Your company appeared in a ChatGPT recommendation on Monday. By Wednesday, the answer looked different. Did something go wrong?
Maybe. But one changed answer won't tell you. A useful tracking routine asks the same relevant questions over time, records the conditions, and keeps the answers available to inspect. That lets you spot patterns without turning every surprise into a website rewrite.
Here's how to build a small routine your team can actually maintain.
What does “ChatGPT rank tracking” measure?
The phrase borrows from search rankings, but an AI answer isn't a fixed list of blue links. An assistant might recommend a shortlist, mention a product in passing, or link to a guide without recommending its publisher.
Decide which of those outcomes you care about before collecting data. Otherwise, a dashboard can look encouraging while answering the wrong question.
- Brand presence: did the answer name your company?
- Recommendation: did it suggest your product for the stated need?
- Citation: did it include a linked source on your website?
- Description accuracy: did it get your features, audience, and limits right?
Save the underlying response. If a recommendation is qualified with “only suitable for large enterprises,” the qualification matters as much as the mention.
Start with questions a buyer might reasonably ask
Use sales and support conversations to identify the decisions people need help making. Then turn them into neutral questions. Don't add your brand to every prompt: that would measure a different task from discovering vendors.
For a hypothetical reporting tool, a useful set might include “What tools help a small agency report results to clients?” and “How can an agency combine website and campaign reporting?” Keep questions about your own pricing or features in a separate group.
These are research questions you have chosen. They aren't a recording of private ChatGPT conversations or a count of actual buyer demand. Check their relevance against customer conversations, search data, and the audience you want to reach.
Start with a manageable set covering a few important buying needs. Keep a record of why each question belongs. A large collection of near-duplicates can give one topic more weight than you intended.
Record the setup along with the answer
OpenAI explains that ChatGPT can search the web and provide links to sources. Record whether you're checking a search-enabled experience or another setup. Don't mix them and call the result a single ranking.
For each observation, keep the exact prompt, date, available model or mode information, language, location assumptions, and whether a fresh conversation was used. Note personalization or conversation history when relevant. If a setting isn't visible, record it as unknown instead of guessing.
An automated API run and a person's ChatGPT session can differ. Describe the method you actually use. Don't label an API sample as what every ChatGPT user sees, and don't invent a consumer-app temperature setting to explain a result.
Consistency helps you compare observations. It doesn't make the sample universal.
Read a sample before trusting the total
Suppose a fictional company appears in six of twenty completed answers. Its answer-presence rate is 30% for that sample. If only two of those answers recommend it, the recommendation rate is 10%. Both calculations are useful; they measure different things.
Keep failed runs separate from completed answers where the brand was absent. A tool outage shouldn't quietly become a visibility loss. Record the numerator and denominator beside each percentage.
Review examples of wins, losses, and ambiguous cases. Check product-name variants and whether the answer is talking about your company or another business with a similar name. Human review is especially useful when a sentence mentions several competing products.
When using WildSEO, inspect the metric definition in the report. Its share-of-voice calculation counts detected brand-mention spans; that differs from the answer-presence example above. Our AI share-of-voice guide walks through the distinction.
Use a weekly review to choose the next action
Look at topic groups before the overall total. Did your recommendations change for one use case, one provider, or the whole sample? Read the answers behind the change, then check the sources they link to.
A wrong feature claim may call for clearer documentation. A question you don't answer may justify a useful page update. A one-day citation drop may simply need more observation. Ahrefs' citation-stability research is a useful reminder that linked sources can turn over within a short tracking window.
Finally, connect the monitoring to business outcomes. Check relevant visits, signups, and sales feedback alongside visibility. Keep a short change log so you know what you published and when. If the numbers move afterward, investigate; don't assume you've proved the cause.
A good report should leave your team with a decision: fix an inaccurate page, investigate a repeated gap, or keep watching. That's more useful than celebrating a number nobody can explain.
