Skip to content

How to run a GEO audit of your website for AI visibility

Learn how to run a GEO audit covering prompts, crawler access, structured data and citations, so you can see where your brand appears in AI answers.

Your strongest category page holds position two on Google, pulls steady organic traffic, and still goes unnamed when a buyer asks ChatGPT which vendors belong on a shortlist. Ranking and being cited have drifted apart. In an updated analysis of 863,000 keywords and 4 million AI Overview URLs, Ahrefs found only 38% of cited pages ranked in Google's top 10 for the same query, against 76% in its July 2025 study.

Use a geo audit to see where your brand stands inside AI answers and trace each gap to a cause you can fix. For a B2B content team, the procedure runs in this order:

  1. Build a prompt set and baseline your visibility.
  2. Check crawler access and rendering.
  3. Validate structured data and entity signals.
  4. Assess citability and off-site presence.
  5. Benchmark competitors and score the results.
  6. Prioritize fixes and re-test on a schedule.

What a GEO audit measures

A geo audit scores how often AI engines name and cite your brand, then explains why. Generative engine optimization (GEO) is the work of making content more visible inside answers written by engines such as ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews.

The term comes from a 2023 GEO paper published at KDD 2024, which framed GEO as black-box optimization of a source's visibility in generative engine responses. AI visibility is the outcome you measure, and the audit is the repeatable check on it.

How a GEO audit differs from an SEO audit

In an SEO audit, you check whether pages rank and earn clicks. In a geo audit, you check whether the answer text includes you, and those are different events.

A mention is your brand named in the answer. A citation is your URL listed as a source. You can earn one without the other. An engine can link your guide while recommending a competitor in the prose, or name you while citing a review site. Log both, because the fixes differ. To earn mentions, improve how people describe your brand across the web. To earn citations, make your pages fetchable and quotable.

The other difference is stability. A rank check returns the same position on repeat. AI answers vary from run to run, so sample answers many times during a geo audit rather than checking once.

The metrics a GEO audit scores

Four metrics cover most audits, and each needs one fixed definition:

  • Citation rate: Answers that cite your URL, divided by total answers tested, times 100.
  • Share of voice: Your brand mentions divided by mentions of every tracked brand in the category, times 100.
  • Sentiment: Whether each answer describes your brand positively, neutrally or negatively, reported as a share of answers that mention you.
  • Visibility score: The share of tracked prompts where your brand appears at all, so 22 appearances across 100 prompts scores 22%.

Commercial tools calculate share of voice with different denominators and weightings (by position, by search volume, by impressions). Pick your formula once and keep it across audits, or quarter-on-quarter comparisons mean nothing.

The GEO audit steps at a glance

Each step depends on the one before it, so run them in sequence:

  1. Prompts and baseline. Write buyer-style prompts, run each one repeatedly per engine, and log mentions, citations and competitors.
  2. Crawler access. Confirm AI search bots can fetch your pages through robots.txt, CDN rules and server-rendered HTML.
  3. Structured data and entities. Validate JSON-LD and check that your company is described consistently across Wikipedia, Wikidata, profiles and listings.
  4. Citability and off-site. Review whether key pages contain extractable answers, and which third-party pages engines cite instead.
  5. Benchmark and score. Compare your metrics with the competitors engines recommend, and read their cited sources.
  6. Prioritize and re-test. Rank fixes by impact, confidence and ease, then repeat the audit on a fixed cadence.

Build a prompt set and baseline your AI visibility

Before you change a single page, baseline visibility with the same prompts across ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews, adding Google AI Mode if your buyers search there. Without a baseline you can't tell a fix from noise.

Choose buyer-style prompts

Write prompts the way your buyers type them into an assistant, then group them by intent:

  • Category prompts: "Best expense management software for mid-market companies."
  • Alternative prompts: "Alternatives to [competitor] for a 200-person finance team."
  • Head-to-head prompts: "[Your brand] vs [competitor] for multi-entity accounting."
  • Problem prompts: "How do I stop employees submitting duplicate receipts?"
  • Fan-out prompts: The sub-questions an engine generates in the background, such as pricing, integrations, security certifications and implementation time.

Map each prompt to a persona and funnel stage. Problem prompts test whether your educational content gets cited, category and alternative prompts test comparison, and head-to-head prompts test validation. I'd start with 30 to 60 prompts. That's enough to cover each persona and stage without drowning the team in logging.

Set run counts to handle answer variability

Run every prompt several times per engine, because one answer is one draw. The University of St. Gallen's Don't Measure Once study, which queried ChatGPT, Perplexity, Gemini and Google AI Mode in March 2026, recommends at least seven runs per prompt per day for brand visibility and at least eight when source-level coverage matters. In the same study, identical prompts sent minutes apart often returned different sources.

