LLM Monitoring
Language models have no Search Console. Here is how you measure and monitor your visibility in ChatGPT, Copilot, Perplexity and Claude with a method that holds over time.
In classic search you do not have to invent the measurement. Search Console, rank tracking and analytics tell you where you stand and what is moving. In AI search none of that exists. ChatGPT, Copilot, Perplexity and Claude ship no dashboard showing how often your brand is mentioned or which of your pages get cited. If you want to know, you have to measure it yourself.
That is what LLM monitoring is: turning your visibility in language models into something you can see, follow and act on, rather than something you hear about when a customer happens to mention it.
The discipline belongs squarely in SEO. Same buyers, same questions, a surface that reports nothing back on its own.
What to measure
Visibility in language models has several layers, and each tells you something different.
- Brand mentions. Does your brand get named when the model answers the questions your customers ask? This is raw visibility, and it can be there without a single one of your pages being cited, because the model also knows you from its training data.
- Citations. When the model searches the web, which pages does it build the answer on? This measures whether your content is quotable, and it is the layer you can influence most directly.
- Share of voice. Who else gets named? Appearing in three out of ten answers means something very different depending on whether your closest competitor appears in nine or in two.
- How you are described. What does the model actually say about you? A model that describes you inaccurately or with two-year-old information is a problem no rank tracker catches, and it is often where the largest and fastest win sits.
- The effect. Referral traffic from AI surfaces shows up in your analytics, and your server logs show whether AI crawlers such as GPTBot, PerplexityBot and ClaudeBot are fetching your pages at all. A small amount of traffic with high intent can be worth more than a large amount without.
The method: a fixed prompt set, run systematically
Reliable LLM monitoring rests on the same principle as all other measurement, which is comparability over time. The method looks like this:
- Build a prompt set that mirrors your customer journey. Start from the questions real people ask an assistant, from the broad ("how do you choose a supplier for X") through the comparative ("who are the best in the Nordics at X") to the specific ("what does X typically cost"). If you have done a classic keyword analysis, it is a good starting point, but prompts have to be written as whole questions.
- Run the set across the models. ChatGPT, Copilot, Perplexity, Claude and ideally Google's AI Overviews. The surfaces pick sources differently, so a complete picture needs all of them.
- Record the same things every time. Were you mentioned, were you cited, who else appeared, and how were you described.
- Repeat on a fixed rhythm. Single measurements are snapshots. It is the curve that tells you whether the work is landing.
One honest caveat: language models are not deterministic. The same prompt can produce different answers from one run to the next, and the models are updated continuously. That does not make measurement pointless. It means you measure patterns rather than individual answers. Run each prompt several times, work with proportions rather than absolute counts, and resist the urge to read a single response as a verdict.
From measurement to action
Monitoring is only half of it. The value appears when the results drive your priorities. In practice the measurements point at three kinds of work.
The first is pages that need to become quotable, because competitors are winning citations on questions you should own. Usually the content is there and the structure is not: no clear answer near the top, no numbers, no headings a model can lift a passage from.
The second is gaps, where none of your pages can carry an answer at all. These are often the most valuable finding, because they are the questions your buyers are asking that you have simply never written about.
The third is misleading descriptions, which require you to make your own story clearer and more consistent across the web. That work is rarely on your own site alone. It runs through the third-party sources the models read, which is why digital PR belongs in the response and not only in the link plan.
At Bonzer this discipline does not run by hand. Our own work sits on Morrison, our AI content ops platform, which reads a client's entire website, learns the brand from their own documents and connects to performance data, so monitoring, analysis and execution stay in one place instead of living in separate spreadsheets. It is the tool behind the delivery rather than a product you buy. The point is that systematic beats occasional: a prompt set run every month tells you something, and asking ChatGPT a question now and then tells you nothing.
Getting started
Start small, but start systematically: ten to twenty prompts, four surfaces, one monthly run, one spreadsheet. That is enough to see the patterns and choose your first actions, and considerably better than waiting for the perfect setup.
Measuring visibility in classic search has its own guide in visibility tracking, and the two belong together. Same customer journey, more surfaces. Reporting the two side by side is what makes the picture usable for the people who approve your budget, which is covered under search analytics and reporting.
If you want a combined picture of where you stand today, across classic search and AI search, it starts with a free SEO analysis. It is built manually and covers both, because measuring one without the other has stopped making sense.
