How-tos · November 7, 2026 · 15 min read

GEO Audit: How to Run an AI Visibility Audit in an Afternoon

A GEO audit checks whether AI engines can reach and read your site, and how they answer about your brand. Steps, a checklist and a scorecard for one afternoon.

A GEO audit, also called an AI visibility audit, checks two things: whether AI engines can find, read and cite your site, and what those engines actually say about your brand when buyers ask about your category. It is a one-time measurement, compared with your competitors, that ends in a short list of actions. The first half is a technical check of crawler access, indexing and page structure. The second half samples the answers: you ask each assistant a fixed set of buyer questions several times and record who it names and which sources it cites. One person can do both halves in an afternoon with a spreadsheet.

This guide gives the steps with time estimates, a checklist, a scorecard template and the report outline.

What you need before you start

  • A spreadsheet with one row per answer.
  • Access to the assistants your buyers use. ChatGPT, Perplexity, Gemini and Google AI Overviews cover most categories. Add Claude, Copilot or Google AI Mode if your buyers use them.
  • A clean session for each assistant: a temporary or logged-out chat, with memory and custom instructions off. Your own history changes the answers you get.
  • Access to your robots.txt file, Google Search Console and Bing Webmaster Tools.

The time estimates below assume 20 prompts, 4 competitors and three assistants. Scale them to your own set.

Step 1: check that AI engines can reach and read your site (30 minutes)

An engine cannot cite a page it cannot fetch. Start with the site, because a block here makes the rest of the audit explain itself.

Crawler access. Open your robots.txt file and look for rules that name AI crawlers. Each company documents its own, and most separate search crawlers from training crawlers:

CrawlerCompanyWhat it does, per the company’s documentation
OAI-SearchBotOpenAISurfaces sites in ChatGPT search
GPTBotOpenAICrawls content that may be used to train models
Claude-SearchBotAnthropicIndexes content to improve search results for Claude users
ClaudeBotAnthropicCollects content that may be used to train models
PerplexityBotPerplexitySurfaces and links sites in Perplexity search
GooglebotGoogleCrawls for Google Search, which includes AI Overviews and AI Mode
Google-ExtendedGoogleControls use in AI training and grounding in some of Google’s other systems, not Search

Sources: OpenAI, Anthropic, Perplexity and Google, checked November 2026. OpenAI says sites that block OAI-SearchBot are not shown in ChatGPT search answers. You can block the training crawlers and still allow the search crawlers.

Check your CDN and firewall as well. Google’s documentation asks site owners to make sure crawling is allowed in robots.txt and by any CDN or hosting infrastructure. A bot rule at the edge can block a crawler that robots.txt allows.

Indexing. Google says a page must be indexed and eligible to show in Google Search with a snippet to appear as a link in AI Overviews or AI Mode. Check your key pages in Google Search Console, and check that no nosnippet or noindex sits on a page you want quoted. Then check Bing Webmaster Tools. OpenAI’s ChatGPT search help page lists Microsoft among the search providers ChatGPT search works with, so a page Bing has not indexed may start at a disadvantage there.

Page structure. Open your homepage, pricing page and one comparison or use-case page. For each one, check:

  • The main text is in the HTML, not only drawn by scripts after load.
  • The first paragraph states what the product is and who it is for.
  • Prices, plans and limits are written in text, not only in images.
  • Headings use the words a buyer uses.

Note what fails. You do not need new files or special markup for Google’s AI features: Google says so on the same documentation page. Fix access and indexing first.

Step 2: write 15 to 25 prompts (30 to 45 minutes)

Now the answer half. Write the questions a buyer asks before they know your name. Spread them across intent types, because each type tests a different part of your reputation.

Intent typeExample shapeWhat it tests
Category“best help desk for a small SaaS team”Whether you are in the default shortlist
Use case“help desk that works inside Slack”Whether the assistant knows your strengths
Problem“how do I stop support tickets getting lost”Whether you appear before the buyer looks for a tool
Comparison“Rival A vs Rival B”Whether you enter a conversation about others
Alternative“alternatives to Rival A”Whether you are the named substitute
Brand“is YourBrand good for a 5-person team”Whether the description of you is accurate

Keep most prompts unbranded. A set full of your own brand name measures recall of your name, not discovery. Our guide to checking if ChatGPT recommends your brand has more on phrasing prompts like a buyer.

Freeze the set when you finish. If you change a prompt halfway through, the numbers before and after the change no longer compare.

