The AI Visibility Audit: A Complete Walkthrough

An AI visibility audit measures whether AI search tools can find your brand, understand what it offers, and include it accurately in answers. It checks more than rankings: you need to test prompts, citations, crawl access, brand identity and the content that AI systems can retrieve. This guide takes you through those checks and shows how to turn the findings into a prioritised action plan.

If you need a broader introduction to AI search and related resources, explore the SearchFIT resources hub. The core idea is simple: treat visibility as something to observe across several answer engines, not as one score or one ranking.

Table of contents

• What an AI visibility audit checks
• Set the scope and baseline
• Pillar 1: Check crawl access and bot governance
• Pillar 2: Check entity clarity and brand facts
• Pillar 3: Build and test a prompt basket
• Pillar 4: Map citations and third-party coverage
• Pillar 5: Find content gaps and improve extractability
• Connect AI visibility to traffic and business outcomes
• Prioritise fixes and repeat the audit
• Common audit mistakes
• Next steps

What an AI visibility audit checks

A traditional SEO audit typically examines whether search engines can crawl and index a site, how pages perform in search results, and where technical or content improvements may help. An AI visibility audit includes some of those checks, but asks a different set of questions too:

• Does an answer engine mention the brand for relevant questions?
• Is the brand recommended, or merely named?
• Is the description accurate and current?
• Which pages and third-party sources appear in the answer?
• Can relevant crawlers access and interpret the site?
• What questions are competitors answering that the brand is not?

AI answer engines do not all retrieve or cite information in the same way. A study by 5W Public Relations found that only 11% of domains appeared in both ChatGPT and Perplexity for the same prompt sets. That is a useful reminder: a result from one engine is not a substitute for checking the others.

An audit is not a guarantee that a model will mention you. Answers can vary with the prompt, the session, the engine and the available sources. The goal is to identify repeatable patterns, correct clear problems, and track whether those patterns change over time.

The five audit pillars

Treat these as connected checks. A good page cannot help if a crawler cannot retrieve it. A crawlable website may still be overlooked if the brand is difficult to distinguish or the answer is missing from the sources an engine uses.

Set the scope and baseline

Before collecting answers, decide what the audit is meant to explain. A SaaS company assessing vendor discovery has different questions from an online retailer checking product recommendations. An agency auditing several clients also needs a consistent method for comparing accounts.

Choose your engines

Select the answer engines that matter to your audience and your reporting. A practical starting set is ChatGPT Search, Google AI Overviews, Perplexity, Claude and Gemini. You may not have equivalent access to every experience in every market, so record exactly which product and search mode you test.

Do not combine everything into a single result called “AI search”. Report each engine separately, then show an overall view only if you explain how it was calculated. Different engines can surface different sources and give different answers to the same question.

Choose markets, products and competitors

Write down the limits of the audit before testing. Include:

• Markets and language: Specify the country, language and, if relevant, region. Results can differ by location.
• Business lines: Decide whether you are auditing the whole company, one product, one range or one service area.
• Audience: Name the buyer or user whose questions you are testing.
• Competitors: Include direct alternatives and any brands buyers commonly compare with you.
• Time period: Record when the test was run and set a planned date for the next check.

For a B2B SaaS brand, the scope might be one product category, a particular buyer type and a short list of competitors. For e-commerce, it might be a product category, key use cases and comparable products. Keep the first audit manageable. You can expand it once you know which questions and markets matter most.

Save the baseline evidence

For each test, keep the prompt, engine, date, answer, cited URLs and any relevant settings. Save the full answer, not just a yes-or-no mention. A brand might appear in a list but be described inaccurately, recommended with a qualification, or cited only through an old page.

A spreadsheet is enough for an initial audit. Use one row per prompt and engine, and keep the raw answer in a separate field or linked record. If you use a tracking platform, check that it preserves the prompt, answer text, date and citations so you can inspect the evidence behind its metrics.

Pillar 1: Check crawl access and bot governance

Start with the technical question: can relevant systems retrieve the pages you want them to use? Crawl access is not the only factor in visibility, but it is a practical prerequisite for real-time retrieval of a page.

Review robots.txt carefully

Inspect the site's robots.txt file and look for broad rules that might block important crawlers or paths. A blanket disallow rule may affect more than the team intended. Check the rules against your publishing goals and ask the technical owner to confirm what each rule does.

Also distinguish between crawlers used for model training and those used for search or user-requested retrieval. For example, OpenAI identifies GPTBot separately from OAI-SearchBot, and Anthropic identifies ClaudeBot separately from Claude-SearchBot. Blocking a training crawler does not automatically mean a search crawler is blocked, and the reverse is also true. Verify current crawler names and policies against the provider’s own documentation before changing production rules.

