A GEO audit is a structured evaluation of whether a website can be retrieved, extracted, identified, trusted and ultimately cited by generative search engines. It runs across five layers — retrievability, extractability, entity clarity, authority and measured presence — and produces a prioritised list of defects rather than a score alone. Its purpose is to establish the gap between what a property could be cited for and what it is actually cited for today.

How is a GEO audit different from an SEO audit?

An SEO audit asks whether pages can rank; a GEO audit asks whether passages can be quoted and whether the brand can be identified as an entity worth quoting. The first three layers overlap heavily with technical SEO, which is why a GEO audit is usually run on top of one rather than instead of it. The genuinely distinct parts are the extractability review — whether each block answers its own question completely — and the presence panel, which measures what engines actually say today.

How do you measure whether AI engines cite a site?

By running a fixed panel of representative prompts across the engines that matter to the market, monthly, and logging which sources each answer credits. Roughly thirty prompts covering the questions a real buyer asks is enough to read a direction. Results must be segmented by language and country, because the same prompt returns materially different answers by locale, and tracked as a trend rather than a snapshot, because individual runs vary.

What should you fix first after a GEO audit?

Always in layer order. A blocked crawler makes perfect content invisible, so retrievability defects outrank everything; entity ambiguity makes authority unattributable, so it outranks content work. The common mistake is to start with content rewrites because they are the most satisfying, while a robots directive or a client-side rendering problem quietly prevents any of it from being read.

Do you need an llms.txt file?

It is cheap to publish and not yet load-bearing. As of 2026 it is an emerging convention with uneven adoption rather than an established standard, so treat it as a low-cost signal rather than a fix: publish a clear one if you have a clean information architecture to describe, but do not expect it to compensate for a site that is hard to crawl or hard to extract from.

How long does a GEO audit take?

The technical layers take a few days for a normal site. The presence panel takes an hour to run but only becomes useful after two or three monthly repetitions, because a single run of generative answers is noisy. That is why we start the panel at the beginning of an engagement rather than at the end — the baseline is worth more than the first reading.

The GEO Audit: How We Evaluate AI Search Visibility Before We Build or Buy

What a GEO audit is

A structured check of whether a property can be used as a source by generative engines, and a measurement of whether it currently is. Not a score with a dial on it — a prioritised list of defects, ordered by which ones prevent the others from mattering.

Definition: GEO auditA structured evaluation of whether a website can be retrieved, extracted, identified, trusted and cited by generative search engines, assessed across five layers: retrievability, extractability, entity clarity, authority and measured presence.

The layer order is the method. Each one gates the next, so a defect high in the stack makes everything below it unreadable — which is why the finding “your content needs work” on a site that blocks AI crawlers is not a finding, it is a distraction.

Why it runs before you build or buy

Because it is the cheapest moment for the answer to change a decision. Run against an asset we are pricing, the audit tells us whether the topical authority we think we are buying is real and legible, or whether it is a traffic number resting on one page and a fragile channel. Run against a company we are starting, it tells us what has to be true in the architecture before anyone writes a line of content.

Run afterwards, it becomes a repair list. Same work, several times the cost, and some of it — URL structure, entity naming, information architecture — is expensive precisely because it was decided early by default.

Layer 1 · Retrievability

Can a machine fetch and read the page at all? This layer is binary and it overrides everything below it.

Crawler directives. What robots.txt and meta robots actually permit — per crawler. AI crawlers are distinct agents and are frequently blocked by an inherited default nobody chose.
Rendering. Does the content exist in the server response, or only after JavaScript? Assume some retrievers do not execute scripts and check the raw HTML rather than the rendered DOM.
Response and status hygiene. Redirect chains, soft 404s, canonical conflicts, and pages that are slow enough to be abandoned.
Discovery. A sitemap that matches reality, internal links that actually reach every important page, and no orphan content.
llms.txt. Cheap to publish, not yet load-bearing. A useful signal in 2026, not a substitute for any of the above.

Layer 2 · Extractability

Given the page, can a usable passage be lifted out of it? This is the layer with the most defects and the least attention.

Answer-shaped blocks. Does each section answer its own question in its first sentences, complete and self-contained? The full discipline is in the answer-shaped content piece.
Heading semantics. One h1, a logical hierarchy, headings that state the question rather than a pun.
Anchors. Stable id attributes on every section, so an engine can deep-link to the passage rather than the page.
Structured data. Article or BlogPosting, FAQPage where there is a real Q&A, Organization and Person for the entity spine, and DefinedTerm where the page actually defines something.
Tables and lists. Comparative information in an actual <table> rather than styled divs — extraction is markedly better from real semantics.

Layer 3 · Entity clarity

Can the engine work out who the source is, unambiguously, and match this site to the same entity everywhere else it appears?

One name, one category, one scope. Stated identically across the site, the schema and every external profile. Drift here is the most common cause of an authoritative site that never gets named.
sameAs coverage. Real profiles that corroborate the entity, not a list of empty accounts.
Per-language resolution. Engines resolve entities by language. If the description is crisp in English and vague in Spanish, you will be cited in one language only — the mechanism behind the citation gap.
People as entities. A named author with a resolvable identity is a stronger trust signal than any amount of brand copy.

