How to measure a brand's presence in ChatGPT, Gemini and Perplexity answers and correct inaccurate information.
How often and how accurately large language models mention a brand when answering relevant questions — "which agency should I choose for SEO in Almaty", "how much does a site audit cost". It is measured on a fixed set of queries and covers three things: whether the brand is mentioned, whether its services, prices and contacts are described correctly, and which sources the model cites.
Core metrics: the share of answers mentioning the brand across the control queries, the share with correct facts, position in the list of recommendations, the share linking to your site, and share of voice against 3–5 competitors. Each check is logged with date, system, language and region and repeated monthly on the same sample, so you see a trend rather than chance.
Generation is probabilistic: the same question can produce different answers. Model versions, the search indexes sources come from and the sources themselves also change. Wording, language, region and conversation history all matter. That is why conclusions are drawn from samples of many queries and repeats, not from one screenshot.
Yes. The control queries are run in Gemini, and Google AI Overviews and AI Mode are tracked separately because they use Google's search index. Checks are done in a clean profile without personalisation, with fixed region and language, documenting date and wording. Automated checks are acceptable only within the service's terms of use.
Yes, and it is one of the most transparent systems: Perplexity almost always shows its sources. For each control question, record whether the brand is mentioned, which sites are cited and which facts are used. If it cites aggregators and competitors' reviews rather than your site, that is a direct hint about which materials are missing and where you need a presence.
Usually not: models use the same pages as search engines. It is more effective to improve existing pages — add direct answers, concrete facts, comparison tables, authorship — and remove contradictions. Separate content for bots that differs from what people see is cloaking: it breaks search engine rules and undermines trust.
Your brand's share of all mentions of a chosen set of competitors in a given sample of answers. For example, if competitors are mentioned 160 times in 100 answers and you 40 times, your share of voice is 20%. The metric is relative: it depends on the query set and competitor list, so only compare it on the same sample over time.
Typical causes: similar names, few independent sources about the company, outdated or contradictory data — an old address in directories, different prices on different pages, several versions of the service description. The model stitches together fragments about different entities. The fix is one set of facts on the site, Organization markup with sameAs links, and up-to-date external profiles.
First find the source of the error: Perplexity and ChatGPT Search show which pages an answer cites. Then fix the facts on your site, in directories, maps, social networks and profiles, request corrections from third-party sites and get pages re-indexed. Answers change with a delay: search-based systems react within weeks, while a model's built-in knowledge updates only with new versions.
No. Neither an agency, nor markup, nor an llms.txt file controls what a model cites. What can be guaranteed is the work: an accessibility audit for AI bots, fixing contradictions, preparing materials and regular measurement with a transparent method. If someone promises guaranteed mentions in ChatGPT, ask exactly how that will be measured.
Nothing found.