If you want a walkthrough of the broader technical audit process, Semrush's AI visibility audit guide is one reference to compare with your own checklist.

Check more than robots.txt

A robots.txt file expresses crawl preferences. It does not prove that a crawler can successfully load a page. Check server and CDN logs for requests, response codes and repeated failures. A web application firewall, bot-management setting, rate limit, authentication wall or security challenge can interfere even when the robots file allows access.

Look for:

• Repeated 403 or 429 responses on important pages.
• Bot requests that receive a challenge or an empty page.
• Access rules that differ between the main website and a subdomain.
• Pages that require a login, cookie choice or region selection before useful content appears.
• A sitemap that omits current priority pages or lists redirected and retired URLs.

Do not assume every unfamiliar request is an AI search crawler. Confirm user-agent strings and request patterns, and involve whoever manages your hosting or CDN. Avoid weakening security controls just to allow a request that has not been verified.

Test rendering and page delivery

Open important pages with JavaScript disabled or inspect their rendered output using your normal technical tools. Check whether the main text, product details, headings and internal links are present in the delivered page. If the content only appears after complex client-side actions, retrieval systems may not see the same information a person sees in a browser.

Check mobile and desktop templates, too. A pricing page, product page or comparison page might render differently across devices or markets. Look for content hidden behind tabs, accordions or scripts, and make sure the essential answer is not trapped in an interface that fails to load.

Review page speed and server responses as part of normal technical hygiene, but do not treat a single speed threshold as a universal AI crawler rule. Test actual responses and compare them with what your users and search engines can access.

Make a crawl access register

Keep an inventory of the routes that matter and the status of each check. For example:

This makes the next step clear. A crawler issue should have a page, evidence, owner and a decision. Do not unblock private, customer or sensitive pages in pursuit of visibility.

Pillar 2: Check entity clarity and brand facts

An entity is a distinct person, company, product or place that sources describe. AI systems need to connect references to the right entity before they can reliably combine information about it. The audit question is whether your own site and credible external sources consistently explain who you are and what you offer.

Write a source-of-truth profile

Create a short internal fact sheet for the brand. Include the official name, product names, website, headquarters if relevant, category, intended customers, core use cases, key features and approved descriptions. Add facts that are often confused, such as a renamed product, a discontinued plan or a distinction between the parent company and its product.

Use this to compare what AI answers say with what the business currently offers. Mark statements as correct, outdated, unsupported or ambiguous. Do not quietly accept a wrong claim because it sounds favourable. Inaccurate product or pricing information can mislead potential customers and damage trust.

Check consistency across first-party pages

Review the pages people and systems are most likely to use as references:

• About and company pages.
• Product and category pages.
• Pricing and plan details.
• Help centre and documentation.
• Leadership and author profiles, where relevant.
• Press or company information pages.

Make names and descriptions consistent. If a product has changed name, explain the change plainly. If the company serves several audiences, describe those audiences on the relevant pages instead of relying on a broad corporate statement.

Review structured data as supporting evidence

Structured data is code that labels information on a page, such as a business, product, person or article. Check that the markup accurately reflects visible page content and uses appropriate types. Depending on the site, that may include Organization, Product, SoftwareApplication, Article or Person markup.

Use sameAs links only for official profiles that genuinely represent the same organisation or person. Examples might include a verified professional profile or an official social account. Do not add unrelated profiles to make a list look larger. Structured data can help clarify a page, but it is not a guarantee of inclusion in an AI answer.

Handle knowledge bases with care

Search for the brand name and product names in reliable reference sources and note whether listings are accurate. Do not create or edit a Wikipedia article simply to influence AI answers. Wikipedia has its own editorial and notability standards, and a promotional entry can be removed. Where a reputable public knowledge base has a correct, appropriate listing, make sure it reflects verifiable information and follows that site's rules.

Test for name confusion

If your brand name is shared with another company, product or common term, test prompts that make the distinction explicit and prompts that do not. Compare the answers. If engines conflate entities, strengthen the public evidence that distinguishes your brand: consistent naming, clear product descriptions, official profiles and independent coverage that uses the right name.

Pillar 3: Build and test a prompt basket

A prompt basket is a fixed set of questions used to test visibility over time. It gives you more useful evidence than a few one-off searches, because it covers different intents and makes repeated checks comparable.