Report visibility as a percentage across runs, read over rolling two- to four-week windows. Ranked position within a list is too unstable to report on its own. Record the conditions of each run as well: logged-out or logged-in session, location, model version where the interface shows it, and the date.

Log mentions, citations and competitors

A shared spreadsheet is enough for a first audit. Use one row per run with these columns:

ColumnWhat to record
PromptExact wording, plus persona and funnel stage
EngineChatGPT, Gemini, Perplexity, Claude, AI Overviews, AI Mode
Date and sessionRun date, logged-in or out, location, model if shown
Brand mentionedYes or no, plus position in the list
URL citedEvery cited URL from your domain
Competitors namedEach brand named, in order
Third-party sourcesEvery non-owned URL cited
SentimentPositive, neutral or negative, with the phrase that set it

Check AI crawler access and rendering

An engine can't cite a page its retrieval bot can't fetch or read. This step checks three layers in turn, from robots.txt to the firewall to the HTML itself.

Allow the right AI crawlers in robots.txt

Each AI company runs separate bots for search, training and user-triggered fetches, and you can treat them differently. Current tokens from the vendors' documentation:

VendorSearch or retrievalTrainingUser-triggered fetch
OpenAIOAI-SearchBotGPTBotChatGPT-User
AnthropicClaude-SearchBotClaudeBotClaude-User
PerplexityPerplexityBotNone statedPerplexity-User
GoogleGooglebotGoogle-Extended (control token)Not applicable
AppleApplebotApplebot-Extended (control token)Not applicable

OpenAI's crawler documentation says the OAI-SearchBot and GPTBot rules are independent, so you can appear in ChatGPT search while opting out of training. Websites that block OAI-SearchBot won't be shown in ChatGPT search answers, and OpenAI says changes take about 24 hours to apply. Google-Extended governs Gemini training and grounding without affecting Google Search inclusion. User-triggered fetchers such as ChatGPT-User and Perplexity-User may not follow robots.txt, because a person asked for the page.

A common policy allows search bots and blocks training bots:

User-agent: OAI-SearchBot
Allow: /

User-agent: Claude-SearchBot
Allow: /

User-agent: PerplexityBot
Allow: /

User-agent: GPTBot
Disallow: /

Confirm AI crawlers are not blocked

A permissive robots.txt means nothing if the CDN or firewall rejects the request first. Cloudflare's July 2026 changelog says that from September 15, 2026, new domains block bots classified as Training or Agent on pages that display ads, while Search stays allowed. Multi-purpose crawlers that combine search and training fall under the training block. AWS WAF's Bot Control has a CategoryAI rule that blocks AI bots by default, and Vercel's AI bots ruleset can be set to Log or Deny.

Work through four checks:

  1. Read the AI bot policy in your CDN dashboard (Cloudflare's AI Crawl Control lists each crawler's allow or block status).
  2. Search server access logs for each user-agent token and look at status codes. Repeated 403s mean a block.
  3. Compare the live robots.txt with what your CMS or CDN says it serves, since Cloudflare can manage the file for you.
  4. Fetch key pages with JavaScript turned off and confirm the main copy, pricing, navigation and JSON-LD are in the raw HTML.

The fourth check matters most for modern front ends. Vercel and MERJ's AI crawler analysis found that none of the major AI crawlers, including GPTBot, OAI-SearchBot, ClaudeBot and PerplexityBot, render JavaScript. Some download JS files and never execute them. Gemini uses Googlebot's rendering, so a client-side page can look fine in Google and still appear empty to ChatGPT. Serve critical content with server-side rendering or static generation.

Decide whether to add llms.txt

llms.txt is a markdown file at your website root, proposed by Jeremy Howard in September 2024, that summarizes the website and links to clean markdown versions of key pages. Google's AI optimization guide says Search doesn't use these files for its generative AI features. OpenAI, Anthropic and Perplexity publish llms.txt files for their own developer docs, and their crawler pages describe robots.txt controls instead. If your docs platform generates the file automatically, keep it. Don't spend audit hours on it, and don't score it.

Validate structured data and entity signals

Structured data and entity profiles tell engines which company a page describes. Google states that generative AI search doesn't require special schema, and Ahrefs' tests of pages that added JSON-LD found no meaningful lift in AI citations from schema alone, so treat markup as disambiguation and hygiene rather than a citation lever.

