AI visibility is how often, how accurately, and how favourably generative AI systems — ChatGPT, Claude, Gemini, and Perplexity — represent your brand when they answer questions about you or your category. It is the AI-era successor to search ranking: instead of where you sit in a list of links, it measures whether the model names you, what it says, whether it recommends you, and which of your pages it cites.
This page explains how AI visibility works, the four layers it breaks down into, and how to measure it. It is a living resource — we update it as the models and the field change.
AI visibility is the degree to which generative AI systems know your brand exists, represent it accurately, recommend it, and cite your content when answering relevant questions. Where traditional search visibility is a public, rankable position on a results page, AI visibility plays out inside generated answers that are produced and read in private — so it has to be sampled and scored rather than simply observed.
Three things make it different from the visibility you are used to:
Generative engines do not invent knowledge about your brand. They assemble it from two places: what they absorbed during training, and what they retrieve from the live web at answer time. Understanding both is the whole game.
Training knowledge is what the model learned from its corpus — the open web, licensed data, and everything written about you up to its cut-off. This is slow-moving and consensus-based: the model believes what most of its sources agree on. If your brand is barely mentioned, or described inconsistently across the web, the model's internal picture of you is thin or wrong.
Retrieval is what the model fetches at the moment it answers. Perplexity retrieves and cites pages. Gemini sits on Google's index. ChatGPT browses for anything current. This is fast-moving: a newly published, well-structured page can be picked up within weeks and used to ground an answer — which is why retrieval is the fastest lever you have.
An answer is then generated by combining the two: the model recalls what it knows, grounds the current and checkable parts in retrieved sources, and writes a single response. AI visibility is the measure of how your brand fares across that whole pipeline — recalled, retrieved, and represented.
This is also why AI visibility consumes SEO rather than replacing it. A page that cannot be crawled, parsed, and trusted never gets retrieved, and a brand that earned no authority rarely gets cited. Good SEO is the substrate AI selection runs on. For the full comparison, see GEO vs SEO.
AI visibility is not a single number. It is a stack. We score it in four layers — the L1–L4 framework, drawn from Aaron Haynes' 2025 research and refined through our own daily observations. Each layer fails differently, and each failure has a different fix.
L1 measures whether a model can name your brand and recall basic, correct facts about it without being fed them. Nothing in the layers above matters if the model does not recognise the entity. An L1 failure looks like the model never mentioning you, confusing you with another company, or saying it has not heard of you.
Measured by unprompted mention rate + basic-fact accuracy.
Once a model knows you exist, L2 evaluates how deep and accurate that knowledge goes — your products, positioning, people, pricing, and proof. An L2 failure looks like outdated claims, missing capabilities, or a generic description that could fit any competitor.
Measured by accuracy + completeness of claims across probing prompts.
L3 tracks whether your brand surfaces when a user asks an answer-shaped, solution-seeking question in your category — "what’s the best tool for X?" — not only when they ask about you by name. This is the layer closest to demand. An L3 failure looks like competitors being recommended while you are absent.
Measured by recommendation rate + share of voice vs competitors.
L4 traces which URLs a model pulls from when it grounds an answer with sources. Citations are the AI-era equivalent of the click: being the named source for a claim. An L4 failure looks like a competitor’s blog post being treated as the canonical source for facts about your own product.
Measured by citation rate + which pages earn the citations.
Together the four layers form a single trend you can watch move. The trace above is a mock for a fictional brand — each line is one layer, climbing as its gaps close.
