A tool category appeared over the last two years that didn't exist before: software that tells you whether AI assistants mention your company. It goes by several names (AI visibility, answer engine optimization platforms, LLM rank tracking), and the pitch is straightforward. Buyers ask ChatGPT, Claude, Gemini, and Perplexity for recommendations. You'd like to know if your name comes up.
That's a reasonable thing to want measured. It's worth understanding what these tools actually do before you price one, because the category is young, the methodology varies a lot, and some of the confidence in the marketing is not supported by what's technically knowable.
What they measure
Strip away the dashboards and nearly all of them do some combination of four things.
Prompt tracking. The core mechanic. The tool maintains a list of prompts relevant to your category, things like "best project management software for agencies" or "what is [category] and who are the main vendors," and runs them against several models on a schedule. It records which brands appear in each answer. Over time you get a mention rate per prompt per model.
This is the honest heart of the category. It's essentially automated sampling: asking the question many times and counting outcomes.
Citation monitoring. For models that show sources, such as Perplexity, AI search results, and assistants with browsing enabled, the tool records which URLs got cited. This is more actionable than mention tracking, because it tells you which specific pages are doing the work. Often it's not your homepage. It's a review site, a directory, a comparison article, or a trade publication that wrote about you.
Share-of-answer comparison. The same measurement applied to competitors, expressed as a percentage of tracked prompts where each brand appears. The AI-era analogue of share of voice, with all the same caveats about how much your prompt set determines the result.
Sentiment and accuracy checks. How the model characterizes you when it does mention you, and whether the description is factually right. The accuracy piece is genuinely useful, because models confidently state wrong pricing, wrong founding dates, and wrong feature sets, and you can't fix what you haven't noticed.
What to look for when evaluating one
Prompt set control. Can you write your own prompts, or are you stuck with generated ones? The prompt set is the measurement instrument. A tool that picks flattering prompts produces flattering numbers. You want full control and visibility into exactly what was asked.
Model and surface coverage. Which models, which versions, and whether browsing was enabled. An answer from a model with live search is a fundamentally different measurement from one generated without it. If the tool won't tell you which mode it used, the numbers aren't interpretable.
Sampling frequency and variance reporting. Model outputs are non-deterministic. Ask the same question five times and you can get five different lists. Any tool reporting a single run as a result is giving you noise. Look for repeated sampling and some expression of variance, not a clean number.
Raw response access. You should be able to read the actual answers, not just the aggregate. The verbatim text is where the useful findings are: the phrasing, the competitors named alongside you, the source it leaned on.
Citation-to-source mapping. The single most actionable output. If you learn that a specific roundup article drives most of your citations, you know where to focus.
Sane pricing. The underlying cost is API calls. Pricing that scales to enterprise levels should buy analysis and integrations, not just more of the same queries. For small companies, a spreadsheet and a monthly hour of manual checking gets you most of the signal, and our walkthrough for tracking brand mentions in AI covers that manual version.
Limitations worth being clear-eyed about
Non-determinism is not a bug you can tune around. The same prompt produces different answers across runs, sessions, and users. Anything reported without a sample size is close to meaningless, and small week-over-week movements are usually noise.
Personalization and context vary. A real user's answer is shaped by their history, location, account settings, and whatever they said earlier in the conversation. A tool queries from a clean state. Those are different conditions, and the gap is unmeasurable from outside.
The prompt set is the whole measurement. Twenty prompts is a narrow sample of how thousands of people phrase questions. You are measuring your visibility on the questions you chose to ask, which is a real but limited thing.
Models change without notice. A drop in your mention rate might reflect a model update rather than anything about your company. There's no changelog to check against, so treat discontinuities with suspicion.
Citations aren't the reasoning. A listed source is not proof the model relied on it, and an answer without citations still came from somewhere. The displayed sources are an interface feature, not an audit trail.
Mentions aren't traffic and aren't revenue. Being named in an answer may produce a visit, a remembered name, or nothing. Attribution from AI surfaces to pipeline remains genuinely unsolved, and most reported attempts are proxies.
What no tool can tell you yet
Worth stating plainly, because this is where the category's marketing gets ahead of reality.
No tool can tell you why a model named one company and not another. The selection process inside a large model isn't inspectable from the outside, and the people who build these systems describe it in probabilistic terms, not rules. Correlations do get observed. Brands with more third-party corroboration, consistent facts across sources, and clear entity definitions seem to fare better, but a correlation drawn from a sample of prompts is not a mechanism.
No tool can tell you what changing X will do. There's no keyword-difficulty equivalent, no reliable before-and-after isolation, because you can't hold the model constant while you run the experiment.
No tool measures the surface where it actually happened. It samples an approximation of the environment your buyer was in.
And no tool can promise placement. Any vendor claiming they can get you cited in ChatGPT is describing an outcome they don't control. The practices that appear to help are the same ones that made sense anyway: clear factual pages, consistent naming, and independent sources corroborating your claims. The checklist version is in our guide on how to get cited by ChatGPT, and the broader framework is in what generative engine optimization means.
A reasonable posture
Measure it, because not knowing how AI systems describe your company is a real blind spot, especially on the accuracy dimension, where errors are fixable once spotted. Start manually if you're small. Buy a tool when the manual checking exceeds an hour a month and you need competitor comparison.
Then hold the numbers loosely. Use them to find factual errors, discover which third-party sources carry weight in your category, and notice large directional shifts. Don't build a quarterly target around a metric whose underlying system can change next Tuesday without telling you.
Third-party coverage is one input you can actually control. [See what distribution costs](/pricing).