--- title: "How to run an AI visibility audit: a step-by-step process" slug: "how-to-run-an-ai-visibility-audit" category: "ai-visibility" canonical_path: "/articles/ai-visibility/how-to-run-an-ai-visibility-audit" meta_title: "How to Run an AI Visibility Audit — Prime AI Visibility" meta_description: "A step-by-step AI visibility audit: define buyer prompts, run them across engines, record citations and sentiment, benchmark competitors, and set a cadence." author: "The Prime AI Visibility editorial team" date: "2026-07-31" last_updated: "2026-07-31" read_time: "11 min" keywords: - AI visibility audit - buyer prompts - baseline - citation gap - competitor benchmark featured_image: "/brand/articles/ai-visibility/how-to-run-an-ai-visibility-audit.png" featured_image_alt: "Concentric rings of translucent amber spheres with the brightest ones connected by thin lines to a central hollow ring" og_image: "/brand/articles/ai-visibility/how-to-run-an-ai-visibility-audit.og.png" cta_mid_headline: "Turn your audit into a repeatable baseline" cta_mid_body: "Prime AI Visibility runs your buyer prompts across ChatGPT, Perplexity, Gemini, Claude, Copilot, and Google AI Overviews on a schedule, and records which sources each engine names." cta_mid_button: "See what the engines cite" cta_bottom_headline: "Ready to run the audit end to end?" cta_bottom_body: "Bring the ten buyer prompts you defined here, and let a workspace record citations, mentions, and sentiment across every engine so your re-audit is one click, not a spreadsheet rebuild." cta_bottom_button: "Start your first audit" --- # How to run an AI visibility audit: a step-by-step process An AI visibility audit is a structured pass that measures how AI answer engines represent your brand for the questions your buyers actually ask. You define a set of buyer prompts, run them across each engine, record whether you are cited or mentioned and in what tone, benchmark that against competitors, isolate citation gaps, and set a cadence to re-measure. The output is a baseline, not a guaranteed outcome. ## AI visibility audit: the short answer 1. **Start from buyer prompts, not keywords.** An AI visibility audit measures conversational questions your buyers ask engines, so the prompt list is the foundation everything else rests on. 2. **Record what you can observe, engine by engine.** Capture citations, mentions, and sentiment per engine at a fixed moment to establish a baseline you can compare against later. 3. **Turn gaps into a prioritized, repeatable plan.** Compare your results to a competitor benchmark, isolate the citation gaps, rank fixes by effort and coverage, and schedule the next audit. ## Why an audit, and what it can and cannot tell you Classic SEO audits check crawlability, indexation, and ranking positions on a results page you can see. An AI visibility audit asks a different question: when someone poses a real buying question to ChatGPT, Perplexity, Gemini, Claude, Copilot, or Google AI Overviews, does your brand appear in the answer, and how is it described? The generated answer is the surface, not a static list of ten blue links. Set expectations before you begin. An audit is a measurement exercise, so it tells you what the engines are doing right now for a defined prompt set. It does not — and cannot — promise that a given fix will earn a citation or move you into an answer. AI answer engines do not document their ranking or selection internals, and their outputs vary between sessions, models, and dates. Treat every reading as a bounded snapshot: accurate for the prompt, engine, and moment you captured it, and nothing more. That honesty is what makes the baseline useful, because you are comparing like-for-like snapshots over time rather than chasing a number the engine never published. If you are still deciding whether to do this by hand or with a platform, the companion piece on [choosing an AI visibility platform that fits your stack](https://primeaivisibility.com/articles/ai-visibility/how-to-choose-an-ai-visibility-tool) walks through the trade-offs before you commit to a workflow. ## Step 1 — Define your buyer prompts The prompt list is the single most important input, because a weak prompt set produces a misleading audit no matter how carefully you run it. Buyer prompts are the natural-language questions a real prospect types into an assistant while researching a purchase — not the head terms you would target in a search engine. "best AI visibility tool for a small marketing team" is a buyer prompt; "AI visibility tool" is a keyword. Build the list from evidence, not intuition: - **Mine your own funnel.** Pull the questions sales hears on calls, the phrasing in support tickets, and the searches in your site's internal search logs. - **Cover the journey.