- In a SparkToro study, ChatGPT and Google’s AI almost never returned the same list of brands twice across 100 runs: an audit needs repeated runs.
- Google AI Overviews and AI Mode shared only 13.7% of cited URLs for the same query in an Ahrefs study, so audit them separately.
- In our US measurement, reddit.com was cited on 59 of 65 buyer questions about GEO and youtube.com on 53 of 65.
- Count named, cited and absent separately, and keep questions that contain your brand name apart.
What does an AI visibility audit cover?
Three layers: what AI engines answer to your buying questions, which sources they rely on, and whether your site can be read and quoted. Many self-run audits stop at the first layer, which shows where you stand but not why.
- Answers. For each question and engine: are you named, is one of your pages cited, which competitors appear instead.
- Sources. The domains cited under those answers, ranked by how many questions they appear on.
- Site. Crawler access, indexing, readability without JavaScript, answer structure, structured data.
The output is a baseline you can repeat later with the same protocol, and a ranked list of actions. If you want to see the method on your own questions before doing it at scale, our free Quick Scan runs a short version.
Which questions should the audit use?
The questions your buyers ask before a shortlist, in their own words: 25 to 50 for a first audit, grouped by intent, with questions that contain your brand name kept apart. Fix the list before the first run.
- Category discovery. “Best ITSM software for a mid-size company.”
- Use case. “Tool to automate employee onboarding requests.”
- Alternatives. “Alternatives to a known vendor for a smaller team.”
- Comparisons. “Vendor A vs Vendor B for a regulated industry.”
- Integrations and compliance. “Works with Microsoft Teams, SOC 2 certified.”
- Branded questions. “Is your product a good fit for…” Kept separate, because they measure reputation, not discovery.
Build the list with sales: the words buyers use on first calls are closer to what they type into ChatGPT than your keyword list. Changing questions between audits breaks every comparison.
How do you run the questions and record the results?
Ask each question several times on each engine, with search enabled where it exists, and keep every answer in full. Then count named, cited and absent separately, engine by engine, and never average engines into one score.
Repetition is not optional. In a SparkToro study, ChatGPT and Google’s AI almost never returned the same list of brands twice across 100 runs [1]. Treat Google’s two AI surfaces as two engines: Ahrefs found only 13.7% of cited URLs in common between AI Overviews and AI Mode for the same query [2]. Record these metrics, each with its definition written down:
| Metric | Definition | What it tells you |
|---|---|---|
| Named rate | Answers that name your brand, over all answers for that engine | Whether you make the shortlist |
| Cited rate | Answers that link to one of your pages as a source | Whether your content is used as evidence |
| Absent | Answers with neither | Where the gaps are, question by question |
| Brands named instead | Every other vendor named, with frequency | Who wins the questions you lose |
| Cited domains | Each domain, with citations and number of distinct questions | Where you need to be present |
| Branded vs unbranded | The same metrics split by whether the question contains your name | Reputation versus discovery |
Which sources do AI engines rely on for your questions?
Log every cited link, then count for each domain its total citations and the number of distinct questions it appears on. A domain present across many questions is part of the engine’s core reading list for your category.
Here is what that looks like on a real corpus. On November 8, 2026, we asked 65 US buyer questions about GEO, this one included, on ChatGPT, Claude, Gemini and Perplexity, 5 runs each, for 1,300 answers + 325 Google searches [10]. reddit.com was cited on 59 of 65 questions (251 citations), youtube.com on 53 of 65 (235), en.wikipedia.org on 36 of 65, linkedin.com on 34 of 65 (98), searchengineland.com on 26 of 65 (108) and medium.com on 21 of 65 (73). The answers named 57 distinct companies. Your market will produce a different list, which is exactly why the audit has to measure it.
What should the technical part of the audit check?
Whether AI search crawlers can reach your pages, whether those pages are indexed and eligible for snippets, and whether the main content is readable without JavaScript and structured as answers.
