You can do this by hand in an afternoon. The method matters more than the tool, so here it is in full — including the parts that make most people's first attempt misleading.
Ask each assistant a dozen buyer-intent questions about your category, record whether your brand was named and whether your site was cited, and repeat it enough times that one odd answer does not move the result. Then look at who was named instead and which sources the answers were built from, because that is where the fix lives.
The single most common mistake is asking "what do you think of [my brand]?". That measures whether the model has heard of you, which is nearly useless. Buyers do not type your name; that is the entire problem. Ask category questions instead:
Get your category wording right. "Project management software" and "project management software for software teams" return substantially different brands.
ChatGPT, Perplexity, Gemini and Claude build answers from different source mixes and will not agree. In our test on Linear, ChatGPT named it in 92% of answers and Perplexity in 42% — the same brand, the same questions, on the same day. A single-assistant check will tell you a story that is true of one surface and false of the rest.
Use a fresh chat each time, and turn off personalisation or memory if you can. An assistant that already knows you work at the company will name it, and you will fool yourself.
Mention rate is the share of answers naming your brand. Citation rate is the share that actually link to your site as a source. They come apart badly. Ghost is named in 91.7% of answers about blogging platforms while ghost.org is cited in 12.5% of them — the assistants know the brand well and are reading other people's pages to describe it. Those two gaps need completely different work, and any score that blends them hides which problem you have.
Count how many separate answers named each rival, not how many times a name appeared. One answer that says "ClickUp" eight times is one vote. Require a name to show up in at least two different answers before you treat it as a real competitor in the assistants' eyes, and most of the noise disappears.
This is the part people skip and it is the most actionable. Every answer with search enabled is assembled from pages. Ours keep returning the same names: G2, Reddit, TechRepublic, Cloudwards, PCMag, Forbes, category round-up blogs. If those pages do not mention you, or list you badly, no amount of rewriting your own homepage will fix the answer.
Assistants are non-deterministic. Ask the same question three times and you can get three different orderings. One run is an anecdote. Twelve questions across two assistants is twenty-four data points, which is enough to act on. A weekly repeat is what turns it into a trend.
Three fixes cover most cases, in this order.
Fix your presence on the sources being cited. If G2 and a handful of round-up articles are building the answer, an incomplete profile or an absence from those round-ups costs you mentions that your own site cannot win back.
Publish the comparison that does not exist. Assistants quote pages that compare options directly, including the cases where the other option wins. A page honestly comparing you to the three names that keep appearing is the highest-yield thing most companies can write.
Answer the specific questions you are invisible for. Take the questions where you were not named, and write pages that answer them plainly, with specifics and numbers rather than marketing language. Assistants extract answers, not adjectives.