GEO Audit Step by Step: From Diagnosis to Action Plan

A GEO audit checks, in six blocks, whether AI systems can access, understand, trust and cite your site: crawlers, content, entity, authority, off-site presence and measurement. The end product is a prioritized spreadsheet and a report with an action plan. It complements the SEO audit. The big picture is in SEO, AEO and GEO: what is the difference.

Quick summary

• Start with a baseline: 20 to 30 questions, several runs each, recording appearance, cited sources and brand description.
• Each assistant has separate bots for search, training and user actions. Blocking the wrong one takes you out of search or changes nothing.
• Audit 10 to 20 priority pages, not the whole site: direct answer, data with source and date, original perspective.
• Entity and off-site presence decide whether AI has anywhere to get facts about you besides your own site.
• Prioritize by impact, reach and effort. Technical blocks come first, because a page the bot cannot read cannot be cited.

How to read this. If you are starting out in GEO, read the step-by-step and the report template; the GEO, AEO and SEO checklist is a simpler version of this process. If you audit client sites, the sections on crawlers, measurement and prioritization have the edge cases and the statistics. The terms are in the SEO, AEO and GEO glossary.

What does a GEO audit cover that an SEO audit does not?

It keeps the SEO foundation (indexing, content, links) and adds four things: control by AI bot, analysis of citations and brand description in answers, the entity's presence outside the site and measurement on platforms other than Google. There is no official standard for a GEO audit, so the method below is our own proposal, backed by public documentation.

Block	Question it answers	Evidence
1. Access and crawlers	Do the right bots read the right pages?	robots.txt, logs, firewall rules, rendered HTML
2. Content	Does the page answer, prove and have something of its own?	Review of 10 to 20 pages scored 0 to 2
3. Entity	Is it clear who the brand is and who signs the content?	About page, authors, schema, profiles
4. Authority	Do relevant sites point to you?	Link profile and comparison with competitors
5. Off-site presence	Do third parties talk about you where your audience asks?	Sources cited by AI, reviews, forums
6. Measurement	Can you tell whether anything changed?	Search Console, Bing, test runs

In plain English: it is like a car inspection with six items: the engine starts (access), the dashboard reports correctly (content), the paperwork is in order (entity), the car has a good reputation (authority), the neighbors speak well of it (off-site presence) and the odometer works (measurement).

How do you build the baseline before changing anything?

Without a baseline, every improvement is an anecdote. Run the questions before any fix and keep the spreadsheet.

1. Write 20 to 30 questions in three groups: brand ("what is company X?"), category ("best tools for Y") and decision ("X vs Z", "how much does X cost"). Use the customer's vocabulary, not yours. The guide to prompts to test your brand in AI helps.
2. Run each question several times in each assistant that matters (ChatGPT, Perplexity, Google AI Mode, Copilot), in sessions with no history.
3. Record whether the brand appears, in which position of the list, which URLs were cited as sources and how the brand was described.
4. Calculate the frequency of appearance per question (appearances ÷ runs).

Why several runs? In a SparkToro study with 2,961 runs on ChatGPT, Claude and AI Overviews, the chance of repeating the same list of brands was below 1 in 100, and the same order below 1 in 1,000. The authors acknowledge there was no peer review and propose appearance percentage as a more useful metric than position.

For SEO professionals: a round of 10 gives little precision. If the brand appears in 3 of 10, the 95% (Wilson) confidence interval runs from about 11% to 60%. With 100 runs, 30 appearances give about 22% to 40%. Our rule of thumb is 10 runs per question and assistant for triage and 30 or more when comparing before and after. It is not a rule from any cited source. Also control for location, language, model and date.

Block 1: can the right bots read the right pages?

The answer is in robots.txt, the firewall and the logs, bot by bot. The main assistants separate search, training and user-action crawlers, and the rule for each one is different.

Bot	Documented function	robots.txt control
OAI-SearchBot (OpenAI)	Show sites in ChatGPT search	Yes; blocking it keeps you out of search answers
GPTBot (OpenAI)	Model training	Yes; blocking it does not remove you from search
ChatGPT-User (OpenAI)	User-initiated actions	May not apply
PerplexityBot	Show and link sites in results	Yes; Perplexity recommends allowing it
Perplexity-User	User requests	Generally ignores robots.txt
Claude-SearchBot (Anthropic)	Search result quality	Yes
ClaudeBot (Anthropic)	Model training	Yes
Google-Extended	Gemini training and grounding	robots.txt token; does not affect Search or AI Overviews

Sources: bot documentation from OpenAI, Perplexity, Anthropic and Google's list of crawlers. To decide case by case, see robots.txt and AI crawlers.

# Example: appear in ChatGPT search without allowing training
User-agent: OAI-SearchBot
Allow: /

User-agent: GPTBot
Disallow: /

Test access in two layers: curl -I -A "OAI-SearchBot" https://yoursite.com/page shows whether the server responds to that user agent, but WAFs and CDNs also filter by IP and reputation. So confirm in the logs that the bots get a 200. Anthropic advises against blocking by IP, because that can stop the bot from reading robots.txt, and Perplexity publishes IP ranges for allowlisting.

Also review: improper noindex and canonical tags, restrictive nosnippet or max-snippet, content that only appears after a click or heavy JavaScript, paywalls and pages behind a login. Without indexing and snippet eligibility, the page does not take part in Google's AI features.

Block 2: does the content answer, prove and have something of its own?

