How to run an AI visibility audit
Run an AI visibility audit in a day: choose buyer questions, save answers, check mentions and sources, then prioritize page fixes and a re-check.
What should an AI visibility audit include?
Choose real buyer questions, ask them across the assistants your customers use, and save the full answers. Record brand mentions, position, rivals and source links separately. Check the relevant pages for access problems and incorrect facts, prioritize fixes using demand, and schedule the same questions for a re-check.
Set a scope you can finish today
An AI visibility audit should end with saved evidence and a short work list. Start with one brand, one market and one buying decision. Put the website, language, country and audit date at the top of a spreadsheet. Decide who will review the findings and who can change the pages.
For a first pass, budget an hour to choose questions, two hours to collect answers, two hours to inspect pages and an hour to rank fixes. Use the remaining time for repeated questions and a written handoff. This is a suggested work plan, not a statistical sampling standard.
The same process works if your brief calls it a GEO audit or an LLM visibility audit. Define what you will observe before collecting anything: which products, which interfaces, which questions and which dates. Keep the initial scope small enough that you can read every answer yourself.
Choose questions from buyer decisions
Write ten questions using sales conversations, support tickets and search queries. Include category discovery, comparison, requirements and objections. For a booking product, examples might be “What booking software suits a two-location salon?” and “Which appointment tools support deposits and recurring bookings?” Make the examples fit your actual customers.
Separate questions that name your brand from questions that do not. “Is our product suitable?” tests how an assistant describes you after being prompted. “Which products are suitable?” tests whether you appear without that help. Keep both groups, but report them separately.
Use ordinary Search Console query and page data to check whether related needs already bring attention to your site. Google documents how to sort queries by clicks or impressions and inspect the pages shown for a query. Those are Google Search observations, not counts of assistant prompts. Google's performance-report guide.
Write the reason for each question beside it. A sales objection, recurring support request or relevant search query gives you a reason to keep it. Avoid filling the sheet with slight rewrites just because they produce favorable answers. Save a separate backlog for questions you cannot cover today.
Record the collection conditions
Choose the assistants your customers actually discuss, and use their browser interfaces for this manual audit. Name the product and mode precisely. A search-enabled conversation and a conversation without search belong in different groups. Record the displayed model when available; otherwise write “not shown.”
For each run, record language, location, sign-in state, relevant personalization settings and time. OpenAI documents that ChatGPT search can use approximate location and, when enabled, saved memories to shape searches. A new conversation alone does not establish identical conditions for everyone. ChatGPT search documentation.
Start each question in a fresh conversation. Paste the exact wording, let the answer finish and save the full text plus a screenshot. Keep follow-up questions in a separate group: asking “What about our brand?” changes the task. Give every saved answer an identifier so another reviewer can find it.
Log mentions, position and sources separately
For this audit, define a mention as an unambiguous reference to your business in the answer. Check the surrounding sentence when your name is also an ordinary word or another company's name. Record approved abbreviations and product names in a small matching note.
Define position as the order of distinct relevant brands in the saved answer. Count each brand once. Record “absent” when yours is missing and “unclear” when the format has no meaningful order. This is your audit convention; show the numbered list or excerpt beside the value so readers understand it.
Use a separate field for source URLs. Preserve the link attached to the answer and note whether it points to your domain, a rival, a directory or an editorial page. Distinguish a brand mention, a returned source link and an explicit citation attached to a claim. OpenAI's API documentation, for example, separates consulted sources from inline citations. OpenAI web-search documentation.
Copy this results-table template into your notes. Use one row per question, assistant and run; put collection settings and full evidence in the linked record. The row below is a placeholder, not a measured result.
| Question and run | Assistant and mode | Mention and position | Rivals | Source URLs | Evidence and next action |
|---|---|---|---|---|---|
| Q01, run 1, date/time | Product; mode; model if shown | Yes / no / unclear; order | Names in answer | Exact links; citation status | Saved answer ID; page to inspect |
Repeat the questions that drive decisions
Run each selected question again in a fresh conversation under the same recorded conditions. If time is tight, prioritize the questions behind your proposed fixes. Save both answers, including a disappointing result. Anthropic documents that identical API inputs can produce different outputs even at temperature zero; repeated observations deserve their own records. Anthropic's glossary.
