How to Run an AEO Audit

An AEO audit is a structured review of whether answer engines can access, extract, trust, and cite your website. Audit retrieval before optimization: check crawler access, content structure, schema accuracy, authority signals, and measured visibility across ChatGPT, Perplexity, Gemini, Claude, and Google AI Overviews, then rank fixes by impact and effort.

Your pages can rank on page one of Google while ChatGPT recommends a competitor by name. Citation overlap also varies sharply by engine. Ahrefs research found that ChatGPT, Gemini, and Copilot showed only about 12% average top-10 overlap with Google’s top 10, meaning a site can pass a rigorous SEO audit and still be largely invisible in AI answers.

How is an AEO audit different from a traditional SEO audit?

A traditional SEO audit optimizes toward keyword rankings and organic clicks. An AEO audit optimizes toward citations: whether ChatGPT, Perplexity, Gemini, Claude, and Google AI Overviews pull your content into generated answers and name you as the source.

• Different crawler set: Training and retrieval crawlers serve different purposes. GPTBot and ClaudeBot are training crawlers. OAI-SearchBot, Claude-SearchBot, Claude-User, PerplexityBot, and Googlebot affect live retrieval or search surfaces, so audit those agents first.
• Different unit of optimization: Answer engines retrieve and rank passages within broader document-retrieval pipelines, so audit both page-level eligibility and passage-level extractability.
• Different measuring instrument: You need a prompt panel run across engines, because traditional keyword rank trackers don’t report whether Perplexity cited you yesterday. That takes a dedicated AI visibility tool.

What does an AEO audit cover?

An AEO audit covers five pillars, and together they answer two questions: can answer engines read your content, and do they choose to cite it.

• Content structure and extractability: Whether each page contains self-contained, front-loaded passages an LLM can lift into an answer, with headings that match real queries.
• Schema and structured data: Whether your markup parses cleanly, matches visible text, and covers the types worth maintaining.
• Crawler access: Whether GPTBot, OAI-SearchBot, ClaudeBot, Claude-SearchBot, Claude-User, PerplexityBot, and related fetchers can access your pages as intended.
• Authority and E-E-A-T: Whether author credentials, entity clarity, and third-party mentions give engines a reason to trust your content over a competitor’s.
• Measurement baseline: Your current citation rate, mention rate, share of voice, and sentiment across the target engines: ChatGPT, Perplexity, Gemini, Claude, and Google AI Overviews, plus AI Mode, which Google runs as a separate product.

How do you run an AEO audit step by step?

The audit runs in six steps: crawler access, content structure, schema validation, authority, visibility baseline, and competitor gap. Run them in that order. Crawler access affects direct retrieval, and the measurement steps only mean something once you understand the technical layer.

Step 1. Check crawler access and rendering

Crawler access affects direct retrieval, but engines can still cite blocked pages through prior indexes or alternate retrieval systems. Those systems include third-party search providers and user-triggered fetchers. Pull your robots.txt and check for blocks against the bots that feed each engine:

• GPTBot and OAI-SearchBot: GPTBot collects content for training, while OAI-SearchBot surfaces ChatGPT search results. A block against OAI-SearchBot removes direct access to your content for that surface.
• Anthropic crawlers: Anthropic assigns separate crawler roles. ClaudeBot handles training, Claude-SearchBot handles search, and Claude-User handles user-triggered retrieval. Blocking Claude-User prevents retrieval when a user asks a question, which may reduce your visibility in Claude’s answers.
• PerplexityBot: Perplexity uses this bot for Perplexity search results, not training, and recommends allowing it.
• Google-Extended: Google-Extended is a robots.txt control token for Gemini training and grounding. Blocking it does not affect AI Overviews or AI Mode, which are part of Google Search and follow standard Googlebot rules.

The check itself is two lines per bot. Look for a pair like User-agent: OAI-SearchBot followed by Disallow: /, and remove it unless the block is a deliberate policy choice. Propagation timing varies by provider and can take time, so re-test after making the change rather than immediately.

Rendering is the second half of access. Testing of major AI crawlers found that GPTBot, ClaudeBot, and PerplexityBot fetched static HTML without rendering client-side JavaScript. Evidence remains incomplete for Claude-User, Claude-SearchBot, and Perplexity-User, so treat server-rendered core content as the safe default.

