How I measured Pepys’s 8.6× ChatGPT referral growth

Pepys had 8.6× as many ChatGPT-referred sessions in September 2026 as in July. That is a useful growth story. It is also easy to tell it badly.

A founder sees a screenshot of their brand in an AI answer. Someone on the team sees ChatGPT in their referrer report. Search Console shows another number. By Friday, all three are in the same slide, labeled "AI visibility."

I would throw out that slide.

An appearance, a visit and a useful product action are different events. They happen in different places, and they need different instruments. I've had to keep them separate while looking at search for QuoteSweep and Pepys, two products I build.

Start with the question behind the number

If the question is "Are our pages appearing in Google's AI features?", use Search Console. Google now has a Generative AI performance report for impressions in AI Overviews and AI Mode. It can show which pages earned those impressions, and how the count changed over time. Google says it rolled the report out globally on 31 August 2026, although an individual property may still lack access or enough data.

That update retired an easy talking point. Saying "GSC doesn't show AI search" is now too broad. It does show one important part of Google AI search.

An impression still isn't a visit. And Google's report does not cover what happened inside ChatGPT, Claude or Perplexity. For those, I look elsewhere.

A ChatGPT referrer is a visit, not a citation

PostHog's web analytics can break sessions down by entry referrer. For Pepys, the published case file counts pageview sessions whose entry referring domain was chatgpt.com or chat.openai.com.

Using that definition, the current ankur.works case file reports 8.6 times as many ChatGPT-referred sessions in September 2026 as in July. Useful? Yes. Proof that Pepys appeared in 8.6 times as many AI answers? No. The session data cannot tell us which answer sent each person, and some AI-influenced visits may arrive without a usable referrer.

The definition is the point. A chart without the referrer rule, timezone and complete date range is an attractive guess. I reran the PostHog query for complete UTC months: 190 sessions in July, 948 in August and 1,637 in September. September is 8.6 times July after rounding.

If you want citations, collect citations

I use a fixed set of buyer questions, then capture the answers on a schedule. Record the prompt, engine, date, location or account context where relevant, visible mention, linked source and screenshot. Repeat the same set. That gives you an observation series someone else can inspect.

QuoteSweep has been observed in Google AI Overview comparisons around commercial raters, EZLynx alternatives and Bold Penguin alternatives, according to the dated captures described in the case file. Those captures establish that the pages appeared in those sampled answers. They do not establish a global citation rate.

One nice screenshot is worth keeping. It is not a denominator.

Follow the visitor into the product

The last line on my scorecard is the one a buyer cares about: did anyone sign up, complete the core job or send an inquiry? Those actions live in product analytics or a CRM. They deserve their own definition.

Pepys also grew in completed transcriptions during the same period as its ChatGPT referrals. I would show both trends. I would not claim the referred sessions caused the product growth without a joined journey that supports that conclusion.

Here is the monthly scorecard I would actually use:

You do not need a composite "AI score" to decide what to do next. If impressions rise and visits don't, inspect the page and question context. If visits rise and activation doesn't, inspect the landing experience. If the brand is absent from important comparison answers, make the underlying source material better and test again.

Four honest lines beat one impressive number with no definition.

If you want me to look at the question buyers ask before they find your product, send me your site. I'll tell you which of these measurements I would start with.

Sources and checks