Salesman Hal

AEO reporting beyond mentions: tying AI visibility to revenue

Most AEO reports stop at “your client was mentioned.” Clients pay for calls, quotes and booked jobs. This guide is for agencies: what visibility metrics can and can’t tell you, the numbers that connect AI to revenue, how to attribute honestly, and a monthly report structure you can use.

Updated

What visibility tracking measures

AI visibility tools answer one question: when people ask AI assistants about this category, does the client come up? Most work the same way. You choose a set of prompts (“best emergency plumber in Seattle,” “who does tankless water heater repair near me”), the tool runs them against one or more assistants on a schedule, and it parses the answers. From that you get:

  • Mentions: how often the client’s name appears in answers.
  • Citations: how often a page on the client’s site is linked as a source.
  • Share of voice: the client’s mentions or citations as a share of all brands in the same answers.
  • Position and sentiment: where in the answer the client appears, and how it’s described.

Some first-party data now exists too. Bing Webmaster Tools’ AI Performance report, in public preview since February 2026, shows how often a site is cited in Microsoft Copilot, Bing’s AI summaries and some partner integrations, which pages are cited, and the “grounding queries” the AI used to retrieve them. Since June 2026 it adds a Citation Share figure per grounding query. Google Search Console counts traffic from AI Overviews and AI Mode in its Performance report under the Web search type, alongside ordinary search.

This is valuable. It shows whether the client is in the conversation at all and whether the content work is moving that. Clients should keep seeing it.

The limits of prompt tracking

Be upfront with clients about what a visibility number is, and isn’t.

  • It’s a sample of prompts you chose. Real people phrase questions in endless ways and add details about their situation. A prompt set is a panel, not a census. Bing describes its own grounding-query data as “a sample of overall citation activity,” and that is first-party data.
  • Answers vary. Run the same prompt twice and you can get different brands. Assistants also break one question into several searches: Google says AI Overviews and AI Mode may use “query fan-out,” issuing multiple related searches across subtopics, so the pages cited can be broader than for a classic search.
  • Answers are personal. Assistants can tailor answers to the person asking. Since April 2025, for example, ChatGPT can draw on a user’s previous conversations to tailor its answers. A scheduled run from a tracking account can’t reproduce that.
  • Share metrics aren’t scoreboards. Bing calls Citation Share an “observational metric – not a ranking system or a competitive scoreboard,” and says it doesn’t represent traffic share.

Ask any visibility vendor how prompts are run: which assistants and models, from which locations, logged in or not, and how many runs per prompt. Report the method with the number, and keep the prompt set stable so month-on-month changes mean something.

The gap between a mention and an outcome

A mention is the start of a chain. For a mention to become revenue, something has to happen next:

  1. The assistant reads the client’s site, live, to answer a follow-up (“are they open Saturdays?”).
  2. The person clicks through, or the assistant acts for them: sends an inquiry, asks for a quote, books a time.
  3. The business gets the lead and wins the job.

Visibility tools see none of this, and web analytics sees little of it. Cloudflare’s data shows how few clicks AI platforms send relative to how much they crawl: in July 2025 it measured about 1,091 crawls for every referred visit from OpenAI, and about 38,066 for Anthropic. Those ratios count only visits that carry a Referer header, and Cloudflare notes that Claude’s native app doesn’t send one (and believes other native apps don’t either), so real referrals are higher than the ratio suggests. Either way, much of what AI does with a client’s business happens inside the assistant, on the client’s site or through its forms, not in a referral session you can see in GA4.

That’s the gap clients feel when they ask, “So what did we get for it?”

Metrics that tie AI to revenue

These are the numbers that sit between a mention and a sale. Most need server-side data; our guide to measuring AI agent traffic covers how to collect it.

Metric What it tells the client Where it comes from
Agent visits by purpose How often assistants read the site because someone asked, separate from indexing and training crawls Server or CDN logs, matched on user agent
Pages agents read Which services and pages people are asking assistants about Logs, filtered to user-triggered fetchers
Questions agents asked What prospects wanted to know, word for word, and which questions the site can’t answer An MCP server or on-site search that logs queries
Actions taken Searches, availability checks, inquiries, quote requests and bookings an agent made, including failures MCP tool-call logs, WebMCP tool calls
AI-sourced leads and bookings The leads with names on them Actions above, plus CRM fields for source
AI referral conversions People who clicked through from an AI answer and then converted GA4’s AI Assistant channel with key events set up

A few notes on collecting them:

  • Questions are the most useful thing you’ll show a client. A list of real questions, ranked by how often the site couldn’t answer them, is a content brief that writes itself, and it’s the clearest link between AEO work and the next month’s deliverables.
  • Actions need somewhere to happen. Most AI fetchers read HTML but can’t submit a form. An MCP server on the client’s site gives assistants a way to send inquiries and book, and logs each action, so “AI-sourced lead” stops being a guess. See our guide to MCP for business websites, and WebMCP for agents in the browser.
  • Referrals need key events. GA4’s default channel group includes an AI Assistant channel for visits from sources such as ChatGPT, Gemini, DeepSeek, Copilot and Grok. Without form, call and booking key events, it only counts sessions. OpenAI says ChatGPT adds utm_source=chatgpt.com to its links, which helps when the referrer is missing.
  • Carry the source into the CRM. Leads from AI referrals usually arrive through the normal contact form, and the source is lost once the form is submitted. A hidden field fixes that:
// Remember where the visit started, so the lead form can carry it into the CRM.
const params = new URLSearchParams(location.search);
const referrer = document.referrer ? new URL(document.referrer).hostname : '';
const source = params.get('utm_source') || (referrer !== location.hostname ? referrer : '');

if (source && !sessionStorage.getItem('first_source')) {
  sessionStorage.setItem('first_source', source);
}

document.querySelectorAll('input[name="first_source"]').forEach((input) => {
  input.value = sessionStorage.getItem('first_source') || '';
});

Add a hidden first_source input to the form and map it to a CRM field. Adding “an AI assistant (ChatGPT, Claude, Perplexity…)” to a “How did you hear about us?” question catches people who copied the business name out of an answer, which no tracking can see.

