ChatGPT Traffic: Can You Grow It? Everything We Did for LLMO

LLMO — what it means

LLMO is the practice of optimizing a site so that large language models such as ChatGPT cite and reference its information. It is also called AIO or GEO.

Hi, I’m Masato. In July 2026 we rebuilt our own site from scratch with AI, and at the same time rolled out “AI search optimization (LLMO)” across the whole blog. This article lays out every measure we implemented and the real numbers as they stand, with nothing hidden.

✓What’s in this article

• The real traffic numbers from ChatGPT (for a blog with 190,000 impressions a year)
• Every measure we implemented to get cited by LLMs
• How to measure AI-driven traffic in GA4

Let me start with the conclusion. I can’t yet say “yes, you can grow it.” Even on a blog with 190,000 impressions a year, AI-driven traffic is around 12 sessions a month. That is the honest state of things as of July 2026. What we have done is finish all the groundwork for being cited, and this article is the record of that implementation plus the method we’ll use to measure the effect from here.

The current baseline: 12 sessions a month from AI

Numbers first. Looking at referral sources in GA4, our blog gets around 12 sessions a month from AI. Almost all of it is ChatGPT.

For comparison, here’s ordinary search. Over the last 92 days in Search Console, the head keyword “webflow” sits at an average position of 3.9 with 681 clicks from 15,084 impressions. Set against search traffic, AI is still rounding-error scale.

Does that mean you can ignore it? I don’t think so, for two reasons.

• AI search will split sites into “the cited” and “the uncited.” My read is that the earlier you claim a seat, the better
• Someone who arrives after asking an AI already knows what their problem is. Small in number, potentially rich in intent

And there’s this: if an AI mentions your site inside its answer but nobody clicks the link, GA4 records zero sessions. So “12 sessions” might be the tip of an iceberg, or it might be the whole thing. That’s exactly why we decided to build the groundwork first and get ourselves into a state where we can measure.

LLMO: everything we did to get cited by LLMs

In July 2026 we rolled this out across all 63 blog posts at once (the per-article implementation detail is in AIO field report #1: the implementation). We had just rebuilt the entire site with AI (Claude Code) in seven days, so this rode on the back of that.

llms.txt

A text file at the site root that tells AI what’s here. It organizes “what this site contains” as Markdown an LLM can read easily, and we set it up alongside the sitemap and structured data during the rebuild.

To be honest, llms.txt is not a spec that the major AI companies have publicly committed to supporting (as far as I could find in July 2026). But it costs essentially nothing to produce, so I put it in the “no reason not to” bucket.

FAQPage structured data

We added an FAQ to all 63 articles and built a system that generates FAQPage JSON-LD automatically. Rather than hand-writing it per article, the article data carries structured fields and the template emits both the visible FAQ and the structured data at once.

The key point is that visible content and structured data must always match. Structured data with no counterpart in the body is not permitted under Google’s guidelines either.

One-sentence definitions (key terms)

For each article’s main term we added a self-contained one-sentence definition — “X is Y.” It’s generally said that when an LLM assembles an answer, it favors self-contained sentences it can quote as-is. And for human readers, having the term explained up front is simply considerate.

Takeaway boxes and Speakable

We put a box of key points at the top of each article and attached Speakable structured data too. The takeaway box exists to make the answer available without reading the whole article — in other words, in a form an AI can lift directly.

We fixed the foundations at the same time

Alongside the measures, we overhauled the quality of the articles themselves in one pass.

• Typo and inconsistency fixes: over 450
• Overlapping articles merged behind 301 redirects: three pairs
• AIO readiness on our own scoring rubric: 47.8 → 97.3 on average

The scoring uses a rubric we built ourselves (points for the presence of takeaways, FAQs, definitions and so on), so it’s an internal metric only. But as that pre-implementation average of 47.8 suggests, we started from a near-untouched state as far as AI search was concerned.

I think first-hand information matters most

All of the above is structure. What I consider most important is first-hand information — my own experience and my own numbers.

One article on our blog holds an average position of 1.7 with 162 clicks over 92 days for the Japanese-language search “address 最悪” (roughly “ADDress is terrible”). It’s just an honest write-up of what I felt after actually using the service. In search engines and in AI alike, “what someone who actually did it says” has no substitute. My position is that LLMO techniques only work once there is something worth citing.

Measurement: how to see AI-driven traffic in GA4

Measures you can’t measure are pointless, so here’s the how.

1. Create a free-form report under “Explore” in GA4
2. Add “Session source” as a dimension
3. Filter with a regular expression such as chatgpt|perplexity|copilot|gemini to catch AI domains together

Clicks that come via ChatGPT are recorded with chatgpt.com as the source (links often carry utm_source=chatgpt.com). Our own “around 12 sessions a month” came out of exactly this method.

One caveat: this method cannot measure zero-click citations. If an AI names your site or describes your content inside its answer and nobody clicks, it never shows up in the numbers. So alongside AI sessions, it’s worth watching the trend in branded searches for your company or product names.

So, can you grow traffic from ChatGPT?

Honestly, I don’t know yet. The implementation only finished in July 2026, so the effect has yet to show up in the numbers.

One thing I can say for certain: we finished the groundwork in a state where it can be measured. If traffic grows, we can work out what worked. If it doesn’t, we’ll have first-hand information that says “a blog with 190,000 impressions a year implemented LLMO in full and AI traffic did not grow.” Either way, it becomes an article.

I report on this experiment early in the newsletter series “A public experiment in growing a blog with AI.” If you want the next installment, head to the newsletter. If you’d like to talk through AI search optimization for your own site, ask us in a free consultation.

FAQ

How do I check traffic from ChatGPT in GA4?

Build a free-form report under “Explore”, use “Session source” as the dimension, and filter with a regular expression such as chatgpt|perplexity|copilot|gemini. Clicks via ChatGPT are recorded with chatgpt.com as the source.

Will adding llms.txt increase traffic from AI?

There is no guaranteed effect. It is not a spec the major AI companies have publicly committed to supporting, but it costs essentially nothing to set up, so we think it is worth having.

What should I do first for AI search optimization (LLMO)?

Write first-hand information — your own experience and your own numbers. Formal measures such as FAQs and structured data only work once there is something worth citing.

How much AI-driven traffic is there in practice?

About 12 sessions a month on the Webharu blog, which gets 190,000 impressions a year (as of July 2026). Almost all of it comes via ChatGPT, and it is still small compared with search traffic. Note that zero-click cases — where an AI describes your site inside its answer but nobody clicks — are not included in that figure.