Cover four useful intent groups

Build prompts from how customers discover, evaluate and validate options. A practical basket can include these groups:

1. Category discovery: Questions about options in a category. Example: “What tools help mid-market teams manage recurring billing?”
2. Problem or use case: Questions describing a pain point. Example: “How can a growing ecommerce team reduce errors in product catalogue updates?”
3. Comparison and selection: Questions that name alternatives or buying criteria. Example: “Compare Brand A and Brand B for a small finance team.”
4. Validation and reputation: Questions about trust, limitations, customer fit or common concerns. Example: “What should a buyer check before choosing a platform for this workflow?”

Include branded and unbranded prompts. Branded prompts help reveal whether the engine knows basic facts about you. Unbranded prompts show whether it brings your brand into the category conversation at all.

A starting basket of 30 to 100 prompts can be useful for a substantial audit, but it is a working range rather than a fixed requirement. Start smaller if you have one product or one market. Increase coverage when the questions can be grouped and interpreted clearly. It is more valuable to test a purposeful set consistently than to collect a large list you cannot review.

Write prompts from real customer language

Use interviews, sales calls, support questions, search queries, review language and internal subject-matter knowledge to build the list. Keep the wording natural. A prompt that reads like a keyword list may not reflect how a buyer asks an assistant for help.

Avoid giving the answer away. A prompt such as “Why is Brand A the best platform?” is biased and cannot measure neutral discovery. Prefer “Which platforms could help with this use case, and what are the trade-offs?” If you want to check a particular capability, ask about the need rather than asserting that your brand has it.

For agencies, record the rules for creating prompts so each client audit is comparable. For example, keep the same number of prompts per intent group and document how competitors were chosen. This helps prevent a client’s apparent improvement from being caused by a changed test set.

Track more than a mention

For each answer, record several dimensions:

• Mention: Did the brand appear at all?
• Position or prominence: Was it in the opening recommendation, later in a list or only in a citation?
• Recommendation: Did the answer actively suggest it for the stated need?
• Accuracy: Were its category, features, audience and current details correct?
• Sentiment and framing: Was the description positive, neutral, qualified or negative?
• Competitors: Which alternatives appeared alongside it?
• Citations: Which URLs or source types supported the answer?

Keep the scoring rules simple and written down. One team member may consider a neutral list entry a recommendation, while another may not. Define the difference before comparing results.

Calculate useful rates

A basic mention rate is the number of tested answers that mention the brand divided by the number of valid answers tested. Calculate it separately by engine and prompt group. You can also calculate recommendation rate using only answers that actively recommend the brand, and share of voice by comparing brand mentions with competitor mentions in the same test set.

State the denominator whenever you report a rate. For instance, say that a brand appeared in a given number of valid answers out of a particular prompt set, on a named engine and test date. A percentage without the scope can look more conclusive than the evidence supports.

Do not overinterpret small differences. Models can produce different answers to similar prompts, and a prompt set may not reflect every buyer. Preserve raw outputs and rerun the same basket before making a claim about a trend.

Keep the test repeatable

Record the exact prompt, engine, date, language, market and any relevant account or search settings. Re-run the basket on a consistent schedule, and retain the earlier results. If you change prompts, create a new version of the basket rather than silently replacing the old questions.

It can help to separate a stable tracking set from exploratory prompts. The stable set is for comparing periods. The exploratory set is for investigating a new product, concern or market opportunity. Label them clearly so exploratory results do not distort the baseline.

Pillar 4: Map citations and third-party coverage

An AI answer can draw on your own site, independent publishers, directories, review platforms, community discussions or other sources. This pillar asks where the information comes from, whether those sources are credible, and what coverage is missing.

Record citation sources, not just citation counts

For every cited URL, record the publisher, page type, subject and the claim it supports. Separate first-party sources, such as your product page, from independent sources, such as a review or editorial comparison. A long list of citations is not automatically useful if those pages are outdated, irrelevant or do not support the answer.

Look for repeated patterns. Are product recommendations grounded in review pages? Are factual descriptions taken from your own documentation? Does an answer cite a dated comparison instead of the current product page? Does the source mention a product name that you no longer use?

Compare citation behaviour across engines

Do not assume the same prompt will draw from the same source set everywhere. An analysis by Qwairy examined 118,101 answers and 669,065 citations. In that analysis, Perplexity averaged 21.87 citations per answer, compared with 7.92 for ChatGPT. Those figures describe that study's dataset, not a permanent rule for every query or current product experience.

A separate analysis by Profound examined 680 million citations and reported different source patterns across platforms. Its reported findings include a substantial role for Reddit in Perplexity's top-source share and for Wikipedia in ChatGPT's top-source share. Use the study as a reason to inspect source types, not as a universal instruction to chase one platform or publisher.

