You can measure Google rankings precisely and AI answers barely at all. Here is a method that works — the prompts to track, how to run them without fooling yourself, what to record, and how to turn the findings into changes that get you cited.
Search Console tells you exactly where you rank, for what, and how that changed. For AI answers there is no equivalent, and the absence causes a specific failure: businesses either ignore the channel because they cannot measure it, or they believe whatever a vendor tells them because they have no way to check. Neither is necessary. A disciplined manual method gives you a real baseline, and this is how to run it.
Start with the questions, not the keywords
Keyword lists do not transfer. Nobody types 'woocommerce developer india' into ChatGPT — they type a sentence describing their situation and ask what to do. So build a prompt list instead, and build it from how customers actually talk. Mine your sales calls, your contact form submissions and your support inbox for the questions people genuinely ask before buying. Twenty is plenty to start. Cover three types: direct recommendation requests, comparison questions, and problem-shaped questions where your service is one possible answer. The third type is the one most businesses forget and the one where citations are easiest to win.
Run them in a way that does not lie to you
This is where most informal checking goes wrong. Use a fresh session with no memory or personalisation — a logged-out or temporary chat — because an engine that has watched you research your own company for a month will flatter you. Run the same prompt across each engine you care about, which realistically means ChatGPT, Gemini, Perplexity and Google's AI Overviews. Do not reword the prompt until you like the answer; the first response is the data. And run from a location consistent with your market, because answers to local questions vary by geography in ways that will mislead you otherwise.
Record four things per prompt
Keep it simple enough that you will actually do it monthly. For each prompt and each engine, log: whether your brand was mentioned at all, its position relative to competitors, the sentiment and accuracy of what was said, and which sources were cited. That last column is the most actionable thing in the whole exercise — it tells you precisely which pages and publications the engine trusts for that question, which is your target list for the next quarter. A spreadsheet is entirely adequate. The value is in the consistency, not the tooling.
Read the citations, not just the mentions
When your competitor is named and you are not, the interesting question is not 'why them' but 'where did that come from'. Open the cited sources. You will usually find a comparison article, a directory listing, a forum thread or a documentation page — and almost always something you could plausibly appear in too. Over a few months this builds a map of the handful of sources that disproportionately shape answers in your category. Appearing in those is the single most direct route to being cited, and it is a far more targeted exercise than general link building.
Correct what is wrong about you
A surprisingly common finding is that engines describe your business inaccurately — wrong services, wrong location, outdated pricing, sometimes conflation with a similarly named company. That is fixable, and it is usually caused by inconsistency in your own public footprint. Make your name, description, services and contact details identical across your site, your Google Business Profile, your social accounts and every directory you appear in, and make sure your structured data says the same thing in machine-readable form. Entity confusion is the most repairable problem in this whole area.
Watch the second-order signal in your own analytics
Two numbers in data you already own corroborate the manual tracking. First, referral traffic from AI user-agents, which is small but growing and tends to convert well because the visitor arrived with a recommendation rather than a search result. Second, branded search volume. The most common behaviour is that someone sees you named in an AI answer and then searches your name — so branded search rising without a campaign behind it is usually the earliest visible evidence that the work is landing. Neither number is sufficient alone; together with the prompt log they make a defensible picture.
Turn the log into a quarterly plan
After two or three months the spreadsheet starts telling you what to do. Prompts where you are never mentioned and the cited sources are all comparison articles mean you need to be in those comparisons. Prompts where you appear but are described wrongly mean an entity and schema problem. Prompts where you are absent and the citations are all competitors' own pages mean your equivalent page is not answer-shaped — the fixes for that are in our GEO guide. Pick the three with the clearest route and work them, rather than trying to move everything at once.
Measuring AI visibility is not difficult, it is just unfamiliar and slightly tedious — twenty prompts across four engines once a month, logged consistently. That is enough to stop guessing, and it is far more trustworthy than a dashboard whose methodology you cannot inspect. If you would rather have it run and interpreted for you, that baseline report is the first deliverable of our GEO service, and it pairs naturally with the SEO work that gets you retrieved in the first place. Tell us what you sell and we will run your first twenty prompts as part of the quote.
Code Craft
Product at Code Craft — the team behind our published work and products and the 39-plugin product suite.