Test JSON-LD schema

Google's structured data guidance separates two tools. The Rich Results Test checks which Google rich results a page qualifies for. The Schema Markup Validator checks any schema.org markup without Google-specific rules. Run both on one template of each type:

  • Organization: No properties are required, but add name, logo, url, address, contact details and sameAs on the homepage.
  • Product: Add price, availability and review data so merchant listings and product snippets read correctly.
  • Article: Add author, datePublished, dateModified, headline and image to blog and guide templates.
  • FAQPage: Google's Search Central updates log says the FAQ rich result no longer appears in Google Search starting May 7, 2026, so FAQ markup now has no visible effect there.

Check where the markup loads. JSON-LD injected by a tag manager after page load is invisible to crawlers that don't run JavaScript.

Strengthen entity information beyond schema

Entity signals live mostly outside your website, so audit them as a set:

  • sameAs links: Point Organization markup to Wikipedia, Wikidata, LinkedIn, Crunchbase and your main review profiles.
  • Wikipedia and Wikidata: Confirm an item exists and that its description, founding date and product names are current.
  • Google Business Profile: Claim and complete it, since Google routes businesses to Business Profile panels for edits.
  • Consistent naming: Use the same company name, product names and one-line description across every profile.

Academic work on factual recall finds that language models answer more reliably about entities that appear often in their training text. Consistent descriptions across trusted sources give a model one version of your company to learn instead of five.

Assess content citability and off-site presence

Engines quote passages they can lift cleanly and trust sources beyond your own domain. This step reviews both sides.

Structure pages into extractable answers

The original GEO study tested content changes on GEO-bench, a large-scale benchmark of diverse user queries. Adding citations, quotations from relevant sources and statistics improved visibility by up to 40%, while keyword stuffing performed poorly. Review your most important pages for these patterns:

  • An answer in the first two sentences under each heading
  • Plain definitions of the terms buyers search for
  • Tables for comparisons, pricing and specifications
  • Statistics with a named source and a date
  • Descriptive headings that match how buyers phrase questions
  • A visible last-updated date

Score each priority page as pass or fail on these patterns, then pair it with your prompt log. A page that fails the checklist and never gets cited is your first rewrite.

Audit third-party mentions

Pull every non-owned URL from your prompt log and group it by domain. Researchers studying citations keep finding Reddit, YouTube, Wikipedia and LinkedIn near the top across engines, alongside review sites such as G2 and editorial outlets such as Forbes. Engines lean differently. ChatGPT tends toward Wikipedia and editorial sites, while Perplexity and AI Overviews draw heavily on Reddit and YouTube.

Then sort the domains into three groups. Pages that cite competitors but not you are outreach targets. Pages that describe you inaccurately need correction. Pages that cite you favorably need refreshing so they stay current.

Benchmark competitors and score the results

Benchmarking starts with the competitors engines name in your place, which may not match the ones your sales team names. Rank brands by how often they appear in your prompt log, then open the sources cited in the answers that named them. Read those sources to understand the gap. A competitor cited through three review sites and a Reddit thread has an off-site advantage, while one cited through its own glossary pages has a content advantage.

Score everyone with the metrics defined earlier. A sample scorecard with illustrative numbers:

MetricYour brandCompetitor ACompetitor BHow to interpret the gap
Visibility score24%61%38%How often you appear at all
Citation rate9%22%14%Whether your pages get used as sources
Share of voice12%34%19%Your slice of category mentions
Positive sentiment70%82%65%How answers describe you when they do

Break each row out by engine as well. A brand can lead in Perplexity and barely register in ChatGPT, and the fix differs for each.

Prioritize fixes and set a re-audit cadence

The audit only pays off when findings become a ranked plan with owners and a date to check results.

Map each finding to a fix

Score each fix on impact, confidence and ease from 1 to 10, then multiply. Sample scores:

Finding typeFixOwnerICE (I × C × E)
CDN blocks OAI-SearchBotAllow search bots in bot policyWeb ops9 × 9 × 9 = 729
Pricing page renders client-sideMove pricing to server-rendered HTMLEngineering8 × 8 × 5 = 320
Absent from top comparison articlesPitch updates with current dataPR or content7 × 6 × 6 = 252
Category page lacks a direct answerRewrite opening and add comparison tableContent6 × 6 × 8 = 288
Wikidata item outdatedUpdate description and product namesContent5 × 5 × 9 = 225

Access fixes almost always come first. An engine has to reach a page before rewrites can matter.

Repeat the audit on a fixed schedule

Re-run your core prompts monthly and run the full audit quarterly. Re-run outside that rhythm after a major model change or a significant website change such as a redesign, a CMS migration or new bot rules.

Model updates shift citations without warning. In mid-August 2026, several trackers recorded Reddit's share of ChatGPT citations collapsing within days, and OpenAI gave no explanation. Mark known update dates on your trend chart, and collect about a week of runs after a change before you rewrite anything in response.

