Key takeaways
- An AI visibility audit tests a fixed set of buyer questions across ChatGPT, Gemini, Perplexity, Claude and Google’s AI features, then scores how often your brand is named, recommended and cited.
- Score unbranded and branded prompts separately, and tag each prompt with the page that should be cited. Mixed scores flatter you.
- Being named inside the answer still pays when clicks fall. Seer Interactive found that pages cited in Google’s AI Overviews earn more than double the clicks of uncited pages on the same results page.
- Since October 8, 2026, many ChatGPT clicks land in analytics as Direct traffic. Check before you report a visibility loss.
Your buyers now ask AI assistants which vendor to hire, and none of those conversations show up in Search Console. An AI visibility audit is how you find out what those assistants say. You run a fixed set of buyer questions across the major AI engines, record every answer under the same conditions, and score how often you are mentioned, recommended and cited against named competitors. Then you trace each gap to a cause you can fix, such as a missing page, a blocked crawler or a wrong fact on a review site. This guide covers the full 7-step process, the metrics, the tools and a 90-day plan.
What is an AI visibility audit?
An AI visibility audit is a repeatable measurement of how AI answer engines describe, recommend and cite your brand for the questions your buyers ask. It covers presence, accuracy, framing and sources, and it compares you with competitors you name in advance. An SEO audit, by contrast, measures rankings and site health on a results page.
| Dimension | SEO audit | AI visibility audit |
| Unit of measurement | Keyword and URL | Prompt and full answer |
| Core question | Where do our pages rank? | Are we named, recommended and cited? |
| Main metrics | Rankings, organic clicks, crawl errors, backlinks | Mention rate, recommendation rate, citation rate, share of voice, accuracy |
| What shapes results | Mostly your site and the links pointing to it | Your site plus reviews, directories, forums, media and roundups |
| Stability | Fairly steady day to day | Answers vary between runs, so repeats are required |
The two audits meet at the technical layer. A page that crawlers can’t fetch or read is a problem for both, so if your last technical review is more than a year old, work through a classic SEO audit checklist alongside this one. For the strategy behind the audit, start with what generative engine optimization is and why it matters.
Why AI visibility needs its own audit in 2026
Rankings no longer tell the whole story, because more searches end before anyone clicks. Studies published in 2026 by Ahrefs, SparkToro and Seer Interactive point the same way: clicks drop when an AI answer appears, and brands named inside the answer keep more value than brands listed below it.
The numbers are hard to ignore:
- Fewer clicks on top results. Ahrefs re-ran its study of 300,000 keywords in 2026 and found AI Overviews linked to a 58% lower average click-through rate for the top-ranking page, up from 34.5% in its first version.
- More zero-click searches. SparkToro’s analysis of Similarweb panel data found that 68% of US Google searches ended without a click between January and April 2026.
- Citations still earn clicks. Seer Interactive’s April 2026 study put organic click-through rate at about 3.3% on results with no AI Overview, about 2.1% for pages cited inside one, and about 0.9% for pages on the same results page that were not cited.
One more figure from Seer explains why this hits B2B companies hard. AI Overviews appeared on about 95% of comparison queries and 86% of question queries, but only about 5% of transactional ones. Comparisons and questions are exactly how buyers build a vendor shortlist. You can hold position 1 for “best [service] provider” and still lose the deal if the answer above you names three other firms.
What an AI visibility audit measures
A useful audit tracks five core metrics with fixed formulas: mention rate, recommendation rate, right-page citation rate, share of voice and accuracy. Sentiment and position add context. Write these definitions down before you run any prompts, so early results can’t quietly change how you score the later ones.
| Metric | What it tells you | How to calculate it |
| Mention rate | Are you in the answer at all? | Answers naming your brand / valid answers in the group |
| Recommendation rate | Are you presented as a good fit for the task? | Answers recommending you / valid answers |
| Right-page citation rate | Is the page built for this question being used? | Answers citing your target URL / valid answers |
| Share of voice | How do you compare with rivals? | Your mentions / mentions of you plus tracked competitors |
| Accuracy rate | Is what AI says about you correct? | Mentions with no factual errors / all your mentions |
A few scoring rules prevent most arguments later. Count each brand once per answer. A warning about your brand counts as a mention, never a recommendation. Log refusals and errors but keep them out of the denominator. An answer that names no vendor at all stays in, because “no one” is a real result.