Layer 4 · Authority

Would a system be right to trust this source on this subject? Almost everything here is classical credibility, weighted more heavily than it used to be.

Topical depth. A defensible cluster on a defined subject, not one article per topic across twenty topics.
Original information. Facts that exist nowhere else — first-party data, documented process, named examples. Restated common knowledge is not citable because it is already held.
Attribution and dates. Real authors, accurate publication and modification dates, sources cited where claims are borrowed.
Honest limits. Content that states what it does not know is measurably more useful to a system trying to decide what to repeat.

Layer 5 · Presence

The only layer that measures reality rather than readiness: what do the engines actually say today?

Build a panel of about thirty prompts. The questions a real buyer asks, in their words — not keyword-tool phrasing. Include category questions, comparison questions, and “who should I use for X in Y” questions.
Run it across the engines that matter in your market, in every language and country you serve, and log which sources each answer credits.
Record three things per prompt: whether you appear, who does appear, and which page of theirs was used. The third is the most actionable and the most often skipped.
Repeat monthly. A single run is noise. The direction over three runs is the finding.

This produces the answer share baseline. It is manual, it is sampled, and in 2026 it remains the most honest instrument available.

Scoring, and what to fix first

We score each layer separately and refuse to average them, because an average hides exactly the thing the audit exists to find: a single blocking defect high in the stack.

Layer	Failure looks like	Cost to fix
1 Retrievability	Invisible regardless of quality	Hours — usually a config change
2 Extractability	Read but never quoted	Weeks — editorial and template work
3 Entity clarity	Quoted without being named	Days — mostly consistency work
4 Authority	Named occasionally, never as the source	Quarters — this is the slow one
5 Presence	You cannot tell whether any of it worked	An hour a month, forever

Swipe to see the whole table →

Fix in layer order. The temptation is always to start at layer four because writing is the enjoyable part — and layer four compounds over quarters while layer one is an afternoon.

How we think about it at Thinktier

This audit is stage 04 of the operating system made concrete, and it is the same instrument whether we are starting a company or rebuilding an acquired one — where it sets the sequence described in the six levers.

The reason we publish the method rather than describing it vaguely is that this site is the working example. Every Insight ships with anchored question-headings, a contents outline, FAQ markup and a DefinedTerm where it defines something — which means the audit can be run against us, by anyone, including you.

Sources

Overview of OpenAI Crawlers — OpenAI
The rise of the AI crawler — Vercel, 2024-12-17
The /llms.txt file, v2 — llmstxt.org (Jeremy Howard), 2024-09-03
AI features and your website — Google Search Central
Creating helpful, reliable, people-first content — Google Search Central
GEO: Generative Engine Optimization — Aggarwal et al., KDD 2024 (arXiv:2311.09735), 2024

Q&A — structured for AI search

What is a GEO audit?: A GEO audit is a structured evaluation of whether a website can be retrieved, extracted, identified, trusted and ultimately cited by generative search engines. It runs across five layers — retrievability, extractability, entity clarity, authority and measured presence — and produces a prioritised list of defects rather than a score alone. Its purpose is to establish the gap between what a property could be cited for and what it is actually cited for today.
How is a GEO audit different from an SEO audit?: An SEO audit asks whether pages can rank; a GEO audit asks whether passages can be quoted and whether the brand can be identified as an entity worth quoting. The first three layers overlap heavily with technical SEO, which is why a GEO audit is usually run on top of one rather than instead of it. The genuinely distinct parts are the extractability review — whether each block answers its own question completely — and the presence panel, which measures what engines actually say today.
How do you measure whether AI engines cite a site?: By running a fixed panel of representative prompts across the engines that matter to the market, monthly, and logging which sources each answer credits. Roughly thirty prompts covering the questions a real buyer asks is enough to read a direction. Results must be segmented by language and country, because the same prompt returns materially different answers by locale, and tracked as a trend rather than a snapshot, because individual runs vary.
What should you fix first after a GEO audit?: Always in layer order. A blocked crawler makes perfect content invisible, so retrievability defects outrank everything; entity ambiguity makes authority unattributable, so it outranks content work. The common mistake is to start with content rewrites because they are the most satisfying, while a robots directive or a client-side rendering problem quietly prevents any of it from being read.
Do you need an llms.txt file?: It is cheap to publish and not yet load-bearing. As of 2026 it is an emerging convention with uneven adoption rather than an established standard, so treat it as a low-cost signal rather than a fix: publish a clear one if you have a clean information architecture to describe, but do not expect it to compensate for a site that is hard to crawl or hard to extract from.
How long does a GEO audit take?: The technical layers take a few days for a normal site. The presence panel takes an hour to run but only becomes useful after two or three monthly repetitions, because a single run of generative answers is noisy. That is why we start the panel at the beginning of an engagement rather than at the end — the baseline is worth more than the first reading.

Keep reading

About the author

Federico Noya

Founder & Operator, Thinktier

Writes about how AI search, distribution architecture and operating discipline compound across the Thinktier portfolio.

More about Federico

Back to all Insights The Glossary