GEO is now on almost every marketing budget. What that budget delivers is a good deal harder to see, because your dashboard barely moves. That is not proof that it is not working, it is a sign that you are looking in the wrong place. Most of what an AI answer sets in motion happens without a click, and that is exactly the part that counts nowhere. Not because it does not happen, but because nobody records it.
What you can do is measure what happens after that answer, and make sure models know enough about your products. These are the seven things to account for, from the figures you put side by side to the fields your product feed needs to fill.
1. Visibility is not your KPI
You appear in an AI answer. And then? The user clicks through, asks a follow-up question, googles your brand name, types in your homepage, or goes off to compare the product on Google. Each of those follow-up actions does leave a trace: LLM traffic and revenue, branded search volume in Search Console, direct traffic to the homepage, and uplift in SEO and SEA on the pages about that topic. Put AI visibility alongside those lines month after month and the shape of the graphs tells you whether GEO is paying off. In one case that came to 31% of revenue, worth 476,000 euros in AI-influenced revenue.
2. Being named is not being recommended
An answer names five brands and recommends two of them. That is the difference between the shortlist and the decision list. In a measurement of the Dutch insurance market, one brand appeared in 72% of answers but was actually recommended in only 8% of them. A smaller player appeared in 24% of answers and was recommended in 80% of those. Highly visible and rarely recommended is a sentiment problem. Rarely visible and often recommended is a visibility problem. Two different jobs, and you only see them if you measure them separately.
3. Report streaks, not percentages
A visibility percentage tells you how often you appeared in the answer, not whether that holds. If it drops from 38% to 30%, there are at least eight possible explanations. A competitor took your place. The page that cited you disappeared from the search results. Or the answer no longer names any brands at all, leaving nothing for anyone to win. Each of those causes calls for different work, while the percentage makes them all look the same.
A streak solves that: you run the same prompt ten times in a row and count how often something holds. A source cited in all ten measurements carries the answer. A source that appeared once was a coincidence. It also shows you whether the citation that named you has just dropped out, and which new source took its place. FolloEngine tracks those streaks, so you answer the question “why are we down” before it is asked.
4. Being cited is not owning the answer
Visibility is almost always measured as: what percentage of answers do I appear in? That is the wrong question. The better question is what percentage of the answer itself is yours. Still a percentage, but one that means something. You work it out by splitting the answer into separate passages, establishing which source each passage comes from, and counting how much text each source contributes in total.
That figure can be steered. Measure it per topic rather than per answer and set two questions against it: is there anything about it on your site, and is that being cited? If there is nothing, write it. If there is something but it is not being cited, rewrite it so a model can take it over, with a clear heading, a list and the answer in the first sentence. If it is already being cited, leave it alone.
5. Win Google Shopping and you win AI Shopping
ChatGPT takes its products from the organic Google Shopping results. As long as that is the case, your visibility in AI Shopping runs through your Shopping feed and your position in it. That changes once ChatGPT has enough product data for a feed of its own, with its own signals such as popularity score, return rate and review data. Until then, the shortest route to AI Shopping is the route you already know.
6. Behind every answer are the criteria the model chooses on
Ask for the best mattress for back pain, or for when your partner tosses and turns, and the model turns that into separate search queries you never typed: pressure relief, medium-firm, motion isolation. This is called query fan-out, and those queries can be read straight out of the response from ChatGPT and AI Mode.
The same goes for the comparison tables a model shows. Their column headings, for running shoes cushioning, stability, feel and price bracket, are the criteria it weighs products against. Pull them out and you know which attributes belong on your product page and in your feed, in the model’s own words.
7. Optional feed fields decide whether you get recommended
Google requires fifteen feed attributes. Alongside those are more than thirty optional ones, and those are exactly the fields an AI agent reads. Most feeds are filled in to around 60%. An A/B test showed that a product feed filled in to 99.9% delivers 32% more clicks, 27% more revenue, 40% more recommendations in ChatGPT Shopping and 50% more impressions in Google Shopping and AI Overviews. Required fields get you on the list. Optional fields get you recommended.
Your analytics shows only a fraction
What you see in your analytics is 1 to 5% of what actually happens. Everything below that is influence that gets no credit: an LLM answer, a video on social, a thread on Reddit, a position in the search results. That is the Dark Search Gap, and it grows as AI Mode becomes the default answer.
The seven tactics have one thing in common: they make part of that invisible share measurable, not by reporting more, but by measuring the follow-up actions that visibility causes. Want to know how visible your brand is? Get in touch.
