The AI Visibility Playbook

What AI visibility (AEO) is

AI visibility is the measure of how often, and how accurately, large language models cite your brand, product, or content when a user asks a question in ChatGPT, Claude, Perplexity, Google AI Overviews, or Gemini. Answer engine optimization (AEO) is the practice of making a page more likely to be selected and quoted by those systems. Generative engine optimization (GEO) is the same discipline applied to broader generative answers, not just direct citations.

The unit of victory in classical SEO is a ranking position on a search engine results page. The unit of victory in AEO is a passage: a specific sentence or paragraph the model chose to reproduce, paraphrase, or link. A page can rank position 1 for a query and still not be quoted; a page ranked position 12 with a well-structured definition can be quoted more often than the winner.

Three signals separate cited pages from ranked-but-uncited pages: an explicit definition in the first two sentences, factual claims formatted for extraction (numbers, dates, named entities), and structured data that names the concept the page is about. Everything else in this playbook is downstream of those three.

Why classical SEO metrics do not measure AI visibility

A rank tracker records the position of a URL for a query at a point in time. It does not record whether an answer engine chose your URL as a source, whether it quoted you verbatim, or whether it paraphrased a competitor while linking to you as a secondary citation. These are different events with different revenue consequences.

Classical SEO metrics also treat the SERP as the surface where the decision happens. In an answer-engine flow, the decision happens in the model. The user sees a synthesised answer, sometimes with citations, sometimes without. If your brand appears in the answer body but not in the citation list, you have earned attention with zero click and zero attribution in Google Analytics. Rank trackers do not surface this. Citation trackers do.

The corollary: a page can lose ten rank positions and gain revenue if it starts being cited by ChatGPT for a buyer-intent query. Rank tracking will read this as a regression. AI visibility tracking will read it as a win. Teams that only look at rank data will make the wrong optimization decision.

How answer engines actually select and cite sources

Every major answer engine combines two stages: retrieval and generation. In retrieval, the system fetches a shortlist of candidate documents. In generation, the model reads those documents and writes an answer. Cited pages are the ones that survive both stages: retrieved with high relevance and easy to quote once retrieved.

The signals engines share

• Explicit entity resolution. Pages that clearly name what they are about (a product, a concept, a company) with consistent naming across schema, title, H1, and body get clustered more reliably. Anthropic and OpenAI have both published on entity clarity as a factor in retrieval-augmented pipelines.
• Factual density. Short, specific, numeric claims are easier to extract than long paragraphs. A sentence like "Google AI Overview appears on 47% of desktop US queries as of Q2 2026" is quotable; the same claim buried in prose is not.
• Structured data. Schema.org types such as DefinedTerm, FAQPage, HowTo, Article, and SoftwareApplication give the engine an unambiguous map of the page. Google publicly documents Article, FAQPage, and HowTo as inputs to AI Overviews.
• E-E-A-T signals. Named author bylines, credentials, dateModified stamps, and links to an About page raise the trust weight of a document. Google published its E-E-A-T guidelines in December 2022 and has referenced them in every Search Central post about AI Overviews since.
• llms.txt. A root-level llms.txt file (spec at llmstxt.org) gives crawlers a summarised map of the site. It is not required but reduces the retrieval cost for engines that support it.

Where the engines diverge

• ChatGPT (with browsing) retrieves via Bing plus OpenAI web crawlers (OAI-SearchBot, GPTBot). Citation coverage is uneven across topics. Confirmed via OpenAI documentation, October 2024.
• Perplexity retrieves from a live index, cites densely (typically 4–8 sources per answer), and preserves inline citations. PerplexityBot is the crawler.
• Claude (with tools) uses first-party search integrations. Anthropic documents the tool-use pipeline; ClaudeBot is the crawler for training and Claude-Web for on-demand fetches.
• Google AI Overview generates in-SERP answers seeded by the top organic results plus Knowledge Graph. Coverage was formally launched at Google I/O 2024 and expanded through 2025.
• Gemini uses Google Search results as retrieval context in most flows; behaviour tracks AI Overview signal set closely.

A concrete AEO audit checklist you can run today

Run this checklist against your top 10 pages ranked by pipeline value, not by traffic. AEO leverage is highest on pages that are already found but not yet cited.

Definitions and factual formatting

• The first paragraph contains a one-sentence definition of the page topic in the shape "X is Y".
• The definition uses the exact term a user would type into ChatGPT (not the branded name of your product).
• Numeric claims include unit, magnitude, and source in the same sentence.
• Dated claims include the month and year of the source, not just the year.
• Named entities (products, companies, versions) are spelled consistently across the page.
• No claim is presented as a stat without either a linked source or a phrase like "based on our internal data covering N events".

