Run an AI search visibility audit: build a prompt set, sample assistant answers, check crawler access and rendering, and turn findings into a prioritized fix list.
An AI search visibility audit establishes whether assistants can reach your content, whether they cite it, and whether they describe you accurately. Without that baseline, any optimization work is guesswork and any reported improvement is unverifiable.
This is a method you can run yourself. It takes a few days of focused work and produces a prioritized fix list.
Step one: build the prompt set
Write the questions buyers actually ask an assistant, in their words. These are longer and more conversational than keywords: "which company should I hire to build an internal chatbot for a mid-size insurer" rather than "AI chatbot development".
Cover four intents: category questions that define the problem, comparison questions between approaches or vendors, selection questions about how to choose, and brand questions about you specifically. Write three or four phrasings of each, because assistants respond differently to small wording changes.
Aim for thirty to sixty prompts. A smaller set measured repeatedly is far more useful than a large set measured once.
Step two: sample the answers
Run every prompt against each assistant your buyers use. Record the full answer, the sources cited with URLs, whether you appear, and how you are described.
Repeat the run at least three times across different days, ideally in a clean session without personalization or prior history. Outputs vary between runs, so a single sample tells you very little. Treat the proportion of runs where you appear as your citation rate, not a yes or no.
Note who is cited when you are not. Competitor citations tell you which sources the model trusts in your category, which is directly actionable.
Step three: check that crawlers can reach you
Visibility failures are frequently technical rather than editorial. Verify each of the following.
Your robots rules allow the AI crawlers you want, checked by user agent rather than assumed.
Your CDN, WAF, or bot protection is not silently blocking those agents. This is the single most common cause of total invisibility.
Key pages render their content in the initial HTML response. Content that only appears after client-side JavaScript execution is frequently missed.
Response times are reasonable under crawl, and you are not rate limiting legitimate agents into failure.
Canonical tags, redirects, and sitemaps are consistent, so the crawler resolves one authoritative URL per topic.
Fetch your own pages using the crawler user agents and read exactly what comes back. Assumptions fail here more often than anywhere else in the audit.
Step four: assess extractability
Read your top pages the way a model does, looking for self-contained claims. For each important question, is there a passage that answers it in one or two sentences without depending on the paragraph before it?
Common problems: the answer is spread across a long narrative, the key number is only in an image or chart, the claim is hedged into meaninglessness, or the page buries its conclusion under six paragraphs of preamble.
Check structured data too. Organization, service, and question-answer markup help the engine resolve what you are and what you assert.
Step five: assess corroboration
For your three or four most important claims, search for independent confirmation. If a claim about your service, pricing model, or capability appears nowhere except your own domain, models have little basis to repeat it.
Inventory where you appear off-site: industry publications, directories, review platforms, documentation, community discussions. Note where competitors appear and you do not. This gap list becomes the authority workstream.
Step six: check accuracy, not just presence
Read how you are described in the answers you did appear in. Record any wrong service descriptions, outdated pricing, incorrect location, or confusion with a similarly named company.
Misdescription usually traces to a stale or contradictory source, sometimes your own old pages, sometimes a third-party profile you forgot existed. These are often the fastest wins in the entire audit.
Turning findings into a plan
Order the work by cost and impact. Crawler access and rendering fixes come first because they are cheap and gate everything else. Correcting inaccurate descriptions comes next. Restructuring existing high-value pages for extractability follows. Authority and corroboration work runs continuously in the background because it is the slowest to compound.
Re-run the same prompt set monthly using the same method, and compare citation rate over time rather than judging by any single check.
Frequently asked questions
How often should we repeat the audit?
Sample the prompt set monthly. Repeat the full technical and corroboration review quarterly, or after any significant site change.
Do we need paid tools?
Not to start. The prompt set can be run manually and recorded in a spreadsheet. Tools help with scale and consistency once the method is established.
Why do results differ between runs?
Generation is non-deterministic and models are updated without notice. This is why the method depends on repeated sampling rather than single observations.
What if we are invisible everywhere?
Check crawler access first. Complete invisibility is far more often a blocked bot or an unrendered page than a content quality problem.
Explore our software and web capabilities or contact Voquarn Code for an audit of your AI search presence.
Written by
Moueen Togarvi
Founder & CEO at Voquarn Code, focused on product engineering, search growth, and practical AI systems.