These terms get used interchangeably, so here is the clean distinction: AI visibility is the outcome you measure; GEO, AEO, and AI search optimisation are the practices you use to improve it.
| Term | What it is |
|---|---|
| AI visibility | The measurable outcome — how AI represents your brand (the L1–L4 stack) |
| GEO (generative engine optimisation) | The practice of optimising for AI-generated answers. Coined in "GEO: Generative Engine Optimization" (Aggarwal et al., KDD 2024) |
| AEO (answer engine optimisation) | Optimising for any direct-answer surface — featured snippets, voice, AI answers. Near-synonym of GEO in 2026 usage |
| AI search optimisation / LLMO | Umbrella labels for the same work; used interchangeably with GEO |
Do not optimise for the acronym. Identify the surfaces your customers actually use — Google, AI Overviews, ChatGPT, Perplexity — and improve your visibility there. The deeper breakdown lives in our GEO vs SEO guide, and the tooling landscape in the best AI visibility tools.
"Do the models like us?" is not a KPI. These are the metrics that map onto the four layers, and the equivalents to the SEO numbers you already track:
| Metric | What it answers | Layer | SEO analogue |
|---|---|---|---|
| Mention rate | How often models name you in category answers | L1 | Impressions |
| Accuracy | Whether claims about you are true and current | L2 | — |
| Recommendation rate | How often you are the suggested solution | L3 | Position |
| Citation rate | How often your URLs ground answers | L4 | Clicks |
| Share of voice | All of the above, relative to competitors | L1–L4 | Share of search |
The catch is sampling. Because answers are private and noisy, a single response tells you almost nothing — the same prompt can name you once and forget you the next run. Robust measurement aggregates across many runs, multiple models, and a stable watchlist of prompts, and reports medians so one unusually good or bad answer does not move the number. Measure weekly; judge monthly.
In practice, that readout looks like this — an assistant reading a brand's four-layer scores and competitive share of voice straight from BotScope:
How did Acme’s AI visibility do this month, and how do we compare to competitors?
Acme’s overall AI-visibility score is 72, up +6.4 on last scan — its strongest month yet. The shape underneath:
On share of voice you’re leading at 58% across tracked queries — ahead of Globex (22%) and Initech (11%), and the gap widened this month. You’re winning at L1/L2 and now pulling ahead on recommendations; the main upside left is L4 citations on Perplexity, where Globex still edges you. Want me to pull the prompts where a competitor gets cited and you don’t, so you can target them?
Fix the lowest broken layer first — there is no point optimising recommendations the model can give if it does not know you exist.
BotScope runs your question watchlist against ChatGPT, Claude, Gemini, Perplexity, and Grok every day, scores all four layers, and shows you which gap to close first.
AI visibility is how often, how accurately, and how favourably generative AI systems — ChatGPT, Claude, Gemini, Perplexity — represent your brand when answering questions about you or your category. It breaks into four layers: whether the model knows you (L1), what it claims about you (L2), whether it recommends you (L3), and which of your pages it cites (L4).
SEO earns you a position in a public list of links you can inspect. AI visibility is about whether a generated answer names, recommends, and cites you — and those answers are produced privately, so they must be sampled rather than read. The fundamentals overlap: generative engines retrieve from search indexes, so weak SEO also weakens AI visibility.
By sampling the models directly: run a fixed set of category and brand prompts against each model on a schedule, then score mention rate, claim accuracy, recommendation rate, citation rate, and share of voice. Because single answers are noisy, aggregate across many runs and report medians. Measure weekly, judge monthly.
Usually one of two layers is broken. Either the model does not have a strong enough entity for you (an L1 problem — too few or inconsistent mentions across the web), or it knows you but does not surface you for category questions (an L3 problem — competitors own the comparison content it retrieves). The fixes are different, which is why you diagnose the layer first.
Citation-level changes (L4) can appear within weeks, because retrieval-backed engines pick up newly indexed content quickly. Entity-level changes (L1–L2) move slower — they depend on consensus across many sources and on model update cycles. Plan in weeks for citations and in months for what models fundamentally believe about you.
This is a living resource, last updated June 2026. AI visibility is a young field; we revise this page as the models, the surfaces, and our own measurements change. Definitions follow the L1–L4 framework BotScope uses in production. Sources: "GEO: Generative Engine Optimization" (Aggarwal et al., KDD 2024); Aaron Haynes (2025) on the four-layer model.