What does an AI visibility audit actually measure?

Why one number is not enough

Start with what the AI companies publish. OpenAI lists separate agents for separate jobs: OAI-SearchBot for search, GPTBot for collecting training data, and ChatGPT-User for pages fetched when a user asks [1]. Anthropic lists ClaudeBot for training, Claude-User for user requests and Claude-SearchBot for search quality [2]. Perplexity lists PerplexityBot for search results and Perplexity-User for user actions [3]. Each one can be allowed or blocked separately in robots.txt, the file that tells crawlers what they may fetch [1] [2] [3]. So “can AI read my site?” is already three questions, not one.

Google says its AI features (AI Overviews and AI Mode) need nothing extra. A page must be indexed and eligible for a snippet to appear as a supporting link. Traffic from these features is counted inside Search Console’s “Web” search type [4]. Google’s AI traffic therefore sits inside your ordinary search numbers.

Layer 1: access

What to measure: which AI user agents (the name a crawler sends with each request) asked for which pages, and what they got back: served, redirected, missing, or refused.

Where it comes from: your server logs, or a tool that reads them.

Limit: a fetch shows that a page was read. It does not show that the page was used in an answer.

Layer 2: mention

A mention is your brand name in the answer text. The GEO paper (generative engine optimization) measured visibility inside generated answers in two ways. One was position-adjusted word count: how much of the answer is attributed to a source, weighted by where it appears. The other was a judged “subjective impression” [5]. Those metrics fit a lab setting. In an audit, the simpler question is how often you are named across a fixed set of questions.

Limit: GEO tested a GPT-3.5-based engine and Perplexity in 2023–24. The results may not carry over to current engines [5].

Layer 3: citation

A citation is a link or reference to your page as the source of a statement. Mention and citation can come apart. A page can be cited without the brand being named, and a brand can be named with no link to it.

Citation quality also needs checking. Liu, Zhang and Liang audited four generative search engines. Only 51.5% of generated sentences were fully supported by their citations, and 74.5% of citations supported the sentence they were attached to [6]. A 2024 audit of three answer engines measured citation accuracy between 49% and 68% [7]. A good audit therefore records whether the cited sentence actually matches what your page says, not only whether your URL appears.

Layer 4: arrival

Arrival means people reaching your site from an AI product, seen as referral sessions in your analytics. Being cited and being visited are different outcomes. Pew Research Center studied the Google browsing of 900 U.S. adults. Users clicked a traditional result in 8% of visits when an AI summary appeared and in 15% when none appeared. Only 1% clicked a link inside the summary [8].

What an audit cannot measure well

• Repeatability. Atil et al. ran five language models, set to be deterministic, ten times on eight benchmark tasks. Accuracy varied by up to 15% between runs [9]. One run of a question is one sample, not a fact about the engine.
• Cause. C-SEO Bench found most conversational SEO methods largely ineffective, with gains shrinking as more sites adopt them [10]. Wan, Wallace and Klein found models lean on a page’s relevance and largely ignore style [11]. An audit shows where you stand today. On its own it cannot prove that a change you made caused a move.
• Demand. An audit of answers does not tell you how often anyone asks the question. We give no volumes here.

What our own dashboard shows

These are real figures measured on our own domains, not customer results. The Axis dashboard, read on 8 October 2026, shows, for a 28-day window: 2.9K non-branded impressions and 14 non-branded clicks, each with “28 of 28 days measured”. It also shows 63 sessions and 0 sessions from AI assistants, each with “19 of 28 days measured” [12]. Another Axis view groups crawler fetches as search, user action and training [13]. That is the same split the vendors document [1] [2] [3]. Read together, the layers stay separate: a crawler reaching a site and a person arriving from an assistant are different numbers, and each needs its own row.

Map

Look. Which AI crawlers reach which pages, and what they get back.

Find. For a fixed set of questions, where you are named, where you are cited, and where the citation does not match your page.

Align. Fix access first. Then work on the pages that answer the questions you want to hold. Then measure again the same way.

References

1. OpenAI. (n.d.). Overview of OpenAI crawlers. Accessed 9 October 2026. https://platform.openai.com/docs/bots
2. Anthropic. (2026, April 7). Does Anthropic crawl data from the web, and how can site owners block the crawler? Claude Help Center. Accessed 9 October 2026. https://support.claude.com/en/articles/8896518-does-anthropic-crawl-data-from-the-web-and-how-can-site-owners-block-the-crawler
3. Perplexity. (n.d.). Perplexity crawlers. Accessed 9 October 2026. https://docs.perplexity.ai/guides/bots
4. Google Search Central. (n.d.). AI features and your website. Accessed 9 October 2026. https://developers.google.com/search/docs/appearance/ai-features
5. Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., & Deshpande, A. (2024). GEO: Generative engine optimization. KDD ’24, 5–16. https://doi.org/10.1145/3637528.3671900
6. Liu, N. F., Zhang, T., & Liang, P. (2023). Evaluating verifiability in generative search engines. Findings of EMNLP 2023. https://aclanthology.org/2023.findings-emnlp.467/
7. Narayanan Venkit, P., Laban, P., Zhou, Y., Mao, Y., & Wu, C.-S. (2024). Search engines in an AI era: The false promise of factual and verifiable source-cited responses. arXiv:2410.22349. https://arxiv.org/abs/2410.22349
8. Chapekis, A., & Lieb, A. (2025, July 22). Google users are less likely to click on links when an AI summary appears in the results. Pew Research Center. https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/
9. Atil, B., Aykent, S., Chittams, A., Fu, L., Passonneau, R. J., Radcliffe, E., Rajagopal, G. R., Sloan, A., Tudrej, T., Ture, F., Wu, Z., Xu, L., & Baldwin, B. (2024). Non-determinism of “deterministic” LLM settings. arXiv:2408.04667. https://arxiv.org/abs/2408.04667
10. Puerto, H., Gubri, M., Green, T., Oh, S. J., & Yun, S. (2025). C-SEO Bench: Does conversational SEO work? NeurIPS 2025 Datasets and Benchmarks Track. https://arxiv.org/abs/2506.11097
11. Wan, A., Wallace, E., & Klein, D. (2024). What evidence do language models find convincing? ACL 2024 (Long Papers). https://aclanthology.org/2024.acl-long.403/
12. Axis. (2026, October 8). Dashboard reading for Axis’s own domains, 28-day window. Internal measurement, not a customer result.
13. Axis. (n.d.). Product view grouping crawler fetches as search, user action and training. Internal; the site and date are not shown on the image.

Which of these four layers can you see for your own site today? Tell us where you want to stand. Or continue with how to check whether an answer cites you.