Attribution you can defend

AI attribution invites overclaiming. A report that credits every new lead to AEO work will be found out the first time a client checks. Sort results into three tiers and label them in the report.

  1. Direct: AI did it. An inquiry sent or a meeting booked through an agent-facing endpoint was made by an AI assistant by definition. Count these as AI-sourced without hedging, and say which assistant sent each one when it identified itself.
  2. Referred: AI sent the visit. A session that arrived from an AI assistant and then converted. That’s last-touch attribution: honest, but partial, because apps without referrers and people who searched the name later are missing.
  3. Influenced: the trend lines moved together. Citations rose and so did branded search, direct traffic or calls. That’s correlation. Seasonality, ad spend, a new review or a competitor closing can produce the same pattern.

Some habits that keep the third tier honest:

  • Use the right verbs. “AI assistants sent 6 leads” is tier one. “Branded search grew 12% in the months citations rose” is tier three. Avoid “drove” unless you can show the mechanism.
  • Compare against something. Year-on-year for seasonal businesses, or pages and services you didn’t work on as a rough control group.
  • Fix definitions and windows. Same prompt set, same agent list, same calendar month. When you change a method, say so in the report and don’t compare across the change.
  • Report the floor. Direct plus referred is a minimum. Say that the true number is higher and why, rather than inventing a multiplier.

A sample monthly client report

Lead with outcomes, because that’s what the client is paying for. Then show the activity that explains them, the visibility that explains the activity, and finish with what you’ll do next. One page per client, the same order every month.

Section What goes in it
1. Outcomes AI-sourced leads and bookings (tier one), AI referral conversions (tier two), each with last month’s figure. List the leads by name and service if the client is happy for you to.
2. What AI did Actions taken by type (searched the site, checked times, sent an inquiry, booked), and any that failed.
3. Who came, and why Agent visits split into answering someone’s question, indexing for search and training, with the top agents.
4. What people asked The top questions and searches agents brought, and the ones the site couldn’t answer.
5. Visibility Mentions, citations and share of voice for the tracked prompt set, with the method stated; Bing AI Performance citations if available.
6. Influenced trends Branded search, direct traffic or call volume next to visibility, clearly labelled as correlation.
7. Next month The content and fixes you’ll ship, drawn from sections 4 and 2: unanswered questions to answer, failed actions to fix, pages agents read that need updating.

Section 7 is what makes the report worth reading. It turns “here’s what happened” into “here’s what we’re doing about it,” and ties the retainer to work the client can see.

How Salesman Hal handles AEO reporting

Hal covers the part after the mention, on the client’s own site. It gives each client site an MCP server that answers from the client’s pages and sends inquiries, quote requests and bookings into the forms and calendars they already use, and it logs AI agent visits on the server. Per client, the dashboard shows visits by agent and purpose, the questions agents asked and the ones Hal couldn’t answer, the pages they read, every action through Hal, step-by-step agent sessions and whether each ended in a lead, and the leads and bookings themselves, tagged with the assistant that sent them when it identifies itself. Every client site sits in one agency dashboard, with a monthly email report comparing each number with the month before.

Hal doesn’t track mentions or citations, so keep your visibility tool for sections 5 and 6. See Hal for agencies, or scan a prospect’s site free.

FAQ

Should we stop reporting mentions and share of voice?

No. They show whether AI recommends the client, which is the top of the chain. Report them with their method, and put outcomes first.

Can we say AEO work drove a client’s revenue?

Only for what you can trace: leads and bookings an AI assistant made, and conversions from AI referral visits. Wider trends that rose alongside visibility are correlation; describe them as such.

Why don’t AI referrals in GA4 match what the client hears from customers?

Many AI apps don’t pass a referrer, so those visits look direct, and people often copy a business name out of an answer and search for it. GA4’s AI Assistant channel is a floor. A “how did you hear about us” option for AI assistants fills some of the gap.

Are Google AI Overviews counted in GA4’s AI Assistant channel?

No. Google says AI Overviews and AI Mode are excluded from it. In Search Console, their traffic is counted in the Performance report under the Web search type, together with ordinary search.

What’s the single most useful new metric?

The questions agents asked that the site couldn’t answer. It shows what prospects want to know, and it’s next month’s content plan.

How many prompts should we track?

Enough to cover the client’s main services, locations and buying questions, and no more than you can keep stable. Consistency month to month matters more than size.

Sources

Further reading

Show clients the leads AI sent them.

Hal shows what AI did on every client site, from first visit to booked meeting. Free during early access.

Or run the free AI-readiness scan on a client’s site first.