Rank tracking AI search, and why the metaphor breaks.

Rank tracking made sense because there was a ranked list and a position in it. A generated answer has neither. You are either mentioned or you are not, sometimes, depending on how the question was phrased and when it was asked.

What replaces position

Citation rate. Across repeated asks of the same prompt, the proportion in which you are named. This is the closest honest analogue to a ranking, and it is a probability rather than a position.

Share of voice. Of all businesses named across your prompt set, what proportion of mentions are yours. More robust than your own rate, because it controls for category-wide shifts.

Mention order. Where in the answer you appear. Weak evidence and worth recording, but not worth optimizing for directly.

Worked example: reading a citation rate

Suppose you ask ten buyer questions five times each in ChatGPT, fifty asks in total, and you are named in eighteen. Your citation rate is 36%. A competitor is named in twenty-six, and six other businesses share the remaining mentions. Share of voice is your mentions divided by all mentions across those answers.

Next month you are named in twenty-one of fifty. Is that progress? With samples this size, a move from 18 to 21 is within the range you would see by chance. If it holds or grows over the following two months while the competitor's count falls, it is probably real. That is why the trend and the sample size need to travel together in every report.

Why the distinction matters in reporting

Presenting a citation rate as a rank invites comparison with search rankings, and those comparisons mislead. A citation rate of 60% is not equivalent to position 3, and a fall from 60% to 50% may be variance rather than decline.

Reports that use ranking language create an expectation of precision the underlying data cannot support. When someone eventually examines the methodology, the credibility cost lands on whoever wrote the report.

What to put in front of a client or a board

Citation rate with the sample size stated. Share of voice against named competitors. A short list of sources gained or lost. Any accuracy problems found. And an explicit note that responses vary between sessions, so small movements are not significant.

That last line costs nothing and prevents a quarter of pointless conversations about noise.

Where traditional rankings still matter

None of this makes ordinary rank tracking obsolete. Google's AI Overviews and AI Mode draw on pages that rank in Google's index, and Copilot draws on Bing. Ranking well for the underlying queries remains one of the more reliable predictors of being eligible for citation in those answers.

The change is in how the rankings are used. They become an input, one of several reasons you might or might not be cited, rather than the outcome being reported. A page can rank third and not be cited; the rank tells you eligibility, the citation rate tells you selection.

Tracking clicks when rankings hold

A common pattern since AI summaries spread is stable rankings with falling clicks. Pew Research measured click-through to traditional results at around 8% when an AI summary appeared, against 15% without one. In Search Console that shows up as steady positions and impressions with declining click-through rate on informational queries.

Track it deliberately. Segment Search Console data into queries that trigger AI Overviews and those that do not, if your tools can identify them, and report click-through separately. It explains traffic changes that rank tracking alone cannot.

Frequently asked questions

Can you track rankings in AI search?

Not in the traditional sense, because a generated answer has no ranked list. The honest analogues are citation rate across repeated asks, share of voice against competitors, and mention order, all of which are probabilities rather than positions.

What is a good AI citation rate?

There is no benchmark, because it depends entirely on your prompt set and category. What matters is the trend on an unchanged prompt set and your share relative to competitors on the same questions.

Why do AI rank tracking tools show different positions?

Because they are converting a probabilistic outcome into a positional metric using their own methodology. Different prompt sets, sampling frequencies and mention definitions produce different numbers from the same underlying behavior.

How many asks do I need for a reliable citation rate?

More than most people use. Five asks per prompt across twenty prompts gives a hundred observations, enough to see large changes. Small month-to-month differences still need confirming over several periods.

Do AI Overviews cite the top-ranking pages?

Often, but not always. They frequently cite pages from the first page of results, and sometimes pages further down that answer the specific question more directly.