** Include problem-aware prompts ("how do I tell if AI engines mention my brand"), solution-aware prompts ("tools that track AI search citations"), and vendor-comparison prompts ("Prime AI Visibility vs a spreadsheet"). - **Write them the way people talk.** Full questions, follow-ups, and comparisons — the conversational shape assistants are built to answer. - **Keep the set fixed.** Ten to thirty prompts is a workable first audit. Once chosen, freeze the list so future audits compare against the same baseline. Group prompts into themes (category education, comparison, use case, objection). Themes let you spot whether you are absent across an entire stage of the journey rather than missing a single question. ## Step 2 — Choose your engines and record conditions Decide which engines are in scope. Most teams start with the assistants their buyers actually use, then expand. For each engine, note the conditions that affect the answer so your baseline is reproducible: the model or mode, whether browsing or a live index is active, the account or region if that changes results, and the exact date. Answers drift, so a snapshot without its conditions is not a baseline — it is an anecdote. A grounding step here saves confusion later: run a bounded, engine-specific access check rather than assuming one universal rule. Where an engine documents a first-party crawler, verify that the page is indexable and that the crawler is allowed to fetch and render it — reviewing [how the AI crawlers fetch and render your site](https://primeaivisibility.com/articles/geo/ai-crawlers-explained) shows what to confirm. But absence from your own crawler logs does not prove a page cannot be cited: some engines retrieve through third-party search indexes (for example, ChatGPT search and Copilot have drawn on Bing's index) and some answers come from pre-existing model training data. The engines do not document a single universal "only what my crawler fetched" rule, so treat crawler access as one access path among several, and diagnose an absent gap per engine rather than declaring it purely technical. ## Step 3 — Run the prompts and capture raw answers Run each prompt in each in-scope engine and save the full answer verbatim. Do not summarize as you go; capture the raw text, any linked sources, and a timestamp. Consistency matters more than volume — use the same phrasing, the same session hygiene (fresh context per prompt where possible), and the same day for the whole pass so nothing in the baseline is confounded by timing. Because engine outputs vary between runs, decide upfront how you handle variance. A pragmatic approach is a single clean pass for the baseline, with a note that any single reading is a snapshot. If you need more confidence on a critical prompt, run it a small fixed number of times and record each result rather than averaging into an invented figure. ## Step 4 — What to record for each result This is the heart of the audit. For every prompt-and-engine pair, record a consistent set of fields. Standardizing the schema now is what makes the competitor benchmark and the re-audit possible later. | Field | What to capture | Why it matters | |---|---|---| | Prompt | The exact question text and its theme | Anchors the reading to a fixed buyer question | | Engine and conditions | Engine, model/mode, browsing on/off, date | Makes the snapshot reproducible | | Cited | Whether a page of yours appears as a linked source (strong / partial / none) | Distinguishes a real citation from a passing mention | | Mentioned | Whether the brand is named in prose without a link | Captures presence that a citation-only view would miss | | Sentiment | Tone of the mention (positive / neutral / negative) | Presence is not the same as favorable framing | | Source URL | The specific page the engine cited, if any | Tells you which content is doing the work | | Competitors named | Which rival brands appear in the same answer | Feeds the competitor benchmark directly | | Notes | Follow-up prompts, hallucinations, stale facts | Surfaces correctable errors and content gaps | Use word ratings, not invented numbers, for judgment fields like "cited" and "sentiment." "Strong / partial / none" is honest about what you observed; a precise percentage on a single run would imply a rigor the reading does not have. The distinction between *cited* (a linked source) and *mentioned* (named in prose) is the one teams most often collapse — keep them separate, because they call for different fixes. ## Step 5 — Establish your baseline The baseline is the deliverable that gives your AI visibility audit lasting value. With the grid filled in, summarize it into a baseline: for each engine and each theme, how often you were cited, how often merely mentioned, and the prevailing sentiment. Do not compute a single composite "visibility score" that hides the variation — the value of the baseline is that it is specific. A