AI visibility is a container word. It covers three different conditions that fail for three different reasons and look identical from outside, which is why so much money gets spent fixing the wrong one.
The three:
- Retrievability. Can an AI system get your content at all?
- Usability. Given your content, will it ground an answer in it?
- Attribution. Having used it, will it name you?
You can pass one and two and fail three. You can pass three occasionally while failing one for 90% of your pages. The symptom in every case is the same: your brand is not in the answer. Diagnosis has to be done in order, because fixing the wrong layer produces exactly zero movement.
Layer 1: retrievability
The unglamorous one, and the one most often quietly broken.
Generative engines mostly retrieve from search indexes built by crawlers. Google states that robots.txt directives for Googlebot are the control for site owners managing access to AI features in Search, which means your AI Overviews policy and your Googlebot policy are the same policy whether you meant that or not. OpenAI documents OAI-SearchBot for surfacing sites in ChatGPT search and recommends allowing it. Perplexity documents PerplexityBot as the crawler that surfaces and links websites in its results.
Three failure modes here, in order of how often I see them:
- Blocking a search crawler while trying to block a training crawler. These are different bots with opposite consequences. Blocking training affects models that do not exist yet. Blocking search makes you ineligible to be cited today. See AI user agents for the split, and the crawler directory for which token does which job.
- Content that only exists after JavaScript runs. Most AI crawler documentation does not mention rendering at all. Perplexity’s does not. OpenAI’s does not. Undocumented is not the same as no, but it is not something to bet a quarter on. What each crawler has committed to.
- Ordinary indexation problems. Orphan pages, canonical mistakes, thin sitemaps. Boring, and the most common cause of an invisible library of good content.
Check this layer in your server logs, not in a tool. Log lines are the only proof that a specific agent reached a specific URL.
Layer 2: usability
Retrieved is not used. The model still has to want your sentences.
Grounding operates on passages, so what gets used is a chunk of text that supports a claim the model is trying to make. That rewards a specific kind of writing: self-contained, factual, specific, checkable. It penalises the house style of most B2B marketing, which is designed to be persuasive rather than quotable.
The practical test is brutally simple. Take any paragraph on your key page and ask whether a model could cite it as evidence for anything. “Trusted by ambitious teams worldwide” supports no proposition. “Available in 14 regions with a 99.9% uptime commitment on the paid tier” supports several.
Original data outperforms everything else here, because grounding rewards the source that carries the fact rather than the fourth site to repeat it.
Layer 3: attribution
The engine used you and did not say so. This happens constantly, it is a product decision on their side, and you have limited leverage over it.
What appears to help: being the identifiable origin of a claim rather than a restatement, having a clean entity identity so the model has a name to attach, and being corroborated elsewhere. Engines name brands that appear consistently across independent sources more readily than brands that only appear on their own domain.
Which is why third-party presence does disproportionate work. A mention in a comparison post, a directory, a forum thread, a review site, or a roundup gives the retrieval layer somewhere else to find you and gives the attribution layer a reason to trust the name.
What does not drive it
Some honesty about the things people spend on that do not move the number:
- Publishing volume alone. Forty more pages of the same undifferentiated content adds forty more documents that ground nothing.
- Keyword density and its descendants. Retrieval is embedding-based. Repeating a phrase is not a signal.
- Any file no vendor has committed to reading. Including llms.txt, which is worth understanding precisely so you can size it correctly.
- Schema as a magic switch. Structured data helps machines parse entities and is worth doing. It does not create authority you do not have.
Google’s own statement is the useful calibration: there are “no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary.” Anyone selling a secret mechanism is contradicting the only primary source in the room.
Measuring it
Three instruments, each partial, and you need all three.
| Instrument | Proves | Blind to |
|---|---|---|
| Server logs | Which agent fetched which URL, and when | Whether it was ever used in an answer |
| Prompt monitoring | Whether you appear in answers | Why, and everything you did not prompt for |
| Referral analytics | Traffic that arrived from an AI host | Every answer that used you without a click |
The gap between the first and the third is the whole story of AI visibility as a business problem. Read citation insights for how to close it, and AI share of voice for why the prompt-monitoring number is less standardised than it looks.
The short version
Diagnose in order. Confirm a search crawler reaches and reads your pages, then make your pages contain claims worth grounding, then earn corroboration elsewhere so the engine has a reason to name you.
Skipping to layer three is the default behaviour, and it is why so many AI visibility programmes produce activity instead of appearances.