Step 3: pick 3 to 5 competitors (10 minutes)

Choose the rivals a buyer compares you with, not the biggest names in the market. Include one you suspect wins in AI answers today. Write down each brand name, its common aliases and its domain. You need the domain to count citations.

Step 4: run each prompt several times on each assistant (90 to 120 minutes)

Ask each prompt at least three times on each assistant, in a fresh session each time. This is the step people skip, and it is the step that makes the audit worth reading.

Repeats matter because assistants do not return a fixed answer. The same prompt can name four brands at one moment and six brands a minute later, in a different order, with different sources. One answer is an anecdote. Three answers show whether a mention is stable or lucky. This is prompt sampling, and every serious AI visibility tool works this way.

The arithmetic for the example: 20 prompts, 3 runs, 3 assistants is 180 answers. At 30 to 40 seconds each to ask and record, that is about two hours. If that is too much, cut prompts before you cut repeats.

Step 5: record four things per answer (during step 4)

For every answer, fill one row:

  1. Named or not. Does the answer text name your brand? Do the same for each competitor.
  2. Position. Among the tracked brands named (you and your competitors), where does yours appear? 1 means first.
  3. Description accuracy. Is what the assistant says about you correct, outdated or wrong? Note the wrong claims word for word. They become actions.
  4. Sources cited. Which domains does the answer cite? Mark the ones that are yours or a competitor’s.

Keep “named” and “cited” in separate columns. An answer can cite your site as a source and never name you, and an answer can name you with no link to you at all. They are different results, and they need different fixes.

PromptAssistantRunNamedPositionAccuracySources cited
best help desk for a small SaaS teamChatGPT1Yes3Correctg2.com, rivala.com
best help desk for a small SaaS teamChatGPT2Nocapterra.com, rivalb.com

The two rows above are illustrative. They show the format, not real results.

Step 6: compute the five metrics (30 minutes)

Compute each metric per assistant and overall, for you and for every competitor.

  • Visibility rate: answers that name you, divided by all answers.
  • Share of voice: answers that name you, divided by answers that name any tracked brand.
  • Average rank: your mean position among the tracked brands named, over the answers that name you.
  • First-mention rate: answers where you are the first brand named, divided by all answers.
  • Citation share: citations that point at your domain, divided by all citations.

Put them in one scorecard. Use this template, with one block per assistant and one overall block.

BrandVisibility rateShare of voiceAverage rankFirst-mention rateCitation shareAnswers (n)
You
Rival A
Rival B
Rival C

Always fill the last column. A rate without its sample size cannot be judged. Below about ten answers, treat any percentage as a direction, not a measurement.

An honest note on “AI visibility scores”

Many tools and reports compress all of this into one composite number: “your AI visibility score is 62”. It is easy to put on a slide. It is also where most of the useful detail disappears.

A single score hides two things. It hides the per-engine split: you can lead on Perplexity and be absent from Google AI Overviews, and a blended number shows neither fact. It also hides the per-prompt split: a good score can sit on top of three category prompts you never appear in, and those three prompts are often the ones that matter most.

If someone gives you a score, ask two questions. First, what goes into it, and with what weights? Second, how many answers is it based on, per engine and per prompt? If neither answer is available, the score is decoration. Report the five metrics instead, and keep the raw answers.

What the audit report should contain

The report is short. Most readers stop after the first page, so put the findings first.

  1. Findings. Three to five sentences. Where you stand overall, where you lead, where you trail and which competitor wins most often.
  2. Access and indexing issues. Anything step 1 found: blocked crawlers, pages missing from an index, key facts that exist only in images.
  3. Gaps by prompt. A list of the prompts where a competitor is named and you are not, sorted by how often the competitor appears. This is the most useful table in the report.
  4. Source map. The domains the assistants cite for your category, grouped by type: review sites, communities, comparison articles, publishers, your own site, competitor sites. Mark which ones mention you today.
  5. Description errors. The wrong or outdated claims, quoted, with the source the assistant likely took them from.
  6. Three to five actions. Each one tied to a gap, a source or an access issue. For example: “get listed on the comparison article cited in 9 of 12 answers to our two biggest gap prompts”. Not “improve content”.

Attach the raw answers. They are the evidence, and they answer the question “are you sure?” faster than any chart. For the work that turns these actions into results, see our playbook for improving AI visibility.

The GEO audit checklist

Use this list to check that the audit is complete.

