What gets measured gets recommended
Prompt testing is the closest thing to a ranking report
Classical SEO has rank-tracking tools that report a site's position for a given keyword on a given search engine. The AI equivalent is prompt testing: a small, repeatable set of representative queries run against multiple assistants — ChatGPT, Claude, Gemini, Perplexity, Copilot — at a regular interval, with the results logged. The right test queries are those a real customer would type: 'best CRM for HVAC contractors in Texas', 'how do I clean a cast-iron pan', 'open-source vector database with low cost'. Recording who appears, in what order, and what is said about each is the basic visibility metric for AI search. A spreadsheet updated monthly is enough to start.
Server logs reveal the agent traffic that analytics misses
Most JavaScript-based analytics products only count visitors that execute their tracking script. Most AI crawlers and many agents do not. The result is that AI traffic is largely invisible in the dashboards most teams look at. Server logs — the raw access logs every web server keeps by default — do see this traffic. Filtering for known AI user-agent strings (GPTBot, ClaudeBot, Perplexity-Bot, OAI-SearchBot, ChatGPT-User, Google-Extended, Applebot, AI2Bot, and others) reveals how often those crawlers visit, which pages they load, and how their visit patterns change over time. A weekly grep through server logs is one of the most underused diagnostic habits in current SEO.
Citation tracking watches what AI assistants actually quote
When an AI assistant cites a page in its answer, the citation often produces a click — a referrer-traceable visit that does show up in classical analytics, tagged as coming from openai.com, perplexity.ai, claude.ai, or one of their domains. Tracking these referrers monthly produces a second visibility signal: not just 'do we appear when prompted' but 'are we being read after appearing'. Combined with prompt testing, citation tracking forms a complete picture of AI search performance. Both are free; together they take less than two hours a month for a small site.
Agent-readability scoring frameworks are the third layer
Vercel's agent-readability framework, the schema validators run by Google and Schema.org, accessibility audits from Lighthouse and axe, the directories of SEO tools and monitoring tools, and a growing set of dedicated AI-SEO scoring services produce a numeric score for any URL. Run regularly, the score acts as a leading indicator: improvements show up in the score weeks before they show up in citations. The score does not replace prompt testing or citation tracking, because it measures readiness rather than outcome, but it is the cheapest and most actionable of the three. Most teams should run all three layers together — readiness, citations, prompts — and trust the combined picture rather than any single number.
Things to consider — attribution from visibility to revenue
Visibility metrics tell you whether AI assistants are citing the site. They do not tell you whether those citations are turning into customers. The bridge is attribution. Three steps make a basic attribution layer. Tag links pushed by AI assistants where possible: most citations arrive with an identifiable referrer (openai.com, perplexity.ai, claude.ai, gemini.google.com), letting analytics tag those visits as 'AI-attributed'. Build a custom segment in classical analytics for that referrer set and report 'AI-attributed sessions' and 'AI-attributed revenue' alongside the usual organic-search figures. Acknowledge the limits: agents that complete a transaction inside a chat (via ACP or AP2) bypass the brand's site entirely and cannot be measured this way. For those, the best available signal is the merchant-side report from the platform itself — Stripe, Shopify, or the ACP partner. Imperfect attribution still beats reporting visibility alone.
Open ChatGPT, Claude, Gemini, and Perplexity. Ask each of them the most obvious question a customer in your category would ask. Does your brand appear? In what position? With what claim attached?