useful baseline reads like "cited on comparison prompts in Perplexity, absent from category-education prompts in Gemini, mentioned neutrally in Google AI Overviews." This baseline is your reference point, and its only job is comparison over time. Resist the urge to interpret a first-audit baseline as good or bad in absolute terms — you have nothing to compare it against yet. Its worth compounds at the second audit, when a like-for-like snapshot shows movement you can attribute to specific changes. ## Step 6 — Build the competitor benchmark An audit of only your own presence tells you where you stand but not what "good" looks like for the category. The competitor benchmark fixes that. From the "competitors named" field, tally which rival brands appear across your prompt set and on which themes. That tally reveals two things: which competitors the engines treat as reference points for your category, and which specific pages they cite for those competitors. Keep the benchmark descriptive and verifiable. Record which competitor was named on which prompt and what page was cited — all observable facts. Do not speculate about *why* an engine favored a competitor; the engines do not document their selection logic, so any causal claim is inference. The benchmark's job is to show you where a rival is present and you are not, which is exactly the input the next step needs. ## Step 7 — Isolate the citation gaps A citation gap is a buyer prompt where the answer is commercially relevant to you but you are absent, weakly mentioned, or framed unfavorably — especially where a competitor is cited and you are not. Working from the grid and the benchmark, tag each prompt into one of a few buckets: - **Absent gap.** You are neither cited nor mentioned, and the prompt matters to your funnel. - **Mention-only gap.** You are named in prose but not cited as a source, so the engine knows you exist but is not pointing to your content. - **Sentiment gap.** You are present but described with stale, incomplete, or negative framing. - **Competitor-exposed gap.** A rival is cited on a prompt where you have equivalent or better content that the engine is not surfacing. For each gap, note the likely lever. Absent gaps often point to missing or unindexed content; mention-only gaps often point to content that exists but is not being cited as a source; sentiment gaps often point to outdated pages. A frequent absent-gap cause is that engines simply have not been pointed at the relevant pages — reviewing [what an llms.txt file does and does not control](https://primeaivisibility.com/articles/geo/what-is-llms-txt) helps you separate a genuine content gap from a discoverability one before you commission new writing. ## Step 8 — Prioritize the fixes You now have a list of gaps and a rough lever for each. Prioritize rather than attacking everything at once. A simple two-axis sort works well: potential coverage (how many buyer prompts or themes a fix could plausibly touch) against effort (how much work the fix requires). Fixes that plausibly touch several prompts for modest effort go first. | Priority | Pattern | Typical first move | |---|---|---| | High | Absent or competitor-exposed gap on a high-intent theme, existing adjacent content | Strengthen and clarify the closest existing page so it directly answers the prompt | | Medium | Mention-only gap across several prompts | Add explicit, quotable answers and ensure the page is reachable and rendered | | Medium | Sentiment gap from stale facts | Update the page, correct the outdated claim, and log it for re-audit | | Lower | Absent gap needing net-new content | Commission new content mapped to the specific prompt cluster | Frame every fix as an input, never a promise. "Clarify this page so it directly answers the prompt" is an honest action; "this will earn a citation" is not, because the engines control selection and do not document it. Log each planned fix against the prompt it targets so the next audit can check whether anything observably changed. ## Step 9 — Set a re-audit cadence An AI visibility audit is not a one-time deliverable, because engine outputs, models, and your own content all change. Treating the AI visibility audit as a recurring measurement, rather than a single report, is what lets you see movement. Set a cadence — monthly is a common starting rhythm for active programs, quarterly for slower-moving categories — and re-run the exact same frozen prompt set under the same recorded conditions. Comparing snapshot to snapshot is the entire point of holding the prompt list fixed. Between full audits, watch for events that warrant an off-cycle check: a major model release, a significant content launch, or a competitor's visible push. When you re-audit, keep the schema identical so the new grid stacks cleanly on the baseline. As your program matures, a glossary of the recurring terms — citation, mention, sentiment, baseline — keeps the team aligned; the [Prime AI Visibility reference glossary for AI-search terms](https://primeaivisibility.com/glossary) is a handy shared vocabulary. And if you want the wider context for why any of this matters as a discipline, the [Prime AI Visibility site overview of the measurement pipeline](https://primeaivisibility.com/) frames where an audit sits in a full program. ## Common misconceptions **"An audit proves what the engines will do."