Tools for running a GEO audit

Tools automate the prompt-running and logging steps. Crawler access, rendering, schema and fix prioritization stay with your team.

Paid platforms run prompts on a schedule and calculate visibility metrics. Their prompt allowances are the main limit, because each prompt is usually checked once per day per engine.

Semrush's AI Visibility Toolkit costs from $99/mo. It tracks custom prompts daily, reports mentions from ChatGPT, Google AI, Gemini and Perplexity, and includes a site audit for AI readiness.

Ahrefs Brand Radar is an Ahrefs feature. Prompt tracking costs from $50/mo and the AI Visibility Index from $199/mo. Its indexes cover ChatGPT, Perplexity, Gemini, Copilot, AI Overviews and AI Mode, and custom prompts add Claude, which uses 8 checks per update.

OtterlyAI lists plans from $29/mo. Its base engines are ChatGPT, Google AI Overviews, Perplexity and Microsoft Copilot, with Claude, Gemini and Google AI Mode as add-ons.

Free checkers and how far to trust them

Free checkers run a handful of prompts once and return a score. Given how much answers vary between runs, that's a snapshot. Use one to spot an obvious problem, such as a brand that never appears, and don't put the score in a report.

Run a manual audit with ChatGPT

You can run a useful first audit with no tools at all. ChatGPT can fetch a page you paste and critique its structure, but it can't see your CDN rules or server logs, and your own logged-in session is one personalized run. Treat it as one engine in the sample.

  1. Open a logged-out or temporary chat for each run.
  2. Paste each prompt from your set word for word.
  3. Log mentions, citations, competitors and sentiment in the spreadsheet.
  4. Repeat each prompt at least seven times across days.
  5. Repeat the same steps in Gemini, Perplexity, Claude and Google.

Pair the log with traffic data. Google's generative AI performance report in Search Console shows impressions from AI Overviews and AI Mode and rolled out to all websites on August 31, 2026. In GA4, check the AI Assistant channel, or build a custom channel group under Admin, then Data settings, then Channel groups. Use a regex such as chatgpt\.com|chat\.openai\.com|perplexity\.ai|claude\.ai|gemini\.google\.com and place the rule above Referral. Expect some AI visits to land as Direct, because apps and copied links often strip the referrer.

A worked GEO audit example

Here's a compressed audit for a hypothetical project management vendor called Acme. All numbers are illustrative. The prompt set covered five intents:

  • "Best project management software for agencies" (category)
  • "Alternatives to [Competitor A] for a 50-person agency" (alternative)
  • "Acme vs [Competitor A] for client reporting" (head-to-head)
  • "How do I track billable hours across client projects?" (problem)
  • "Does Acme integrate with QuickBooks?" (fan-out)

Each prompt ran eight times per engine over two weeks, for 40 runs per engine:

EngineAcme mentionedAcme URL citedMost cited third-party source
ChatGPT10 of 402 of 40Comparison article on a software blog
Perplexity18 of 409 of 40Reddit thread on agency tools
Gemini14 of 406 of 40G2 category page
Claude8 of 401 of 40Competitor A's documentation
AI Overviews16 of 407 of 40YouTube review

The crawler check explained the weak ChatGPT and Claude results. The CDN blocked OAI-SearchBot and Claude-SearchBot, and the integrations page loaded its content through JavaScript. The ranked fix list followed:

  1. Allow OAI-SearchBot and Claude-SearchBot in the CDN bot policy.
  2. Server-render the integrations page and add a QuickBooks answer in the first paragraph.
  3. Pitch an update to the comparison article ChatGPT cites most, with current pricing.
  4. Refresh the G2 profile and fix the outdated product name on Wikidata.
  5. Re-run the full prompt set four weeks after fixes ship.

Once the audit becomes a monthly habit, the hard part is keeping prompt logs, page fixes and search data in one place instead of five spreadsheets. GrowthX brings shared company context, portfolio priorities and production into one place, with your team making the calls. To see it on your own pages, book a demo

Related articles

  1. Google November 2026 core update: What is confirmed and what to monitor Learn · 12 November 2026
  2. Can B2B and service businesses use Businesses in ChatGPT today? Learn · 10 November 2026
  3. Guide to Businesses in ChatGPT: From Business Hub to Business Agent Learn · 10 November 2026
  4. How to measure your business's visibility with ChatGPT Business Hub Learn · 10 November 2026
  5. How to claim and verify your business in ChatGPT Learn · 10 November 2026
  6. How to create, test, and publish a Business Agent in ChatGPT Learn · 10 November 2026
  7. Keeping your business details accurate in ChatGPT, step by step Learn · 10 November 2026
  8. ChatGPT AI search console: What OpenAI’s Business Hub measures Learn · 9 November 2026