How to run an AI visibility audit in 7 steps
The full process has seven steps: build a frozen prompt set, choose engines and conditions, run and record answers, score them, benchmark competitors and sources, check technical access, and prioritize fixes. LLM Pulse estimates 7 to 9 hours of focused work for 30 prompts across 5 engines, before analysis begins.
Step 1: Build a frozen prompt set
Write 30 to 50 prompts the way buyers actually ask, spread across five types, and tag each with the one page on your site that should answer it. Keep branded prompts in their own group. Then freeze the set with a version number and date, because edited prompts make results impossible to compare over time.
| Prompt type | What it tests | Example |
| Problem | Do you appear before buyers know the category? | “Why are our organic leads falling when our rankings look stable?” |
| Category | Do you make the shortlist? | “Best white-label SEO providers for small US agencies” |
| Comparison | How are you positioned against rivals? | “[Competitor A] vs [Competitor B] for a 20-person SaaS company” |
| Branded | Is what AI says about you accurate? | “What does [Brand] do, and who is it for?” |
| Fact check | Are key details right? | “Does [Brand] offer [service], and where is it based?” |
Unbranded prompts carry the main score. Branded prompts get their own score, since naming your brand makes an appearance almost certain. New prompt ideas go into a parking list and join the core set only at a dated revision, such as v1.1 next quarter.
Step 2: Choose engines, locales and test conditions
Test at least ChatGPT, Gemini, Perplexity and Google’s AI Overviews, because each builds answers differently and a brand can be strong on one and missing on another. For US buyers, add Google AI Mode and Microsoft Copilot, plus Claude if you sell to technical teams. Record locale, login state and search setting every time.
| Engine | How it builds answers | Audit note |
| ChatGPT | Model knowledge plus web results from OpenAI’s search index | Largest audience; record whether search ran |
| Google AI Overviews | Summaries on Google results pages, drawn from Google’s index | Appears on some queries only; test from a US location |
| Google AI Mode | Conversational search that splits a question into related sub-searches | Google says it can use different models and links than AI Overviews |
| Gemini | Google’s assistant app | Test separately from AI Overviews |
| Perplexity | Search-first answers with visible source lists | Useful for seeing which domains get cited |
| Claude | Model knowledge plus web search when enabled | Worth testing for technical and B2B buyers |
| Microsoft Copilot | Microsoft’s assistant, backed by Bing | Relevant where buyers work in Microsoft 365 |
Run every prompt in a fresh, logged-out session where the engine allows it, with a US locale. Run each prompt at least twice on different days. AI answers vary from run to run, so one answer is a sample, not a result.
Step 3: Run and record every answer
Log one row per prompt, engine and run, with the full answer, every source link and the test conditions. Good records let a colleague repeat the reading and let you re-score later without re-running prompts. Capture everything first and analyze afterward, so early answers don’t bias how you judge the rest of the set.
| Field | Example entry |
| Prompt ID and set version | P12, v1.0 (frozen October 20, 2026) |
| Date, time and timezone | 2026-10-21, 10:15, EDT |
| Engine and mode shown | ChatGPT, search on |
| Locale and session | en-US, logged out, new chat |
| Status | Answered, refused or error |
| Brands named, in order | Competitor A, Competitor B, your brand |
| Cited URLs | Every link shown with the answer |
| Target page cited? | Yes, other own page, or no |
| Accuracy notes | “Lists a service we stopped offering” |
| Evidence | Screenshot file name, such as p12-chatgpt-r1.png |
Step 4: Score mentions, citations and share of voice
Apply the metric definitions to every valid answer, by prompt group and by engine. Report unbranded and branded results separately, and read the target-page citation rate next to the mention rate. A brand that is mentioned often but rarely cited from its own pages has a source problem, not an awareness problem.