  • robots.txt allows the search crawlers of the engines your buyers use.
  • No CDN or firewall rule blocks those crawlers.
  • Key pages are indexed in Google and Bing, with no nosnippet or noindex on pages you want quoted.
  • Key pages state what you are, who it is for and what it costs, in text.
  • 15 to 25 frozen prompts, mostly unbranded, across intent types.
  • 3 to 5 competitors, with aliases and domains.
  • At least three runs per prompt per assistant, each in a fresh session.
  • Named, position, accuracy and cited sources recorded for every answer.
  • Five metrics per assistant and overall, each with its sample size.
  • Report with findings, access issues, gaps, source map, errors and three to five actions.
  • Raw answers attached.

Audit vs ongoing tracking

An audit is a photo. It shows your position on one afternoon. Answers drift after that: model updates change which brands an assistant prefers, retrieval indexes refresh, and new articles get cited. An audit cannot tell you whether a change in the next month is your work, a competitor’s work or the model.

Use the audit as a baseline. Then either repeat it on a fixed schedule with the same prompt set, or move to a tracker that runs the same prompts every day. The prompt set from the audit becomes the tracking set, so the baseline and the trend line compare. For what a tracker can and cannot measure, see what AI visibility tools actually measure.

How CitePrism does the answer half in a first run

CitePrism automates steps 4 to 6. You add your brand, its aliases and domain, your competitors and your 15 to 25 prompts. The first run asks every prompt on each engine you enable. All seven are on every plan: ChatGPT, Claude, Gemini, Perplexity, Google AI Overviews, Google AI Mode and Microsoft Copilot. It stores every raw answer and scores each one for named, position and cited sources, for you and for each competitor by the same rules.

After that first run you have the per-answer records, the overall and per-engine metrics, the Gaps view (prompts where a rival is named and you are not) and the cited sources, grouped by type. Every chart clicks through to the stored answers behind it. Per-prompt rates stay greyed out until each one has at least ten answers behind it, because a rate on three answers is not a rate. Scheduled runs then turn the photo into a trend line, at separate hours of the day.

The GEO tool page explains the full workflow, and AI citation tracking covers the source map. The free plan runs on your own keys: model keys for the four API engines, plus a key for the engines that have no API, priced per request by its provider. Sampling costs about $0.046 per prompt across the four API engines on the models we measured (measured August 2026), paid to the model providers. Plans and limits are on the pricing page.

FAQ

What is a GEO audit?

A GEO audit checks whether AI engines can find, read and cite your site, and how they answer buyer questions about your brand and its competitors. The first half covers crawler access, indexing and page structure. The second half samples the answers of assistants such as ChatGPT, Perplexity and Google AI Overviews, and turns them into metrics and a short list of actions.

Is there a free GEO audit tool?

Free tools cover the first half well. Many check your robots.txt rules, page structure and markup at no cost. The answer half is harder to get for free, because it needs repeated sampling of each assistant, and one check of one answer can tell you the opposite of the truth. You can do that half by hand with this guide and a spreadsheet. CitePrism Free runs the sampling on your own API keys, so you pay only the providers of those keys.

How many prompts does an AI visibility audit need?

Use 15 to 25. Fewer than 15 leaves whole intent types untested, so one lucky prompt moves the result too much. More than 25 makes a manual audit too slow to finish in a day, and the extra prompts are usually variations of ones you already have. Cover each intent type at least twice.

How many times should I ask each prompt?

At least three times per assistant for a manual audit, each time in a fresh session. Assistants vary their answers from one request to the next, so a single answer cannot tell a stable mention from a lucky one. For per-prompt rates you can trust, you need about ten answers per prompt and assistant, which is why ongoing tracking accumulates them over days.

What is a good AI visibility score?

There is no universal benchmark, and a single composite score hides the per-engine and per-prompt detail that tells you what to do. Compare your share with each rival’s in the same prompt set. Leading the tracked brands matters more than any fixed figure. Shares across brands can sum above 100 percent, because one answer can name several brands.

How often should I repeat the audit?

Repeat it at least once a quarter with the same prompt set, and after any major model release in the assistants your buyers use. If you are actively working on gaps, quarterly is too slow to see what worked. Daily tracking with the same prompts gives you the trend line and keeps the audit as its starting point.

See where you stand in AI answers

CitePrism tracks whether ChatGPT, Claude, Gemini, Perplexity, Google AI Overviews, Google AI Mode and Microsoft Copilot mention your brand, who they recommend instead, and which sources they trust. Free to start with your own keys, or Pro from $29 a month on ours.