Research published by AirOps reported that 84% of AI citations in its analysis came from domains outside the web's top 100 cited domains. The practical lesson is not that authority no longer matters. It is that a brand should inspect the specific sources appearing for its category instead of assuming only the biggest websites can contribute.

Find citation gaps

For each important prompt, compare the sources that support competitor recommendations with the sources that mention your brand. Make a simple gap list:

Treat the gap as a research lead, not an instruction to seek every mention. Assess whether the source is credible, relevant to the audience and editorially independent. Do not pay for undisclosed endorsements or attempt to manipulate community discussions.

Review community and directory sources responsibly

Check relevant review platforms, professional directories and communities where your actual audience evaluates products. Review the rules of each site before participating. Ask customers for honest feedback through normal, transparent processes. Respond to legitimate criticism with useful information, not scripted comments designed to insert brand keywords.

For SaaS companies, review sites and category comparisons may matter to buyers. For ecommerce, relevant publishers, product reviews and community discussions may be useful. The right source set depends on the product and audience. Use citation evidence to guide the investigation rather than assuming a particular platform matters equally for every brand.

Pillar 5: Find content gaps and improve extractability

The aim is not to write for a machine at the expense of people. It is to make useful, accurate information easy to find, understand and quote. An AI visibility audit should connect each missing or inaccurate answer to a page that can responsibly address it.

Identify questions your site does not answer

Group prompts where the brand is missing, inaccurately described or supported by weak sources. Then check whether your site answers the underlying question. Look for gaps such as:

• The page names a feature but does not explain what it does.
• The content explains a product but not who it suits.
• A comparison page lists benefits without stating relevant limitations.
• Pricing information is scattered or out of date.
• Documentation describes setup but omits requirements or prerequisites.
• A high-intent question is answered only in an old blog post.

Map each gap to the most suitable page. If the information belongs on an existing product page, improve that page rather than publishing another near-duplicate article. If it needs a distinct guide, plan a page that answers a real customer question and adds clear value.

Put the answer where readers expect it

Use descriptive headings and answer the heading's question near the start of the section. A concise first answer helps readers decide whether to keep reading; the rest of the section can explain the reasoning, steps, examples and exceptions.

For example, a section titled “Does the platform support Shopify?” should answer yes or no in its opening, if that is accurate, then explain the scope, setup and relevant limitations. Avoid opening with a broad paragraph about innovation before addressing the actual question.

Use clear page structure:

• One main topic per page.
• Descriptive headings that reflect the content below them.
• Short paragraphs for distinct ideas.
• Lists for steps, requirements or options.
• Tables when readers need to compare attributes.
• Definitions for technical terms that may not be familiar.

These choices improve usability and make the content easier to scan. They do not guarantee that a passage will be cited.

Add evidence and useful detail

Replace general claims with information a reader can verify. That might include a clear process, documented product requirements, examples, original research or a transparent comparison of options. Use named sources for external claims and show when information was last reviewed if freshness matters.

Avoid unsupported superlatives such as “best”, “leading” or “most advanced”. If you compare products, explain the criteria and date of the comparison. If you state a product feature, make sure the product and support teams agree that it is current.

A strong informational page often answers several connected questions:

1. What is the thing or problem?
2. Who is it for?
3. How does the solution work?
4. What are the main steps or requirements?
5. What are the trade-offs, limits or alternatives?
6. Where can a reader verify the details?

The right answers depend on the page. Do not force every page into the same template if it makes the content repetitive.

Improve freshness without cosmetic updates

Review high-value pages when a product, price, policy or process changes. Update the actual content, examples and supporting links. If you show a revised date, it should reflect a meaningful review or change, not a date reset that gives readers a false impression.

Keep retired pages useful where appropriate. Redirect pages that have been replaced, and correct internal links that still point to obsolete information. In product documentation, identify which version or plan a feature applies to. These steps reduce conflicting signals for people and answer systems alike.

Use internal links to give context

Link related pages with descriptive anchor text. A product page can link to the setup guide; a guide can link to the relevant comparison or pricing details. Make sure the linked page expands on the subject rather than simply repeating it.

For more guidance on the connection between content and AI search, see SearchFIT's article on content strategies for AI search visibility. Ecommerce teams can also review the practical considerations in this guide to product schema for AI answers.

Connect AI visibility to traffic and business outcomes

Visibility is an important observation, but it is not the same as a sale, lead or qualified visit. Build a view that connects answer presence with referral traffic and business outcomes without claiming that every conversion came from an AI answer.