Here is an illustrative example with made-up numbers, not client data. Say you ran 30 unbranded prompts on 4 engines, twice each, for 240 answers. Twelve were errors or refusals, leaving 228 valid answers.
| Brand | Mentions | Mention rate | Share of voice |
| Competitor A | 120 | 52.6% | 42.9% |
| Competitor B | 74 | 32.5% | 26.4% |
| Your brand | 46 | 20.2% | 16.4% |
| Competitor C | 40 | 17.5% | 14.3% |
Your target pages were cited in 14 answers, a right-page citation rate of 6.1%. The reading is clear. Competitor A is the default answer, and AI describes you mostly from other people’s pages. That points straight at Step 5.
Step 5: Benchmark competitors and map cited sources
Pick the 3 to 5 competitors you lose deals to and record each of their appearances. Then pull every cited URL into one list, group it by domain and flag the sites that cite competitors but not you. That gap list, usually review platforms, roundups, directories and forums, is often the quickest route to more mentions.
Also note the label each engine attaches to each brand: “enterprise option,” “budget pick,” “best for agencies.” Those labels are your positioning in AI answers, and they often trace back to one or two influential sources. Reviews carry real weight here, which is why online reviews shape AI-powered search more than many teams expect.
Step 6: Check technical access
Confirm that AI search crawlers can reach and read the pages you want cited. Check robots.txt rules for each vendor’s search agent separately from its training agent, test key pages with JavaScript turned off, and keep structured data consistent with visible content. Google says its AI features need no special markup or AI text files.
| Vendor | Search and retrieval agents | Training agent | Effect of blocking the search agent |
| OpenAI | OAI-SearchBot; ChatGPT-User for user-triggered visits | GPTBot | Opted-out sites are not shown in ChatGPT search answers, though they may still appear as navigational links |
| Anthropic | Claude-SearchBot; Claude-User | ClaudeBot | Can reduce your visibility and accuracy in Claude’s search results |
| Googlebot (AI Overviews and AI Mode draw on the Search index) | Google-Extended | Blocking Googlebot removes you from Search entirely; Google-Extended does not govern AI Overviews or AI Mode |
The most common mistake is blocking every AI-related agent with one rule. Blocking GPTBot keeps your content out of OpenAI model training but leaves ChatGPT search alone. Blocking OAI-SearchBot is what removes you from ChatGPT search answers. Decide each line on purpose.
Then run two quick checks:
- Firewall and CDN rules. Bot-protection settings sometimes challenge AI crawlers even when robots.txt allows them. Look for search-agent hits in your server logs.
- JavaScript. Load key pages with JavaScript disabled. If service descriptions or pricing disappear, server-render that content, since some AI crawlers may not render pages the way Googlebot does.
Step 7: Prioritize fixes
Score each finding by likely impact and effort, then work the high-impact items in order: quick fixes this month, bigger content and PR projects this quarter. Most gains come from three places: answer-first pages for prompts where you are absent, corrected facts on third-party sources, and earned mentions on sites AI already cites.
| Impact | Low effort (this month) | High effort (this quarter) |
| High | Remove accidental crawler blocks; correct wrong facts on your About, service and pricing pages; update directory and review profiles; add a direct answer at the top of each tagged page | Build pages for unanswered problem and comparison prompts; earn coverage on sites that cite competitors; publish original data |
On your own pages, lead with the answer, then support it with specifics: who the service fits, what it includes and how it compares. Named authors and verifiable credentials help, which is why E-E-A-T signals belong in this work, and the page format overlaps heavily with answer engine optimization. If AI repeats wrong or damaging claims about you, treat it as a reputation issue as well as a content one; online reputation management covers the third-party side. For a team to run the full program, see our generative engine optimization services.
How to track AI traffic after the October 2026 ChatGPT change
In early October 2026, links in ChatGPT answers stopped carrying the utm_source=chatgpt.com tag and stopped passing a referrer. Many sites then saw ChatGPT sessions fall while Direct sessions rose on the same pages. Before reporting a visibility loss, annotate the change and check Direct traffic on the pages ChatGPT used to send visitors to.
