"LLM rank tracking" is a phrase people type into search engines, so it deserves a straight answer rather than a redirect. The answer is that a generative answer has no ranks in it, and any product that reports one has either invented it or renamed something else. What can be measured is real, useful and considerably more modest: rates of mention and citation, over a controlled prompt set, with sample sizes and intervals attached.
This page exists because the vocabulary decides the analysis. A team that thinks it is tracking positions will interpret noise as movement, will optimise for a number that does not exist, and will eventually have to explain to somebody why the number went backwards for no reason.
Why rank tracking works for search and not for answers
Conventional rank tracking works because a search engine result page is a deterministic ordered list. Ask twice under the same conditions and you get the same links in the same order, subject to personalisation and location. Position is a property of the page, so "position 4" is a fact about the world that two observers can independently confirm.
A generative answer has none of those properties.
- It is prose, not a list. Three companies named in a paragraph have not been assigned places. The order they appear in is a property of sentence construction.
- It is assembled at answer time. A retrieval step whose results vary feeds a generation step that samples. Two identical requests can produce different sets of named entities.
- The population is not fixed. Rank is relative to a set of competitors. In a generative answer the set is whatever the model chose to mention this time.
- There is no shared instrument. The model version, orchestration and retrieval stack behind a consumer chat product are undisclosed and change without notice.
Remove ordering, stability and a fixed population and nothing is left for the word "rank" to attach to.
What a single answer can actually tell you
One response supports a small number of binary facts, and only these:
- Was the target entity named in the answer text?
- Did the answer carry citations at all?
- Was the target domain among the cited sources?
- Which tracked competitors were named alongside it?
Each is a yes or no about one response. That is the entire raw material. Everything an honest tool reports is built by repeating those questions under recorded conditions and counting the yeses — or it is fabricated.
What SiteRank AI reports instead
Repeat a prompt enough times under identical conditions and the proportion of yeses becomes a rate with an uncertainty attached. SiteRank AI reports six of them, each with its numerator, its denominator, its sample size and a 95% Wilson score interval: mention rate, citation rate, domain citation rate, share of voice, citation share and prompt coverage. The definitions and arithmetic are on the AI visibility monitoring page.
Two of those look superficially like a ranking and are not.
Share of voice is the proportion of answers mentioning any tracked entity that mention yours. It is a proportion inside a prompt set you wrote and a competitor list you chose. Add a competitor and everyone's share drops without anything changing on any website. It is not market share, and it is not a league table.
Volatility is reported next to every rate: the proportion of samples that disagree with the majority answer for the same prompt under an identical condition key. Where a rank tracker would draw a confident line, this tells you how much any single observation deserves to be trusted. A 40% mention rate with high volatility is a coin flip, and the chart says so.
Why "ChatGPT rank #3" fails as a claim
Four independent problems, each sufficient on its own.
- There is no ordered list. Rank presupposes positions that the artefact does not contain.
- There is no stable population. The set of named entities changes between samples of the same prompt.
- It hides the sample size. "Rank #3" from one API call and from two hundred are presented identically, though only one of them is a measurement.
- It implies a mechanism nobody has documented. Presenting a position implies something is ranking, which implies criteria you could influence. No provider publishes such criteria.
There is a fifth problem specific to consumer products. A number obtained by driving a chat interface with browser automation cannot be reproduced, cannot be attributed to a known model version, is shaped by account memory and personalisation, and is generally contrary to the provider's terms of service. API-based AI visibility tracking vs ChatGPT works through the trade-off in full.
Terms that get used interchangeably and should not
| Term | What it means here |
|---|---|
| Crawled | A crawler was permitted to fetch the page |
| Retrieved | The page was pulled into a model's context for a specific answer |
| Mentioned | Your entity appears in the generated text |
| Cited | Your domain appears in the response's attribution data |
| Recommended | The answer advises the reader to choose you |
Each step is a precondition for the next and none of them implies it. A page can be crawled and never retrieved, retrieved and never cited, cited and never recommended, mentioned unfavourably and counted as a mention all the same. Crawled is not cited develops the first half of that chain; robots.txt access in particular buys you nothing beyond eligibility.
What to do instead of chasing a position
The work that plausibly helps is the same work that has always helped, done properly. Make the site readable — that is the technical SEO audit. Make it unambiguous about who published what, and structure sections so a passage can stand alone — that is GEO and AI readiness. Then measure what answers actually say, with a prompt set that reflects real buying questions, through prompt monitoring.
Judge progress on the terms the measurement supports: a mention rate that moved outside its confidence interval across several runs, on unbranded prompts, with volatility low enough that the movement is not noise. That is slower and less satisfying than watching a position number, and it has the advantage of being true.
Who this page is for
Anyone who has been asked for "our ChatGPT ranking" by someone who assumed it exists. The useful reply is that the underlying question — are we present in AI answers, and is that changing — is answerable, just not in that unit.
Official sources & further reading
- Web search tool — OpenAI
- AI features and your website — Google Search Central
Related reading
Frequently asked questions
So SiteRank AI cannot tell me how I compare with a competitor?
It can, within stated limits. Share of voice and citation share compare you against the competitor list you defined, across your prompt set, with both denominators shown and intervals on every figure. When two intervals overlap, there is no ordering to report, and the product does not manufacture one.
Does any tool track LLM rankings?
Some products use the word. What sits behind it is generally a mention or citation rate, sometimes an order-of-appearance heuristic, sometimes an aggregate with undisclosed weights. Ask any vendor for the numerator, the denominator, the sample size and the raw responses. If those cannot be produced, the number is not evidence.
Is order of mention worth anything?
It is observable, so it can be recorded, but no provider documents it as meaningful and it varies between samples of the same prompt. SiteRank AI does not report it as a position.
What about tracking Google AI Overviews?
Google reports clicks from its AI features within the "Web" search type in Search Console rather than as a separate breakout, so no first-party report isolates AI feature performance. Anyone claiming an AI Overview position is inferring it from their own sampling, and should say so.