The measurement gap behind "why doesn't ChatGPT recommend us?"
Every client conversation this quarter eventually lands on the same question. Why doesn't ChatGPT recommend us. We can never answer it properly, because the measurement underneath is broken.
Referrers from ChatGPT and Perplexity are inconsistent, a lot of it arrives as direct traffic, and so the honest answer to "how many of our signups touched an AI tool first" is usually that we don't know.
What shifted my thinking: one of the GEO tools that launched here recently publishes a source-mix breakdown of what AI answers actually cite. Reddit was 30% of it. Review sites and comparison pages took another large share. Owned content was the smallest slice of the chart.
If that's even roughly right, then the work isn't publishing more of your own pages. It's showing up in the places the model already trusts. That's an uncomfortable strategy to accept, because you control almost none of it.
So two things I'd genuinely like to hear from people here. Has anyone attributed signups to an AI tool with any real confidence, and how did you do it? And are you doing anything deliberate to appear in those answers, or leaving it to chance?
Replies
16,976 outbound clicks logged, 11,155 flagged as bots, 5,821 left over. That looks like an audience until you bucket the survivors by day: 4,336 of them land inside a single three day window, hitting 674 distinct products from 768 addresses. Something walked the catalogue. The total was never the instrument, the distribution was, and every figure now gets date bucketed before it reaches a report, with a single window spike labelled as crawling instead of folded into the baseline.
Which is the measurement gap sitting behind the ChatGPT question. A visit that starts in a chat arrives with no referrer, so it cannot be attributed directly and you are left inferring from source mix. On ours, chat has passed search as the largest single source, and that reading is only ever as good as the bot filter underneath it. So before asking why a model does not recommend you, date bucket whatever number convinced you that you had visibility, because the question cannot be answered on top of a baseline made of crawlers.