The evidence lines up. DevTune tracked the share of links in ChatGPT answers that carried the tag and saw it fall from 99.7% to 0.1% overnight on October 8. Search Engine Journal reported that ChatGPT-attributed GA4 sessions fell by 22% to 57% for sites tracked by Graphite, while one analyst found Direct traffic rose on 2 out of 3 pages that lost ChatGPT visits.
Four steps keep your reporting honest:
- Annotate October 8, 2026 in GA4 and every dashboard that counts ChatGPT traffic.
- Compare ChatGPT plus Direct sessions on the landing pages that received ChatGPT visits in September.
- Report three stages separately: visibility (from your audit), visits (from analytics) and qualified leads (from your CRM). A citation can happen without a click.
- Add “AI assistant (ChatGPT, Gemini, etc.)” as an option in your “How did you hear about us?” form field.
If traffic is falling and you can’t tell whether tracking or demand is to blame, our guide on why website traffic goes down walks through the other usual causes. Connecting visibility to pipeline is the job of performance marketing and analytics.
AI visibility audit tools: manual vs platforms
A spreadsheet and a browser are enough for a first audit, and doing it by hand teaches you how each engine behaves. Platforms earn their cost once you track more prompts, competitors or engines than one person can re-run each month. Compare them on engine coverage, prompt limits, citation-source reporting and links to your SEO data.
| Option | Best for | Strengths | Watch-outs |
| Manual spreadsheet | First baseline, small prompt sets | Free; you read every answer yourself | Slow; hard to run weekly |
| Semrush AI Visibility Toolkit | Teams already using Semrush | AI visibility next to SEO data; site audit for AI crawler issues | Prompt limits vary by plan |
| Ahrefs Brand Radar | Competitive research | Large database of search-backed prompts; ties into backlink and keyword data | Fixes happen outside the tool |
| Profound | Enterprise programs | Deep citation and competitor reporting | Enterprise pricing |
Your 30-60-90 day AI visibility plan
Spend the first 30 days on the baseline and quick fixes, days 31 to 60 on owned content, and days 61 to 90 on third-party sources and your first re-audit. Judge progress on trend lines from the same frozen prompt set, not on any single answer, and expect some movement to come from model updates outside your control.
| Phase | Focus | Deliverables |
| Days 1 to 30 | Baseline and quick fixes | Prompt set v1.0 frozen; two full runs; crawler access fixed; wrong facts corrected on owned pages and profiles; GA4 annotation for October 8 |
| Days 31 to 60 | Owned content | Answer-first rewrites of tagged pages; new pages for the top unanswered prompts; markup matched to visible content |
| Days 61 to 90 | Sources and re-audit | Outreach to sites that cite competitors; review requests to recent clients; full re-run against the baseline; monthly cadence agreed |
Running AI visibility audits for clients (for agencies)
Agencies can turn this audit into a repeatable deliverable: a frozen prompt set per client, a monthly scorecard and a fix backlog that feeds existing SEO, content and PR retainers. The method stays the same. What changes is the packaging, the client-facing reporting and the capacity to deliver the fixes the audit uncovers.
Set expectations early. No one can guarantee that an assistant will name a brand, and answers vary between runs, so sell the measurement and the improvement process rather than a fixed outcome. If the backlog outgrows your bench, a white-label SEO partner can deliver the fixes under your brand.
Frequently asked questions
How often should you run an AI visibility audit?
Run a full audit at least quarterly, and re-run your 10 most important prompts monthly. AI engines update their models and sources often, so quarterly data ages quickly. If you’re actively investing in AI visibility, weekly tracking through a platform gives you faster feedback on what is working.
How many prompts does an AI visibility audit need?
For a first audit, 30 to 50 prompts usually covers your main services and buyer questions. Fewer than 30 can miss patterns, and more than 50 gets hard to run by hand. A business with one service can start smaller and expand at the next dated revision.
What is a good AI visibility score?
There isn’t a universal benchmark, because category competition varies widely. Compare your mention rate and share of voice with the 3 to 5 competitors you lose deals to, and focus on the trend across re-runs. A rising share of voice on unbranded prompts is the clearest sign of progress.
