Most teams now assume that customers will meet them through a search box. For years that meant Google. In 2024, it increasingly means a question typed into ChatGPT, Claude, Perplexity or Gemini, then a short paragraph that feels confident enough to skip clicking through.
That shift matters in a way that is oddly hard to see from analytics dashboards. Traditional search sends you referrers and query strings. Large language models often do not. A buyer can ask an AI tool about “best suppliers for X”, read a summarised answer, make a short-list and never touch your site. If your company is missing or misrepresented at that first step, you rarely get to find out.
You cannot directly log what an AI model has said about you, but you can probe it in a structured way. Treat it like user research: set up a simple, repeatable exercise to see whether the models mention you at all, how they spell your name, what roles they assign you and which sources they appear to lean on.
This is less about catching the tools out and more about understanding what public signals they are picking up. The output will not be perfect or fair, but it is a useful proxy for your visibility in the parts of the web these systems regard as trustworthy.
Why “LLM visibility” is not the same as SEO
It is tempting to fold this into search strategy and move on. After all, these models are trained on a lot of web content, so surely better SEO means better AI exposure.
The overlap is real but incomplete. Traditional SEO is about documents, links and query matching. Large language models are trained to reproduce patterns of language, then often paired with newer “retrieval” systems that pull in a handful of sources at answer time.
Several consequences follow.
First, models are much more opinionated. They try to answer from a single voice. Instead of ten blue links, you get one narrative, perhaps with a handful of citations. If your company falls just outside that small window, you are invisible to the user, even if you rank well in search.
Second, the systems are biased toward entities and concepts that appear in certain types of sources, such as high authority media, knowledge bases and structured data. If your presence is mostly in unstructured or local content, the model might not “remember” you clearly enough to name you.
Third, models have refresh cycles. A web page you updated last week can appear in organic search quickly, while a base model that was last trained on data from early 2023 might never have seen it. Even when tools say they “browse”, they often mix older internal knowledge with a thin layer of recent references.
So if you want to know whether an LLM knows your business, you cannot infer it entirely from search rankings. You need to ask the models directly, but in a way that lets you compare.
Designing a repeatable test
Think of this as a small research protocol. You want prompts that reflect how a sensible user would look for you, without feeding the model your name, and a way of recording answers across tools and over time.
Pick three types of prompt.
1. Category prompt without location
Something like: “List well known companies that provide [your product or service] for [your audience].”
This checks whether you are part of the generic mental map of the category. Most models will return a mix of global examples and a few mid-sized names. If you are nowhere in that list, it suggests the model does not treat you as a default reference.
2. Category plus location
For instance: “Which companies provide [your product or service] for [your audience] in [your city or region]?”
This is closer to how users search in practice. Here you are testing whether the model has tied your entity to the right geography, and whether local sources are strong enough to surface you.
3. Problem and use case prompts
Something like: “I run [type of organisation] and need help with [problem A and problem B]. Which companies specialise in this?”
Many models are tuned to respond to needs rather than industry labels. This type of prompt checks whether the way you describe your work in public content matches the language your buyers use.
Run each of these, unedited, in the four tools you care about: ChatGPT, Claude, Perplexity and Gemini. Use the same browser and account each time you repeat the test, since personalisation can creep in. Do not correct the model or argue with it; simply copy each answer into a document with the date, prompt, tool and output.
You will quickly see patterns. Some systems will decline to name companies at all and will instead offer generic advice. Others will list specific firms and sometimes provide short summaries. Treat refusals as data too: if a model will not name companies for that query, then you know that particular route will not surface you for users either.
What to look for in the answers
Once you have a set of responses, read them as you would a competitor report.
Start with **presence**. Are you named at all? If yes, in which tools and for which prompts? Is your name spelled correctly? Are you assigned to the right country, sector and audience? If a model invents details, note those carefully, because users may repeat them back to you.
Next, check **category fit**. Models like to place entities in clear boxes. If you are repeatedly described in a way that feels slightly off, that is a hint that your own messaging and public profiles are fuzzy. For example, a Toronto-based marketing agency and software platform that serves medical aesthetics and healthcare practices might see itself mainly as an operating system vendor, while the models keep describing it as a general marketing agency. That gap matters. It tells you which aspects of your offer are reinforced most strongly by your visible content.
Third, scan the **company lists** around you. Models often cluster similar entities together. If you sit in a list with firms that address a quite different audience or price point, that can explain odd leads you receive. It can also highlight publishers and directories that carry more weight than you expected.
Finally, note the **sources the tools cite**, especially in Perplexity and Gemini, which tend to show their working more clearly. Are they pulling from your own site, from review platforms, from trade press, from your GitHub, from government or regulatory registers? This is your window into the parts of the web that count most for these systems when they answer questions about your space.
For a worked example of how to structure this exercise and log results over time, it is worth reading how to test whether an LLM knows your business, which walks through one marketing agency’s approach to tracking its own visibility and that of its clients.
What the results imply about your content and data
If none of the tools mention you, do not treat that as a verdict on your actual market position. It is an indicator that public, machine-readable signals about your company are thin or inconsistent.
Several levers are usually in play.
– **Structured data**: Clear company profiles on your own site and on key directories help models anchor basic facts. Schema markup, consistent NAP (name, address, phone) data and complete profiles on industry or professional registers can all improve the odds that you are recognised as a distinct entity.
– **Third-party coverage**: Trade press articles, conference listings, podcasts and association pages tend to be overrepresented in the sources these models trust. If you are active in a niche but almost all mentions of you live on your own domain, the models may treat you as less “real” than peers with thinner businesses but broader citation.
– **Clarity of positioning**: If your website tries to address many audiences at once, the model may pick one at random or blend them. Tight, specific language about who you serve and what you do makes it easier for the system to slot you into the right mental bucket.
– **Recency and stability**: Frequent rebrands, name tweaks or domain changes can fragment the trail. Since model training lags the live web, an identity that has shifted repeatedly is harder to pin to a single entity.
You cannot force an LLM to say you are prominent, but you can ensure that when it goes looking for you, it finds clean, consistent signals.
Making this part of your regular checks
Once you have run the test once, treat it as a periodic check rather than a one-off stunt. Quarterly is often enough. Repeat the same prompts in the same tools, save the outputs and compare.
You are not chasing perfection in the answers. You are watching for movement. If a model that previously misnamed you starts to get your description right, it suggests that your recent content and data changes are beginning to register. If a tool that used to list you stops doing so after a site restructure, that is a flag to investigate.
Over time, you will build a quiet log of how general purpose AI systems talk about your part of the world. That will not tell you everything about your buyers, but it will show you what many of them are likely to see before they even think about typing your domain into a browser.
