For twenty years, search success had a simple shape: a position in a list, and the clicks that came from it. Generative engine optimisation (GEO) breaks that shape. When someone asks ChatGPT or Perplexity which tool, clinic or agency to choose, the answer is the result. There may be no click at all, and nothing in your analytics to say you were in the running.
So the first question in any GEO programme isn't "what should we do?" It's "how will we know it's working?" Here are the measures we use, how to collect them, and the ones that look useful but mislead.
Why the old numbers fall short
Rankings still matter, because many AI tools search the web before they answer and lean on what ranks. But a page-one ranking doesn't guarantee a mention, and a mention often comes from a page you don't own: a forum thread, a review site, a trade article.
Traffic falls short too. A buyer can read an AI answer that names you, then type your name into Google a week later. That visit looks like branded search, not AI. If you only measure clicks, the work that earned the recommendation stays invisible.
The six measures that matter
1. Share of answer
The headline number. Take a fixed set of questions your buyers actually ask, put them to several AI models, and count how often you're named. If you're named in 9 of 40 answers this month and 17 of 40 in three months, that's progress you can see.
Keep the questions, the models and the settings fixed, or the comparison means nothing. Our baseline guide covers how to build the question set.
2. Position and framing in the answer
Being named fifth in a list of eight is not the same as being the first recommendation. Note where you appear and how you're described. "A cheaper option" and "the one most people recommend" are both mentions, with very different value.
3. Accuracy
Models repeat old facts: a discontinued product, an outdated price, a complaint you fixed two years ago. Track how many answers contain something wrong about you. It's one of the fastest things to improve, and one of the most damaging to leave alone.
4. Sources cited
Most AI tools show or link their sources. Record which sources appear in answers about your category, and which of them mention you. This tells you where the answer is actually being built, and which places to work on next. It's usually a short list: a few forums, a couple of review platforms, a reference page, one or two publications.
5. Coverage of the sources that matter
Once you know the sources, measure your presence on them. How many of the threads models quote mention you? What's your rating and review volume on the platforms they cite? Is your reference page accurate? This is the leading indicator: coverage usually moves before share of answer does.
6. Business signals
Finally, connect it to things the business cares about:
- Referral visits from AI tools. Some assistants pass a referrer, so visits from chatgpt.com, perplexity.ai and similar domains appear in your analytics. They undercount, but the trend is useful.
- Branded search. People who meet you in an AI answer often search your name next.
- What leads tell you. Add "ChatGPT or another AI tool" as an option to "how did you hear about us?" on your forms and in sales calls. It's the most direct measure there is.
How to collect them
A spreadsheet is enough to start. One row per question and model, with columns for: named or not, position, framing, anything inaccurate, and the sources cited. Run the same questions on the same day each month, with the same settings (signed in or not, web search on or off), and keep the raw answers.
Software can automate the asking and charting at volume. If you're weighing that up, we compared the main options in tracking tool or audit?
The measures to ignore
- A single "AI visibility score". Useful as a trend inside one tool, meaningless across tools. Ask what's inside it before you report it.
- One-off screenshots. A model's answer to one question on one day proves very little. Answers vary between runs.
- Daily movement. Week to week, answers wobble. Look at monthly trends against a fixed baseline.
- Mentions without context. Being named in a warning ("avoid X, the support is slow") is a mention too.
Reporting it
Keep the monthly report to one page: share of answer against your two closest competitors, the sources that moved, the inaccuracies fixed, and what's planned next. If you need a benchmark against competitors, set it up once and repeat it.
What this doesn't mean
None of this guarantees a recommendation. Nobody controls what a model says, and anyone who promises a place in ChatGPT's answer is guessing. What you can control is the work on the sources models trust, and an honest measure of whether it's changing the answer.
If you'd rather start with someone else doing the first measurement, that's what our visibility audit is: five models, your buyers' questions, you and two competitors, every source traced, and a ranked plan.