Fetch raw HTML for key pages with curl or view-source and confirm the core commercial facts are present. Googlebot renders JavaScript through its Web Rendering Service, so a page can appear complete in Google while retrieval crawlers receive an empty shell. Move citation-critical copy into server-rendered HTML.

Step 2. Audit content structure and extractability

Structured, front-loaded passages make content easier to match and extract, but structure alone does not guarantee selection. Write concise, self-contained answers that keep each claim and its supporting evidence together. AI systems extract meaning more reliably when a point doesn’t depend on surrounding context to make sense. Put your most important claims early, and use structure as an extraction aid without assuming it guarantees citations.

Audit your top pages against that pattern:

• Headings: Phrase them as the questions buyers ask.
• Answer placement: Put the direct answer in the first one or two sentences under each heading.
• Definition block: Give each primary query a front-loaded, self-contained definition built from short declarative sentences a model can lift whole.
• Answer shape: Keep each answer concise and self-contained. Pair claims with their supporting evidence so a model can quote the passage intact without needing surrounding context.

Vague prose fails extraction for a mechanical reason: a paragraph that circles its point contains no self-contained sentence worth quoting. Audit freshness and page hygiene alongside structure. Keep timestamps accurate and break paragraphs at natural topic shifts so each one stays concise and focused. Reduce template chrome on thin pages or add enough substance to make the body copy dominant.

Step 3. Validate schema and structured data

Use schema to keep markup valid and aligned with visible content. A controlled study of 1,885 pages found no significant citation lift on Google AI Overviews, AI Mode, or ChatGPT. Google also states that AI Overviews require no additional technical work beyond indexing and snippet eligibility.

Validate Article and Organization markup on your key pages using Google’s Rich Results Test and the Schema Markup Validator. Passing markup parses without errors and describes content visible on the page. Failing markup contains parse errors, omits required properties, uses unsupported properties, or claims content the page doesn’t show, which Google’s guidance advises against.

Deprioritize FAQ and HowTo types. Google removed HowTo results from desktop and mobile by September 2023. Its 2026 search updates ended FAQ rich results in May 2026 and dropped FAQ support from the Rich Results Test. Neither type belongs at the top of your schema backlog.

Step 4. Assess authority and E-E-A-T signals

Treat documented expertise and third-party mentions as trust signals, not a published AI ranking formula. Google’s guidance says trust is the central element of E-E-A-T and explains that E-E-A-T is not a specific ranking factor. Raters compare a site’s self-description with independent reputation evidence. Separate research found earned media appeared more often than brand-owned and social content in AI search, so audit bylines, author pages, About-page clarity, and third-party coverage as distinct signals.

• Bylines: Every substantive page carries a byline with stated credentials alongside the author’s name.
• Author pages: Each author page documents real expertise beyond a job title.
• About page: The About page establishes who the entity is and what it does.
• Earned mentions: Your footprint across third-party sites and buyer communities.

Inventory branded mentions and backlinks as separate signals. Use the findings to prioritize trust signals, not to infer a published ranking formula.

Step 5. Measure your AI visibility baseline

Build a prompt panel from the buying journey and weight it by funnel stage. Panels under 50 prompts are exploratory, while a recurring program needs broader category coverage. Prompt-panel guidance suggests 200–500 prompts as a practical range, stratified by intent:

• Problem-aware prompts: The questions your buyer asks before they know solutions exist.
• Category prompts: Solution-aware queries like “best X tools.”
• Comparison prompts: Queries such as “X vs Y.”
• Brand prompts: Direct questions about your product and your named competitors.

Run the panel across ChatGPT, Perplexity, Gemini, Claude, and Google AI Overviews. Answer engines produce non-deterministic outputs, so the same prompt returns different sources on different runs. Budget multiple runs across several days.

The IAB measurement framework treats panels under 50 queries as exploratory rather than measurement and grades citations through Presence, Prominence, Portrayal, and Persuasion. Expect a thin baseline on the first run.

• Citation frequency: Cited response runs ÷ total response runs for a given prompt set.
• Prompt coverage: Unique prompts producing a citation ÷ total prompts you test.
• Mention frequency: Branded responses ÷ total responses, regardless of whether a URL appears.
• Competitive share of voice: Your brand mentions ÷ all brand mentions you track across the same response set.
• Sentiment: A qualitative signal capturing what engines say about you when they name you. Keep tone and framing qualitative and avoid reducing them to a single score.

