AI Search Rank Tracking: What You Can Measure

You cannot track a rank position in ChatGPT or Perplexity the way you track a keyword in Google, because an AI answer is generated for each request and there is no fixed list of results to hold a position in. What you can track is how often your brand is mentioned and your pages are cited when the same questions are asked repeatedly, and which competitors and third-party pages show up instead.

That is a rate, not a rank, and it comes with noise. This guide shows how to measure it by hand with a fixed prompt set, repeated runs and a simple log, how to tell signal from variance, and where paid tools fit. Everything here that is a number from outside this site names its source. The worked example later on is invented and labelled as such.

Tracking is the feedback loop for the optimization work covered in the generative engine optimization guide, the ChatGPT guide and the Perplexity guide. Without it you cannot tell which changes helped.

Two sources describe the limits from the practitioner side. WebFX's guide to tracking AI search rankings (fetched 9 October 2026) says there is no single dashboard that shows where a brand appears across AI search, that results vary by platform even for the same prompt, and that location and search history personalize what people see. It also notes that Google Search Console has no AI-specific report and that analytics only capture clicked referrals, since an answer can summarize your content without linking to it.

On the referral side, OpenAI's Publishers and Developers FAQ says ChatGPT search includes the UTM parameter utm_source=chatgpt.com in referral URLs, so analytics can separate that traffic, and that publishers who allow OAI-SearchBot can measure it in tools such as Google Analytics. I did not fetch an equivalent first-party statement for Perplexity, so check your own analytics for how its clicks arrive instead of assuming.

Reading the current top results for this query, the common pattern is a definition of AI visibility, a list of metrics (mentions, citations, share of voice, sentiment) and a list of tools, many of them written by tool vendors. The WebFX guide covers methods well but, in its own text, gives no sampling protocol, no number of runs per prompt and no logging template. A vendor post from Topify (fetched the same day) recommends 20 to 30 prompts and a manual spreadsheet audit, and itself admits manual audits are point-in-time snapshots. Neither explains how to read variance, which is the part that decides whether your numbers mean anything.

That gap is what the rest of this page fills: a protocol you can run this week, the columns to log, and the arithmetic for deciding whether a change is real. I have not tested these tools' accuracy, and I make no claim that any of them measures better than a careful manual process.

What this example teaches. The headline overall rate looks healthy, but it is propped up by the two branded prompts, where the brand is named in every run. The non-branded slice, the one that reflects discovery by people who do not already know you, is much lower and its range is wide. Reporting only the blended number would hide that.

How to read variance. P01 shows 3 of 5 mentions. With only five runs, the plausible true rate for that single prompt spans most of the scale, so a move from 3/5 to 4/5 next week is not evidence of anything. Three habits help: judge changes on the aggregated slices rather than on single prompts, treat a change as real only when the new range no longer overlaps the old one or when it persists over several rounds, and add runs where a decision depends on a close call. More runs narrow the interval, but slowly, and they do not remove the bias of choosing your own prompts.

What to do with it. In this invented case you would look at P03 and P06, where the brand never appears, read which third-party pages the engine cited instead (the other_sources column), and decide whether to improve your own page for that question or earn a mention on the pages that are being cited. Then re-run the same prompts after the change.

A different job: Website Verdict is an audit tool. Its homepage describes a scan of SEO foundations, AI discoverability and agent readiness based on site signals such as AI-crawler access, llms.txt and structured data. It does not describe monitoring real AI answers, so use it to check that your pages can be found and read, not to measure mentions. It does list an "AI visibility tracker" as a credit-metered Pro feature without explaining it, so I cannot say what that measures.

For classic tools: WebFX's guide says Semrush Position Tracking and Ahrefs Site Explorer can filter for keywords that trigger a Google AI Overview, and that these cover AI Overviews only, not chat platforms. I did not open either vendor's page, so treat that as second-hand. WebFX also reports that AI Overviews appeared in 25.8% of U.S. searches in its study of 2.37 million keywords; I have not seen the methodology, so use it as context rather than a benchmark.

For a broader comparison of optimization tooling, see the AI search optimization tools guide.

Build the prompt set and the sheet first, run one full round, and only then decide whether a paid tracker is worth it. EchoVerse does not sell a tracking product, so there is no product page to point you to; browse all EchoVerse guides or the EchoVerse homepage for the projects behind them. If your site has not been checked for basic AI discoverability, an audit such as Website Verdict is a sensible first step before you measure anything.