Review referral traffic cautiously

In your analytics platform, inspect referral traffic from identifiable AI services and compare it with your wider acquisition reporting. The exact source names and reporting behaviour can change. Some users may arrive through an app, a copied link or a browser flow that does not retain an obvious referrer. As a result, referral sessions are only one part of the picture.

Check landing pages, engagement and conversions for those visits where the data is available. Do not treat the absence of a referral session as evidence that nobody saw or acted on an AI answer. Also do not label every unusual direct visit as AI search traffic.

Build a balanced measurement view

Pair answer-level measures with website and business measures:

Keep the definitions stable. When the prompt basket or scoring method changes, note it in the report. Share examples of raw answers alongside summary measures so stakeholders can see both the trend and its limits.

Prioritise fixes and repeat the audit

An audit is useful when it leads to work that can be owned, completed and checked. Create a register of findings rather than ending with a collection of screenshots.

Score findings by impact and effort

For each finding, record:

• The evidence: prompt, engine, date, answer and source.
• The affected page, product or audience.
• The likely impact if it remains unresolved.
• The proposed action and the person responsible.
• Any dependency, such as legal, product or engineering review.
• How you will verify the fix.

Then sort work into practical groups:

• Fix now: A crawler is blocked from a public priority page, a key fact is wrong, or an important page is broken.
• Plan next: A useful answer or comparison is missing and needs subject-matter input.
• Monitor: Results are inconsistent, the evidence is weak or the best response is not yet clear.
• Do not pursue: The source is unsuitable, the prompt is irrelevant or the proposed change would be misleading.

Do not prioritise solely by how dramatic a model answer looks. A single odd response may not represent a repeatable issue. Check whether the same problem appears across related prompts, engines or sources before making a substantial change.

Assign owners by type of work

Crawl and rendering issues usually need a developer, platform owner or hosting provider. Entity and product facts need input from brand, product and communications teams. Citation gaps may involve customer marketing, PR, partnerships or review-site owners. Content work needs an editor and a subject-matter reviewer.

Agencies can make this easier by presenting each finding in a consistent format: evidence, business relevance, recommended action, owner and measurement. SearchFIT's agency comparison page can help teams considering how to organise ongoing visibility work alongside other marketing activity.

Set a useful review rhythm

Choose a repeat schedule that matches the rate of change in your market and content. Re-run the fixed prompt set after meaningful site changes and at planned intervals. Review crawler access when publishing infrastructure or security policies change. Recheck factual accuracy when products, plans or positioning change.

Keep versions of the prompt basket and the audit report. A simple change log can explain why a result moved: a page was updated, a source changed, a prompt was revised or the engine experience changed. Without that context, teams can mistake changes in the test for changes in visibility.

Common audit mistakes

Treating one model answer as a verdict

A single answer is a useful example, not a complete measurement. Repeat the prompt, inspect related prompts and compare engines. Preserve the answer so you can distinguish a recurring pattern from an isolated response.

Reporting one blended visibility score

A combined score can conceal important differences. A brand might be visible in one engine and absent in another, or appear in answers but be inaccurately described. Report engine-level findings and explain any calculation used to combine them.

Focusing only on your website

First-party content matters, but cited answers can draw on independent sources too. Audit your own pages and the external sources that appear around your category. Make sure any external coverage work is accurate, transparent and appropriate to the publisher's rules.

Confusing technical permission with actual retrieval

A robots.txt allowance does not guarantee that a page loads successfully. Check server responses, CDN behaviour, page rendering and actual logs. Keep private areas protected, and confirm crawler identity before changing security settings.

Publishing thin content to fill every gap

A prompt without a matching page does not automatically justify a new article. First decide whether the topic matters to the audience, whether you can provide a useful answer and where it belongs in the site structure. One strong, maintained resource is often more useful than several pages that repeat the same points.

Overlooking accuracy and trust

Visibility is not a win if the answer is wrong. Correct outdated prices, product names, capabilities and company details before trying to increase mentions. Do not invent claims, create artificial reviews or push promotional edits into independent sources.

Next steps

Start with a defined market, one product or service, a small competitor set and a fixed prompt basket. Save the answers and citations, then work through crawl access, entity clarity, prompt performance, source coverage and content gaps. Turn each finding into an action with an owner and a way to verify the change.

If you want to see how SearchFIT supports ongoing AI visibility tracking and content work, review SearchFIT pricing. Use the audit as a baseline, not a one-off score: repeat the checks, preserve the evidence and focus on making your brand easier to find, understand and describe accurately.