I measured whether ChatGPT recommends my SaaS. It didn't. Here's what actually decides.

I measured whether ChatGPT recommends my SaaS. It doesn't. Here's what actually decides.

Your product is probably invisible to AI assistants right now, and your website is not the main reason. I know because I measured mine, and the result was a zero I want to walk you through, because the mechanics I found apply to every early-stage product.

I run Respondyr, a review management tool for local businesses. This week I ran an experiment: ten buyer-intent questions ("best review management software for a small business," that kind of thing) against ChatGPT, Claude, and Gemini, several repetitions each, 103 answers total, with every brand mention and cited source logged.

My product appeared in zero of the 94 non-branded answers. My competitors appeared in almost all of them. When I asked about Respondyr by name, all three engines described it accurately and cited my site. So the engines know what I am. They just never think of me.

Here is what the data taught me about why, in three parts.

The engines mostly answer from memory, and you cannot optimize memory

The biggest surprise in my 103 answers: ChatGPT ran a live web search on only 4 of 35 queries. The other 31 times it answered from whatever it learned in training. Claude searched almost every time. Gemini was in between.

Think about what that means. For the engine with the most users, your beautiful landing page is irrelevant to most recommendation answers, because no retrieval happens at all. The recommendation comes from how often your product appeared across the wider internet months ago: forums, directories, comparison posts, reviews. The products getting recommended from memory are the ones that were everywhere back when the model was trained.

You cannot fix that with on-site work this quarter. You fix it by accumulating third-party presence that will be in the next training run, which is slow, which is exactly why starting now matters. The listings you create this month are seeds for answers a year out.

When engines do search, they read a tiny set of pages, and you can be on them

For the answers that did involve live search, I logged every cited source. The pattern was narrow. G2 got cited constantly. A handful of comparison blogs got cited repeatedly. Random homepages almost never appeared. Other research backs this up: one study of 548,000 retrieved pages found only about 15 percent of what engines fetch ever survives into the final answer.

Two practical details from the logs. First, the engines append the current year to their own searches ("best X software 2026"), so pages with a visible current year and a real updated date have a structural edge. Second, cited pages overwhelmingly answer the question in their opening section. If your comparison page builds suspense before the verdict, the engine quotes someone else.

So the checklist for the searchable half is short: be present on the two or three directories your category's answers actually cite (measure this, do not guess), and make sure the pages about you state their answer in the first hundred words.

Your baseline is free to measure, and unmeasured means invisible by default

Everything above cost me about four dollars in API calls and an afternoon. Ten questions your buyers would ask, three engines, a few repetitions each, count who gets named. That is the whole method. One repetition is not enough, by the way. Identical questions produce different brand lists run to run, which also means any tool selling you a one-shot "AI visibility score" is selling you a screenshot of a dice roll.

What the baseline buys you is direction. Mine told me my brand-name recall was already solved, my category presence was zero, and the surfaces that would change that were directories and answer-shaped comparison content, in that order. Without the measurement I would have guessed wrong and polished my homepage.

The zero stung. But a measured zero beats an imagined ranking, because a zero with a mechanism attached is a to-do list. Six months from now I will re-run the same 103 answers and publish the delta, whichever way it goes.

One offer, since the harness is sitting right here: I run this same scan for other products now. Your buyer questions, three engines, repeated runs, who gets named and which sources got cited. If you want your baseline without building it yourself, email me at [email protected] and I will run it.

Travis Bridle is the founder of Respondyr (respondyr.com), review management software for local businesses, and runs AI visibility scans (respondyr.com/features/ai-visibility). He is documenting the climb out of the zero as it happens.