Check each engine’s search crawler as its vendor documents it: OpenAI [3], Anthropic [4] and Perplexity [5]. For Google’s AI features, Google says there is no requirement beyond being indexed and eligible for a snippet [6]. Then the rest of the list: CDN and firewall rules that block bots despite robots.txt, main content and pricing in plain HTML, an answer near the top of each key page under question-style headings, a named author, a visible update date, valid structured data, and an up-to-date sitemap submitted to Google and Bing, with IndexNow for changes [7]. An llms.txt file is a 2024 proposal [8]: cheap to add, but no measured effect has been published.
What do you do with the audit results?
Turn each absence into one action: a page to write, a source to work, or a technical fix, ranked by how many questions and engines it affects. Then re-measure with exactly the same questions, engines and number of runs.
Technical blockers come first, because they cap everything else. Next come the questions where competitors are named and a cited source is within reach: a list you can join, a review profile to complete, a comparison to publish. Record AI-sourced traffic from the start as well: Google Analytics added AI Assistant traffic measurement on May 13, 2026 [9]. Re-measure only after new content has had time to be picked up, then compare against a noise threshold set before the comparison, not after.
How does ZivRank run an AI visibility audit?
At two levels: a free Quick Scan on up to 5 of your buying questions, with a video walkthrough by an expert within 48 hours, and a baseline audit of 50 questions on 5 engines, 5 runs each, included in the 6-month program.
The baseline audit covers 1,000 answers + 250 Google searches, on ChatGPT, Claude, Gemini, Perplexity and Google AI Overviews. It combines the answers, the cited sources, the competitors named instead of you and a technical audit of your site. It costs $2,000, billed at the start and included in the program. You receive the raw answers, and the noise threshold of ±15% is set before any comparison. Programs start at $2,300 per month, with our team writing, fact-checking and publishing the content the audit calls for. To track visibility without production, the Scan Continu plan (waitlist) covers 25 questions on 3 engines from $390 per month. Start with the free Quick Scan.
Frequently asked questions
Is there a free AI visibility check for my brand?
Yes. Several tracking tools offer free self-serve trials, and some agencies run a free diagnostic. The most useful test uses your own buying questions, several engines and several runs per question. At ZivRank, the Quick Scan is free: up to 5 of your buying questions, with a video walkthrough by an expert within 48 hours.
What should a technical GEO audit check?
Crawler access for each engine’s search bot, in robots.txt and in CDN or firewall rules. Snippet eligibility and indexing on Google and Bing. Main content readable without JavaScript. An answer near the top of key pages, question-style headings, a named author, an update date and valid structured data. An up-to-date sitemap, with IndexNow for changes.
How often should I repeat an AI visibility audit?
Often enough to see the effect of what you publish, with the same protocol each time. New content takes weeks to be picked up, so a monthly re-measurement is a common rhythm. Keep the questions, engines and number of runs identical, or the comparison means nothing.
Can I run an AI visibility audit myself?
Yes, on a small scale. A spreadsheet, 20 buying questions, four engines and five runs each give you 400 answers to log by hand. Beyond that, manual work becomes slow and error-prone, which is where tracking tools or an agency help. Whatever you choose, keep the raw answers.
Sources
Every figure in this article comes from the sources below, checked on November 18, 2026. Bracketed numbers in the text refer to this list.
- [1] SparkToro (Rand Fishkin), NEW Research: AIs are highly inconsistent when recommending brands or products, January 27, 2026.
- [2] Ahrefs (Despina Gavoyannis), Are AI Mode and AI Overviews Just Different Versions of the Same Answer?, December 15, 2025.
- [3] OpenAI, Overview of OpenAI Crawlers, checked November 9, 2026.
- [4] Anthropic, Does Anthropic crawl data from the web, and how can site owners block the crawler?, checked November 9, 2026.
- [5] Perplexity, Perplexity Crawlers, checked November 9, 2026.
- [6] Google Search Central, AI features and your website, last updated December 10, 2025.
- [7] IndexNow, IndexNow protocol, checked November 9, 2026.
- [8] Jeremy Howard (Answer.AI), The /llms.txt file, 2024 proposal.
- [9] Google Analytics Help, What’s new in Google Analytics (entry “New AI Assistant traffic measurement”), May 13, 2026.
- [10] ZivRank, US measurement: 65 buyer questions about GEO, ChatGPT, Claude, Gemini and Perplexity × 5 runs (1,300 answers) + 325 Google searches, November 8, 2026.