on what date. Grounded retrieval and generation from weights are different systems and averaging them produces a number about neither.

None of this makes the category fake. It makes most of the reporting in it softer than it is presented, while the hard, checkable half — the part where a server either answers a crawler or does not — is sitting there almost entirely unsold.

What moved from the inferred column into the measurable one #

The split at the top of this guide is not fixed. Work crosses it whenever somebody builds an instrument, and two checks have crossed recently.

Crawl waste. How much parameterised and duplicated URL space a crawler has to walk before it reaches anything worth reading. This was always describable and never countable. It is now counted per site.

Topicality. Whether a page’s own visible text supports the subject its structured data claims. A site can declare a service in schema and never mention it in a sentence, and until this check existed that gap was an opinion. It now produces a number.

Two more now appear on every report without affecting the grade: whether a declared parent organisation declares its locations back, and whether the services a site names line up with the recognised service intents behind them.

“Reported, not scored” is a real category #

Of the 35 sections on a report, 24 count toward the grade and 11 are reported without touching it. A check that can only run where a particular file or platform exists would, if scored, grade most sites on their hosting rather than on their work; reporting it unscored gives the finding without distorting the comparison.

Two scores from different versions are not comparable #

Check this before you compare any two AI-visibility numbers: when a tool changes how it scores, every previously stored result goes stale. Scores produced before and after the change came out of two different instruments, and the delta between them describes the instrument rather than the site. Which is why cross-version deltas are never published here at all.

The three outcomes a row can have #

Pass, fail, and not measured. The third is what separates an instrument from a checklist.

A checklist has two states, so an item it could not test has to be filed as one of them — almost always a fail, because a blank looks like an oversight. That produces the familiar audit output where a site is marked down for something nobody actually checked. Here an unmeasured component is excluded from the calculation rather than scored zero: a zero says we looked and it was absent, a blank says we could not look. Collapsing the two is how a score stops being about the site.

How to tell which half of the work you are being sold #

One question does most of the sorting: can I reproduce this number myself, and what would I need?

Neither of the first two is illegitimate. The difference is that one hands you the method along with the number, and the other hands you the number and keeps the method as the product.

What none of the three acronyms can sell you #

A guarantee of citation. Nobody holds the distribution of what buyers ask, nobody controls which sources a model composes from on a given day, and no amount of optimisation makes an answer name you. What any of the three can honestly deliver is the input side — reachable, readable, quotable, resolvable — plus a dated record of when each of those was true, so that a later change has something to be measured against.

If that sounds narrower than the pitch, it is. It is also the part that survives being checked, which is the only durable test of whether you bought a measurement or a label.

Intermittent problems and permanent ones look identical in one audit #

This is the concrete version of the stale-audit argument, and it is the strongest thing a repeated measurement says that a single one cannot.

Take a fixed panel of sites and measure them every day for a window. For each finding you then have two numbers: how many sites hit it ever during the window, and how many are hitting it now. The gap between the two tells you what kind of problem it is.

Absence findings — no llms.txt, no sitemap, no entity map — have a gap of roughly zero. A file that is missing today was missing yesterday, so ever and now agree, and one audit captures them perfectly.

Delivery findings behave nothing like that. Stale cache served, refusals, challenge pages: these appear and disappear on the same site inside a fortnight, so ever runs well ahead of now. A single audit catches them only if it happens to land on a bad day, and reports a clean site if it does not. In the panel, twelve sites that looked clean at the end of the window had not been clean throughout it.

Which is the real answer to “is an AI visibility audit worth it”. For the absence half, yes, and once is enough. For the delivery half, a one-off engagement at any budget is measuring the weather on a single day and calling it the climate — not because the auditor is careless, but because the finding does not exist as a single reading.

The agentic web: when the machine acts instead of answering

GEO, AEO and LLM SEO all describe being read and cited. The agentic web is the next step: AI systems that act on a person's behalf, opening pages, comparing options, filling in forms and calling tools. An answer engine needs to read you; an agent needs to read you, trust you and complete something on your site. That adds questions none of the three acronyms covers, and each one can be measured from outside.

CrawlCheck is built for that list: one signed answer about a site that an agent can check before it reads, cites or acts. The same discipline applies as for the acronyms above: measure what the site serves, never infer what a model concluded.

Findings behind this #

Every figure above comes from a measurement written up in full, with the date, the method and the raw numbers.

• AI visibility scores measure an API, not what anyone sees in ChatGPT
• 33 AI visibility tools measured: a third publish llms.txt, none an entity map
• Site audits go stale: the day you audited it is not the finding

Questions this post answers

What is the difference between GEO, AEO and LLM SEO?

Very little in practice. All three describe optimising for answers generated without a click. The more useful distinction is between work that produces a verifiable measurement, such as whether a crawler can fetch and parse your site, and work that produces an inferred one, such as a share-of-voice percentage over a chosen prompt list.

Is generative engine optimization real, or is it rebranded SEO?

The underlying shift is real: answers increasingly replace result pages, and there is no ranked position to hold. Much of the practice carries over from SEO. What is genuinely new and mostly unmeasured is the cost side, meaning what a machine is actually handed when it requests your page.

Where should I start with GEO or AEO?

With the checkable layer, in order: whether the fetch succeeds for each crawler, whether what comes back is parseable, whether enough of the payload is quotable text, and whether your entity and its specific facts are declared. Buying answer monitoring before that measures an outcome that is not yet possible.

Is GEO different from AEO?

Not in practice. Both name the same work: making a site reachable, readable and quotable by the clients that assemble answers. Which label somebody uses tells you roughly when they entered the field. The distinction worth drawing runs across both, between work that produces a number you can reproduce and work that is inferred from prompting a model.

What part of GEO can actually be measured?

The retrieval side. Whether a non-browser client gets a response at all, whether there is readable text in that response before anything executes, whether any span of it stands alone well enough to be quoted, and whether the entities named on the page resolve to something. Each is a request you can repeat yourself, and each produces a status, a ratio or a count rather than a percentage over somebody else's prompt list.

Why do AI visibility scores from different tools disagree?

Because they measure different things and report them under one name: different surfaces, different prompt lists, different dates, and often different scoring versions of the same tool. Two numbers are only comparable when the surface, the sample, the date and the version all match. That is rarely stated, and it is the first thing to ask for.

What is the agentic web?

The part of the web used by AI agents that act for a person rather than only answering: opening pages, comparing, filling in forms and calling tools. For a site it adds four measurable questions: permission (what may be fetched), delivery (what is actually served), identity (who the site and the requester are) and tools (whether declared actions behave as described).

Related findings

Comments

Comments are read before they appear. Nothing is published automatically, and no account is needed.

Writing about this? Facts, live figures and marks — every number on that page is dated and traceable to a scan.

All guides · The dataset · How the dataset works