** It does not. An audit records what engines *did* for a specific prompt set at a specific time. Because selection internals are undocumented and outputs vary, treat results as observations, not predictions. **"One run per prompt is enough forever."** One clean pass is fine for a baseline, but a single reading is a snapshot. Variance is real, so note it, and re-audit on a cadence rather than trusting a lone result indefinitely. **"A citation and a mention are the same win."** They are not. A mention means the engine names you; a citation means it points to your page as a source. They call for different fixes, so record them as separate fields. **"More prompts is always better."** A focused set of genuine buyer prompts beats a sprawling list of keywords. Depth on the questions that drive purchases is worth more than breadth across queries no buyer asks. > **Turn your audit into a repeatable baseline** > > Prime AI Visibility runs your buyer prompts across ChatGPT, Perplexity, Gemini, Claude, Copilot, and Google AI Overviews on a schedule, and records which sources each engine names. > > **[See what the engines cite](https://app.primeaivisibility.com/sign-up)** ## References 1. Google Search Central, *AI features and your website*. 2. Google Search Central Blog, *Top ways to ensure your content performs well in Google's AI experiences on Search* (21 May 2025). 3. OpenAI, *ChatGPT search* documentation. 4. Perplexity, *What is Perplexity?* Help Center. ## Next steps 1. **[Weigh a DIY audit against a platform-run one](https://primeaivisibility.com/articles/ai-visibility/how-to-choose-an-ai-visibility-tool)** before you decide how to run the next cycle. 2. **[Confirm which engines can actually reach your pages](https://primeaivisibility.com/articles/geo/ai-crawlers-explained)** so an absent gap is diagnosed correctly rather than mistaken for a content gap. 3. When you are ready, **[create a Prime AI Visibility workspace](https://app.primeaivisibility.com/sign-up)** and bring 10 buyer prompts. ## Frequently asked questions **How many buyer prompts should an AI visibility audit start with?** Ten to thirty focused buyer prompts is a workable first audit for most teams. Choose full, conversational questions your buyers actually ask across the journey, group them into themes, and then freeze the list so every future audit compares against the same baseline. **What is the difference between a citation and a mention in an audit?** A citation means the engine points to one of your pages as a linked source in its answer, while a mention means your brand is named in the prose without a link. Record them as separate fields, because a mention-only result and an absent result usually call for different fixes. **How do I find a citation gap?** Tag each prompt in your grid by whether you were cited, only mentioned, framed unfavorably, or absent — then cross-reference the competitor benchmark. A citation gap is a commercially relevant prompt where you are absent or weak, especially one where a rival is cited and you are not. **Can an AI visibility audit tell me why an engine cited a competitor?** No. AI answer engines do not document their selection or ranking internals, so any claim about why a specific answer named a competitor is inference. An audit can reliably record that the competitor was named and which page was cited, but not the causal reason behind it. **How often should I re-run the audit?** Monthly is a common cadence for active programs and quarterly for slower categories. Re-run the exact same frozen prompt set under the same recorded conditions, and also run an off-cycle check after a major model release or a significant content launch. **Do I need a tool, or can I audit with a spreadsheet?** Both work. A spreadsheet is fine for a small, infrequent audit if you are disciplined about recording conditions; a platform helps when you re-audit on a cadence across many prompts and engines. The right choice depends on prompt volume, engine count, and how often you plan to re-measure. > **Ready to run the audit end to end?** > > Bring the ten buyer prompts you defined here, and let a workspace record citations, mentions, and sentiment across every engine so your re-audit is one click, not a spreadsheet rebuild. > > **[Start your first audit](https://app.primeaivisibility.com/sign-up)**