A buyer used to open Google, scan ten links and pick one. Increasingly they ask an assistant instead, and the assistant answers with three names. If yours is not one of them, you were not outranked. You were left out of the sentence, and there is no position report that will tell you it happened.
Why rank tracking misses this
Rank tracking assumes a list. A query goes in, ten blue links come back, and your page sits at some number in that list. Every tool built in the last fifteen years reads that number, and the whole discipline is downstream of it.
An AI answer has no list. The model reads a handful of sources, writes a paragraph, and names two or three products inside it. There is no third result to be, so there is nothing to look up. The only question that survives is a different one: when someone asks the question your buyers actually ask, are you in the answer?
You are not trying to rank third. You are trying to be named at all — and being named twice out of ten runs is a real, measurable number.
The ten-minute check
Before you buy anything, run this by hand. It takes about ten minutes and it will tell you more than a demo will.
- Write five questions a buyer asks.Not your brand name — the problem. “Best tool for tracking brand mentions in AI answers” is a question. “RankThroughAI review” is not; you already win that one.
- Ask each one in a fresh chat. ChatGPT, Gemini, Perplexity and Claude, web search on, memory and personalisation off. Your own history will quietly hand you a flattering answer otherwise.
- Record three things per run: were you mentioned, were you linked, and who else was named.
- Run each question twice. The same prompt returns a different answer on a second pass. One run is an anecdote; two is the beginning of a rate.
Reading the result
Forty runs gives you a grid. Fill it in and three numbers fall out of it, and those three are the whole scoreboard:
| Metric | What it counts | What a weak score looks like |
|---|---|---|
| Mention rate | Runs where your brand is named | Under 20% on your own category terms |
| Citation rate | Runs where a page of yours is linked | Named often, linked almost never |
| Share of voice | Your mentions against the competitors named beside you | One competitor in every answer, you in two |
The gap between mention rate and citation rate is the interesting one. Being named without being linked means the model knows the category consensus about you and is not reading your site to say it. That is a content problem, and it is fixable.
What actually moves it
Nothing exotic. The pages that get pulled into answers tend to share four properties, and all four are things you control:
- The answer is in the first two lines. A retriever takes a chunk, not a page. If the definition arrives in paragraph nine, the chunk that gets quoted is your introduction.
- Terms are defined before they are used. Models quote sentences that stand on their own. A sentence that depends on the previous three does not travel.
- Comparisons are explicit. If you never say who you are an alternative to, you are not in the set the model is choosing from.
- Numbers carry a source and a date. An unsourced claim is exactly the kind of sentence a model declines to repeat.
None of that is new advice. What is new is that you can now measure whether it worked, instead of arguing about it.
Doing it on a schedule
The manual check is the right way to start and the wrong way to continue. Forty prompts a month across four engines is an afternoon you will stop spending by March, and a rate you only sample once is not a trend.
That loop is what RankThroughAI’s AI visibility tracker runs for you: the same prompts against the same engines on a cadence you set, with the mention, the citation and the competitor set recorded each time — and the content editor pointed at the pages that were missing when the answer was written.
Run the manual version first. Ten minutes, five questions, two passes. Whatever the grid says, you will know something about your market that your rank tracker has never once mentioned.