How is an AI visibility audit different from an SEO audit?
An SEO audit measures rankings, crawlability and backlinks for keywords on results pages. An AI visibility audit measures whether AI engines name, recommend and cite your brand in full answers to buyer questions. The two overlap on technical access, so run both.
Which AI platforms should a US business audit first?
Start with ChatGPT, Google AI Overviews, Gemini and Perplexity. Add Google AI Mode and Microsoft Copilot for broader US coverage, and Claude if your buyers are technical. Test each one separately, because a brand can be visible on one engine and missing on another.
Can I just ask ChatGPT to audit my brand?
You can ask ChatGPT what it knows about your brand as a quick accuracy check. That isn’t an audit, though. A single conversation doesn’t show how the engine answers unbranded buyer questions, how often competitors appear or how much results vary between runs.
Do schema markup or llms.txt guarantee AI citations?
No. Structured data helps engines read the facts on your pages, but no markup guarantees a mention or citation. Google says no special files or markup are needed for its AI features, so treat llms.txt as optional and put your effort into content and third-party sources.
How long does an AI visibility audit take?
A manual first audit of 30 prompts across 5 engines takes roughly 7 to 9 hours of execution, based on LLM Pulse’s estimate, plus time for analysis. Repeat runs go faster once the prompt set and log template exist. Platforms cut execution down to setup and review.
Conclusion
An AI visibility audit replaces guesswork with a baseline. Build a frozen prompt set, test it across the engines your buyers use, score the answers with fixed rules, and trace every gap to a cause you can fix. Keep branded and unbranded results apart, map the sources AI trusts, and read traffic data carefully after the October 2026 ChatGPT tracking change. Then re-run the same set on a schedule, because a single snapshot tells you where you stand, while the trend shows if your work is paying off.
Work with XCEEDBD
If you want help building your prompt set, running the baseline or fixing what it finds, our generative engine optimization services cover the audit through to content, technical and source fixes. Agencies can deliver the same work to their clients through our white-label SEO program.
Since 2008, XCEEDBD has delivered 3,100+ projects for 1,500+ clients across 38+ countries.
Book a free consultation – we respond within 8 business hours.
References
- Ahrefs (via Business Wire). “New research: Google’s AI Overviews now cost websites 58% of their clicks.” May 2026. https://www.businesswire.com/news/home/20260518322756/en/
- Search Engine Land. “Google AI Overviews CTR shows early signs of recovery: Study” (Seer Interactive data). April 2026. https://searchengineland.com/google-ai-overviews-ctr-recovery-study-475566
- TechWyse. “68% of U.S. Google Searches Ended Without a Click in Early 2026, SparkToro Data Shows.” 2026. https://www.techwyse.com/news/reports-data/google-zero-click-searches-68-percent-2026
- DevTune. “ChatGPT’s Tagged Links Fell From 99.7% to 0.1% Overnight.” October 16, 2026. https://devtune.ai/blog/chatgpt-stopped-tagging-links
- Search Engine Journal. “ChatGPT Traffic Drop Appears To Be A Tracking Problem.” October 2026. https://www.searchenginejournal.com/chatgpt-traffic-drop-appears-to-be-a-tracking-problem/593157/
- LLM Pulse. “How to Audit Your Brand’s AI Visibility (10-Step Guide 2026).” Updated September 14, 2026. https://llmpulse.ai/blog/ai-visibility-audit-guide/
- Google Search Central. “AI features and your website.” https://developers.google.com/search/docs/appearance/ai-features
- OpenAI. “Overview of OpenAI Crawlers.” https://developers.openai.com/docs/bots
- Anthropic. “Does Anthropic crawl data from the web, and how can site owners block the crawler?” https://support.claude.com/en/articles/8896518
- Mangools. “Google AI Mode and SEO: What is it and how it works?” 2026. https://mangools.com/blog/?p=11953
- The Commerce Shop. “10 Best AI Visibility Tools to Track Your Brand in AI Search Engines.” 2026. https://www.thecommerceshop.com/blog/10-best-ai-visibility-tools/