Calculate each metric per platform before aggregating. Cross-platform comparisons are methodologically unreliable because denominators differ by tool. Keep run counts and prompt weights fixed across measurement cycles so results remain comparable.

As an optional qualitative check, log factual errors the engines make about your product. A confidently wrong answer calls for a content fix. Visibility fixes solve a different problem.

Vendors compute these metrics with different formulas. Some weight share of voice by impressions and search volume, while others count raw mentions. The same brand can therefore post very different share-of-voice numbers on two dashboards in the same week, so pick one tool and keep it for every future cycle.

Step 6. Run a competitor citation gap analysis

Run the same panel with competitor tagging. For each prompt, record which named competitors get cited or mentioned, then compute share of voice per engine. The deliverable is a list of prompts where a competitor earns citations and you’re absent. Those are your gap queries.

Compare the winning pages’ heading structure, evidence, attribution, and answer depth. Use those patterns to diagnose what your page lacks, then produce an original revision.

Then triage the list. Sort it by commercial intent first, and split what remains by why you’re losing. If the winning page beats you on structure, that is a rewrite you can ship this week. If it beats you on earned mentions, expect your team to spend quarters closing the gap. Work the structural ones first.

Citation ownership changes as answer outputs and source indexes change. A prompt a competitor owns today remains winnable if your page becomes more extractable and better attributed.

How do you read your AEO readiness score and prioritize fixes?

Readiness scores typically span four dimensions: content, technical, authority, and measurement. Use the score as a baseline for comparisons across audit cycles. No industry-standard benchmark for AI visibility exists yet, so the number only means something against your own prior runs and against competitors you measured using the same tool.

Sequence the fixes by highest impact per hour of work, and put binary technical fixes ahead of the signals that compound slowly:

• Crawler blocks: Fix these first. They’re binary, and the fix is a robots.txt edit or a server-side rendering change that improves direct access for every downstream change.
• Extractability on money pages: Retrofit answer-first structure onto your highest-intent pages before touching long-tail content. A focused, structured rewrite of a handful of pages that map to gap queries beats a shallow pass across many low-intent pages.
• Authority: Start now and expect it to compound slowly. Bylines and author pages ship in a week, but the earned-mention footprint builds over quarters.
• Schema: Fix errors and mismatches last. The cited studies found no reliable citation lift from generic schema, so schema work should not consume the budget that extractability needs.

What should an AEO audit checklist include?

Work this list top to bottom. Each block maps to a step above.

• Crawler access: Confirm robots.txt allows GPTBot, OAI-SearchBot, ClaudeBot, Claude-SearchBot, Claude-User, and PerplexityBot unless a block is a deliberate policy choice. Fetch raw HTML for your top pages and verify key content appears without JavaScript execution.
• Content structure: Review your highest-intent pages for question-phrased headings and answer-first sections built around a concise, self-contained, front-loaded definition for each primary query. Build those definitions and section openers from short declarative sentences a model can quote intact.
• Schema: Run key pages through the Rich Results Test and validator.schema.org. Fix parse errors and confirm that Article and Organization markup matches visible text. Remove markup describing content the page doesn’t show.

For a full breakdown of available options, use the AEO audit tools roundup.

• Authority: Add credentialed bylines to every substantive page, build out author and About pages, and inventory your third-party mentions separately from your backlinks.
• Measurement baseline: Build a representative prompt panel stratified by funnel stage (a recurring program might use 200–500 prompts as a starting range, but the right size depends on your topic coverage). Run each prompt across ChatGPT, Perplexity, Gemini, Claude, and Google AI Overviews with a fixed number of runs per prompt, and record results by platform using the exact formulas defined in Step 5: citation frequency, prompt coverage, mention frequency, and competitive share of voice.
• Competitor gap: Tag competitor citations on the same panel, list every prompt where an engine cites a competitor and omits you, and reverse-engineer the pages earning those citations.

Ongoing measurement begins when the audit ends. Re-run the same panel after each fix and record which prompts flip. Use How to Track Your Brand in AI Search to build a repeatable baseline, schedule consistent checks, and compare movement across engines.