Show counts before percentages. If your brand appears in four of ten completed answers, write both numbers. Keep the denominator tied to the same question group and assistant. Record timeouts, refusals and incomplete answers separately from completed answers without your brand. Plan a retry for collection failures instead of silently converting them into absences.
Check page access and factual accuracy
Open your highest-priority landing pages while signed out. Can you read the main explanation without a login or an obstructive overlay? Follow the links the answers returned. Record missing pages, unexpected redirects and outdated destinations for the website owner to inspect.
For Google, use URL Inspection to compare the indexed page with a live test. The indexed report describes Google's stored version; a successful live test does not confirm that the current page is indexed. Google's URL Inspection documentation.
For each material claim in an answer, find the page that should establish the fact. Check pricing conditions, supported countries, product capabilities and availability against your approved records. Classify the issue as “our page is wrong,” “our page is unclear,” “another source is wrong” or “source unknown.” Keep the exact disputed sentence.
Turn vague findings into specific edits. Replace “improve category relevance” with “state whether deposits are supported on the booking page and link to the setup instructions.” If an external directory contains an obsolete fact, record its correction process as a separate task. Do not assume you can edit it today.
Rank fixes using demand and consequence
Give each proposed fix three short notes: evidence of buyer demand, the decision affected and the effort required. Put access failures and materially incorrect facts first. Then favor useful changes on pages connected to repeated buyer questions. Leave cosmetic rewrites until those items are resolved.
For example, a missing explanation of deposit refunds may matter more than adding another broad category article. Your evidence might be three sales questions, relevant Google queries and an answer that describes the policy incorrectly. Keep those observations together instead of inventing one opaque priority score.
Choose no more than three fixes for the first handoff. Name the page, desired change, source of the correct facts and responsible role. Write what you expect to observe next: a corrected explanation on the page, successful access, or different wording in a later answer.
Copy the audit checklist
- Record the brand, website, market, language and date.
- Choose ten buyer questions and label branded questions separately.
- Select assistants and record their modes and collection settings.
- Save each complete answer, screenshot and collection time.
- Log mentions, position, rivals and source URLs separately.
- Repeat important questions and retain unsuccessful collection attempts.
- Inspect priority pages for access problems and incorrect facts.
- Rank three fixes using demand, consequence and effort.
- Assign each fix and record its publication date.
- Set a re-check date using the original questions and conditions.
Set a re-check date and preserve the baseline
Choose a date two weeks after the planned edits as an administrative checkpoint, not a promised response time. Keep the original questions unchanged for that comparison. Add new buyer questions to a separate cohort. Save the revised page text and publication date beside the original evidence.
Treat the audit as a sample of defined conditions. API and browser observations are different samples, and answers can vary. A before-and-after difference does not prove an edit caused it. Report the saved evidence, completed checks and unresolved gaps together so the next review starts from the same facts.
SearchSeal currently monitors AI Overview, ChatGPT, Gemini and Perplexity, saving answers to tracked questions with brand mentions, position, rivals and returned source links. Optional, read-only Google Search Console and Bing Webmaster connections add page demand to help rank fixes. To prepare your comparison set, find who you're compared with using competitor suggestions from your website.
Frequently asked questions
Can I run an AI visibility audit without paid software?
Yes. Use the assistant interfaces available to you, a spreadsheet and saved answers. Keep the question wording, collection conditions and evidence consistent enough for another reviewer to repeat your process.
How many questions should my first audit include?
Ten is a practical starting scope for this one-day workflow. Choose questions that represent distinct buyer decisions, then repeat the important ones. It is a work-planning suggestion, not a statistically representative sample size.
What is the difference between a GEO audit and an AI visibility audit?
Use the brief to define the scope. This workflow covers observed answers, page access, factual accuracy and prioritized fixes. Either label can describe that work, so specify the assistants, questions and evidence rather than relying on the label.
What should I deliver at the end of the day?
Deliver the question list, saved answers, results table, three prioritized fixes with responsible roles, and a scheduled re-check. Keep the original evidence available when someone reviews a proposed edit.