Evaluate 10 to 20 priority pages (the ones that generate revenue or answer the baseline questions). For each one, score 0 (absent), 1 (partial) or 2 (ok):

• Direct answer in the first 40 to 70 words.
• Headings written as questions, with the answer right below.
• Data with a linked source and date; no loose numbers.
• Original perspective: a test, data, a real example, a criterion. Text that repeats what everyone says gives no reason to be cited.
• Table or list where there is a comparison or steps.
• Visible update date and reviewed content.
• Identified author, when the topic calls for it.

The pattern is detailed in LLM-citable content.

Block 3: is it clear who the brand is and who signs the content?

Check that the brand name, description and contact details are identical on the site, on Google, on social networks and in directories; that an "About" page exists with verifiable information; that authors have a bio and real credentials; and that there is Organization and Person markup with sameAs pointing to official profiles. The markup only helps when it matches what the page shows. See Organization, Person and sameAs.

In plain English: the entity is the brand's "ID card" on the internet. If each profile says something different (name, address, what the company does), the system does not know which version is the true one.

Block 4: do relevant sites point to you?

Compare your link profile with that of 3 to 5 competitors that appear in the baseline answers. Look for relevant domains that link to them and not to you, using link gap analysis, and check whether your backlink profile looks natural. Backlinks do not guarantee citation, and a correlation study does not prove cause. Google treats links bought to manipulate ranking as a policy violation, so record the origin of each link.

Block 5: do third parties talk about you where your audience asks?

The most useful way to audit this is to look at the cited sources in the baseline. List the domains AI uses as support in the questions of your category and mark where your competitors are and you are not: comparisons, review sites, forums, video channels, niche press. That is your list of presence gaps. See the channels in Reddit, YouTube and Quora in AI citations. Avoid fabricated "mentions": Google says that chasing inauthentic mentions on the web is not as useful as it seems.

Block 6: can you measure whether anything changed?

Set up three data sources before you start the fixes:

• Search Console: since June 2026 there are generative AI performance reports, with impressions, pages, countries, devices and dates, rolled out to all sites by August 31, 2026. According to Google's announcement, the report covers Google's features (AI Overviews and AI Mode), and the text does not mention clicks or queries in these reports.
• Bing Webmaster Tools: the AI Performance report, in public preview since February 10, 2026, shows total citations, average cited pages, grounding queries and per-page activity, in Copilot and Bing summaries, according to Microsoft's announcement.
• Test runs and logs: for ChatGPT, Perplexity and Claude we know of no site-owner report. Use the baseline repeated monthly and server logs. See AI visibility metrics.

How do you record the findings in the spreadsheet?

One row per finding, with evidence. The minimum spreadsheet has these columns:

Block	Item	Evidence (URL, screenshot, log)	Score 0 to 2	Impact 1 to 3	Reach 1 to 3	Effort 1 to 3	Owner	Deadline
1	OAI-SearchBot blocked by the WAF	Log with 403	0	3	3	1	Infra	7 days
2	Pricing page without a direct answer	URL	1	2	2	1	Content	14 days
5	Absent from comparisons cited by AI	List of domains	0	3	2	3	Marketing	90 days

The data above is a hypothetical example. Calculate priority like this:

priority = (impact x reach) / effort

In a tie, block 1 goes first: if the bot does not read the page, nothing else has any effect.

How do you deliver the report and the action plan?

Deliver a short document with clear decisions. The template below fits in 4 to 6 pages.

1. Executive summary (3 to 5 lines): situation, 3 priorities, risk
2. Baseline: table of questions x assistants, with appearance frequency
3. Findings by block: item, evidence, score, recommended action
4. 30/60/90-day plan: who does what and how it will be measured
5. Limits: what could not be verified and why
6. Next measurement: date, questions, number of runs

In the 30/60/90 plan, put crawler blocks and indexing fixes in the first 30 days, pages and entity in the 60, and off-site presence and authority in the 90, which are the slowest actions. No report should promise citation. Promise what is under your control: fixes made, data collected and a new measurement on the scheduled date.

Frequently asked questions

How long does a GEO audit take?

By our estimate, on a small or medium site, 1 to 3 weeks, most of it spent building and repeating the baseline. The time grows with the number of assistants, languages and priority pages. Auditing the whole site rarely pays off; focus on the pages that generate revenue or answer the target questions.

How often should I repeat the audit?

Redo the baseline measurement every month, with the same questions, and the full audit every six or twelve months, or when there is a major change to the site, the product or the bot policy.

Do I need a paid tool to audit?

Not to get started. Search Console, Bing Webmaster Tools, logs, a spreadsheet and manual runs cover the six blocks. Tools help automate the runs and record cited sources, which saves hours of manual work when there are many questions. Compare options in tools to monitor AI citations.

Can I block GPTBot and still appear in ChatGPT?

According to OpenAI, yes: GPTBot handles training, and OAI-SearchBot handles ChatGPT search. If you block OAI-SearchBot, you do not appear in search answers. ChatGPT-User, used for user actions, may not follow robots.txt. Review the documentation, because names and rules can change.

Does the audit guarantee AI will cite me?

No. It removes barriers and improves the quality of the sources you offer, but the decision to cite belongs to the system, with criteria companies do not fully publish. Measure the change in appearance frequency over time and do not promise results.

Next step

If you do not have a measurement routine yet, start with AI visibility metrics and close the block 1 blockers this week. For a leaner version, use the GEO, AEO and SEO checklist.

Read also:

• SEO audit step by step
• robots.txt and AI crawlers
• Organization, Person and sameAs