AI Search Is Now Sending Us Leads. Here's What We Changed.

In the last 30 days, AI assistants sent LUNA Systems 20 sessions and 3 leads, close to a 15% conversion rate against 4.6% for organic search. Every lead came through a blog post answering a specific question. Not one came through a service page. Here is what we measured, and what we changed.

Key takeaways

• AI assistants sent 20 sessions in the last 30 days (measured July 2026), producing 3 leads at close to a 15% session-to-lead rate, against 4.6% for organic search on the same site.
• Almost every citation lands on a blog post answering a cost, comparison, industry-specific, or compliance question. Service pages earned almost nothing.
• Two posts are each independently cited by three different AI engines: a virtual receptionist comparison and a tradies automation guide.
• A separate citation check across five question types found the site cited on four of five, measured 28 July 2026.
• What worked was answer-first openings, a takeaways box, question-form headings, a named author, and weekly tracking, not more service pages.

How much traffic do AI assistants actually send?

Twenty sessions in 30 days. That is a small sample and we would not stake a strategy on the raw number alone, the trend underneath it is what matters.

ChatGPT sent 15 sessions, up from 2 in the previous 30 days. Claude sent 4, up from 1. Gemini sent 1. Perplexity sent 0, down from 5 the month before, a reminder that these numbers move around and one engine's drop is not the whole picture.

All 3 leads, two completed voice demos and one contact form submission, came from ChatGPT sessions. That gave AI-assisted traffic a conversion rate close to 15%, against 4.6% for organic search over the same window. Small numbers, but a real gap worth watching rather than a fluke worth ignoring.

What pages do ChatGPT and Claude actually cite?

Almost entirely blog posts, and almost entirely posts answering one of four question shapes:

• Cost questions. "How much does X cost in Australia."
• Comparisons. "X vs Y", "X services compared."
• Industry-specific questions. "Business automation for tradies," not "business automation" in general.
• Compliance explainers. Plain-English answers to regulatory questions people actually type.

Two posts stood out because three different engines cited each of them independently: our virtual receptionist services compared piece and our business automation for tradies guide. Both answer a specific, narrow question in the title and open with the answer.

Service pages, the ones written to pitch a service rather than answer a question, told a different story: two of them earned one session each in the same period. That is not a rounding error, it is a pattern. A page built to sell reads to an AI assistant as marketing, not as an answer. A page built to answer a question reads as a source.

What did we change?

In order, this is what we did:

1. Wrote posts in the four winning shapes, and kept writing them, instead of adding more general service pages.
2. Moved the answer to the first 40 to 60 words. No scene-setting, no throat-clearing, the actual answer up front, written so it could be lifted straight into a chat reply.
3. Added a Key Takeaways box to every high-value post, a handful of self-contained sentences a model can quote without needing the rest of the page.
4. Switched headings to question form, matching the exact phrasing people type into a chat box rather than a marketing label.
5. Named a real author with a visible updated date on every post, ourselves included, rather than leaving posts anonymous.
6. Started tracking citations weekly instead of guessing which pages were being picked up.

None of it is exotic. It is the same discipline this post is trying to demonstrate: answer the question first, make the good part liftable, and say who wrote it.

What didn't work?

Two things we tried and dropped.

General service pages, written as a pitch for what we do rather than an answer to a specific question, did not get cited, regardless of how well they were written for a human reader. AI assistants are not extracting a persuasive paragraph, they are extracting a claim.

Chasing "best agency" or "top provider" style queries also went nowhere. Those questions invite a listicle answer pulled from review aggregators and directories, not a single business's own page, so writing for that phrasing is effort spent on a format that was never going to cite us.

How do you measure it?

Two signals, checked together, because either one alone is misleading.

GA4 session source data shows what AI referral traffic is actually worth, sessions, and what those sessions turn into, leads. This is the outcome number, and it is what caught the ChatGPT jump before we would have noticed it any other way.

A citation check is the other half: asking each engine the target questions directly and recording whether the site gets named, and for which page. Ours ran across five question shapes and found the site cited on four of five (ChatGPT and/or Perplexity), measured 28 July 2026. This is the diagnostic number, it tells you which pages and passages are working, not just that something is.

Neither number by itself is a strategy. Session data with no citation check tells you traffic moved but not why. A citation check with no session data tells you a page got named but not whether anyone acted on it.

The honest read

Twenty sessions and three leads will not change a business on their own. What makes them worth writing about is the shape underneath the numbers: a new source of traffic that did not exist in the same volume the month before, converting better than our best existing channel, and confirmed independently by a citation check on the pages we expected. That is worth a change in what gets written, not a headline.

If you want the pages that decide whether AI search names your business built the same way, that is core SEO services work now, not a separate discipline. And if you are curious what the version of this looks like on the phone rather than the page, our AI voice options show the same answer-first thinking applied to a call instead of a search result.

Frequently Asked Questions

How much traffic do AI assistants actually send to a website?

On LUNA Systems' own site, AI assistants sent 20 sessions in the last 30 days, measured July 2026, up sharply from ChatGPT and Claude the previous 30 days. Those 20 sessions produced 3 leads, close to a 15% session-to-lead rate, against 4.6% for organic search on the same site.

How do you get cited by ChatGPT?

Answer the question directly in the first 40 to 60 words, add a short takeaways box of self-contained sentences near the top, use question-form headings, name a real author, and show a visible updated date. On LUNA Systems' own posts, pages built this way are the ones ChatGPT and Perplexity cite.

Do service pages or blog posts get cited more by AI search?

Blog posts, by a wide margin, on LUNA Systems' own measurement. Posts answering a cost, comparison, industry-specific, or compliance question earned almost all of the citations. Two general service pages earned one session each in the same period.

What kind of questions get a page cited by AI search?

Four shapes: cost questions, comparisons, industry-specific questions, and compliance explainers. A page written as a general pitch for a service did not fit any of the four and was not cited.

Is 20 sessions from AI search a meaningful number?

On its own, no, it is a small sample. What makes it worth acting on is the shape of it: three assistants sending traffic where two sent almost none the month before, close to a 15% conversion rate against 4.6% for organic search, and a separate citation check confirming the site is cited on 4 of 5 tested question types.

How do you measure whether your site is being cited by AI search?

Two ways, used together. GA4 session source data shows how much traffic each AI assistant sends and what it converts to. A citation check, asking each engine the target questions directly, shows which specific pages are being picked up. Neither alone is enough.