Structured data

• A single Article or WebPage type wraps the page-level content.
• FAQPage schema is present if the page has 3 or more Q&A blocks.
• HowTo schema is present if the page has numbered steps.
• DefinedTerm schema wraps every definition on glossary and concept pages.
• SoftwareApplication schema is present on product pages with featureList populated.
• Author is a Person type with a URL to a real author page, not the organisation.
• dateModified is set and matches the last real edit.
• BreadcrumbList schema reflects the URL path.

Crawler access

• robots.txt explicitly allows GPTBot, OAI-SearchBot, ClaudeBot, Claude-Web, PerplexityBot, Google-Extended, and Applebot-Extended.
• No bot-detection challenge (Cloudflare Bot Fight Mode, Turnstile) is served to those user agents.
• The page returns a 200 with the full body in HTML without JavaScript rendering required.
• llms.txt and llms-full.txt are served at the root with a text/plain or text/markdown content type.
• The XML sitemap contains the URL with an accurate lastmod.
• Server-side rendering delivers the H1, definition, and structured data in the initial HTML response.

E-E-A-T and trust

• The page has a visible author byline with credentials.
• The author page exists at /about/authors/[slug] and includes a Person schema.
• A "reviewed by" line names a subject-matter reviewer for YMYL-adjacent topics.
• Original data or first-party numbers appear at least once on the page.
• The organisation is described consistently across the site (About page, footer, schema).
• A public changelog or newsroom demonstrates ongoing activity.

What to measure and what "good" looks like

AEO measurement rests on four numbers. Track them monthly at a minimum.

1. Citation rate. For a defined panel of prompts, the share where your domain appears as a cited source. Report per engine and as a weighted blend.
2. Mention rate. For the same panel, the share where your brand name appears anywhere in the answer, whether cited or not.
3. Answer-position for cited results. When cited, are you the primary source, secondary, or one of many? Perplexity in particular exposes ordering.
4. Prompt coverage. What share of buyer-intent prompts in your category surfaces any answer at all? Zero coverage means the model does not yet consider the query answerable.

A useful benchmark for a new AEO programme is 10% citation rate on the buyer-intent prompt panel within 90 days of shipping structured, definition-led pages and llms.txt. 25% within six months is competitive. 40% is category-leadership territory and typically requires original data, not just formatting fixes.

AEO vs classical SEO – a side-by-side

The prompt-panel methodology

You cannot measure AI visibility with a keyword list. You measure it with a prompt panel: a fixed set of natural-language questions that map to buyer intent. Build the panel once, freeze it, and re-run monthly. Rotating the panel destroys trend data.

1. Start from your product's job-to-be-done. For each job, write 3–5 questions a buyer would actually ask an LLM. "What is the best rank tracker with an API" is a prompt; "rank tracker api" is not.
2. Add category questions that do not name a vendor. "How do I track ChatGPT citations for my brand" surfaces category coverage.
3. Add competitor questions. "Alternative to Semrush for agencies" is a defensive prompt.
4. Aim for 40–120 prompts total. Fewer than 40 is noisy; more than 120 is expensive to sample at monthly cadence.
5. Sample each prompt against each engine 3 times per run to reduce variance from model non-determinism.
6. Score each response for: your domain cited (yes/no), your brand mentioned (yes/no), position of citation, presence of competitor mention.
7. Store the raw response text. Aggregate metrics change; the raw text is your evidence base for narrative reports.

The panel doubles as a product research tool. Prompts where no answer exists yet are content gaps. Prompts where competitors are cited but you are not are the highest-ROI edits.

The seven most common AEO mistakes

1. Optimising for keywords, not questions. LLM prompts are sentences. Pages structured around one keyword phrase underperform pages structured around one clear question.
2. Burying the definition. The definition of the page topic belongs in sentence one. Models truncate context. If the definition is in paragraph four, it is not the passage that gets quoted.
3. Fluffy claims. "Faster than the competition" is unquotable. "Returns rank data in under 400ms at p95 based on our internal benchmarks in June 2026" is quotable.
4. Missing dates. Undated numeric claims degrade trust and get quoted less. Every stat needs a month and year.
5. Blocking AI crawlers by accident. Cloudflare Bot Fight Mode blocks GPTBot and PerplexityBot by default in some configurations. Verify with a log audit.
6. Client-side-rendered schema. If your Article JSON-LD is injected by React after hydration, retrieval-time crawlers may miss it. Server-render structured data.
7. Treating llms.txt as a substitute for on-page structure. llms.txt lowers retrieval cost but does not replace clear definitions and schema on the pages themselves.

FAQ

Next actions

Three things to do in the next two weeks, in order:

1. Generate a first llms.txt for your site with the llms.txt Generator. Ship it at /llms.txt with text/plain.
2. Run your top 10 buyer-intent queries through the AI Overview Checker and record which of your pages are cited today.
3. Connect your site to the SEM Optimiser MCP server so Claude, ChatGPT, and Cursor can query your live rank and citation data during your team's workflow.

Who this playbook is for