AI rank tracker: track AI rankings across four models

An AI rank tracker measures whether AI assistants name your brand when someone asks them for a recommendation — and where in the answer you appear. It is a rank tracker with no ranks in the familiar sense. There is no results page, no ordered list, no URL sitting at position three. There is a paragraph of prose, and either your name is in it or it is not.

§What is an AI rank tracker?

An AI rank tracker replays a fixed set of buying questions against AI assistants on a schedule and records, for each answer, whether your brand was mentioned, where your mention falls against the other brands named, which sources were cited, and how the framing read. SimplyRank runs that against four assistants — ChatGPT, Claude, Perplexity and Gemini — on the same day, from the same prompt set. The category is young enough that it is still sold under several names: AI ranking tracker, AI visibility tracker, AI performance tracker. They describe the same job.

The category exists because a chunk of vendor research now happens inside an assistant rather than on a results page. OpenAI's own usage research splits ChatGPT conversations into “doing”, “asking” and “expressing” buckets, with the “asking” bucket — decision support — the dominant one. That is the surface where a shortlist gets made, and it produces no impression, no click and no query report.

It also exists because that attention did not move to one place. OpenAI's CEO Sam Altman told the U.S. Senate Commerce Committee in May 2025 that ChatGPT will “probably not” replace Google as the primary search engine, while conceding some queries are “definitely better done on a service like ChatGPT”. Google DeepMind's CEO Demis Hassabis, in a 2025 Wired interview, said the summary and conversational modes are both “going to be growing and necessary”. Two of the people best placed to know are describing a fragmented surface — which is precisely why one blended “AI visibility score” is the wrong shape for the measurement.

§AI rank tracker vs a normal rank tracker

A normal rank tracker measures where a URL sits in an ordered list, and the search engine publishes that order for it to read. An AI rank tracker has no order handed to it. Position has to be derived from the text, the unit of tracking is a whole question rather than a keyword, and four assistants can disagree about the same question in the same week. Nearly every other difference falls out of those three.

The practical consequence: the two tools rarely replace each other. Your SEO rank tracker watches your pages. An AI rank tracker mostly ends up watching other people's — the comparison posts, review listings and category round-ups the assistants lean on when they assemble a shortlist.

§What an AI rank tracker measures

Four things per question, on every model, in the same shape so they can be compared:

• Mention. Did the answer name you at all? Your brand name and its aliases are matched against the response text. This is the floor: while mentions are rare, position and sentiment are describing answers almost nobody sees.
• Position in the response. Where your first mention falls against every other tracked brand named in that same answer. SimplyRank finds the first mention of each tracked brand, orders them by where they appear in the text, and takes your rank in that ordering; averaged across the prompt set, that is your average position. The assistant never reports a rank — it is derived, which is why it only means anything when the prompt set and model version are held still.
• Citations. The source links carried in the answer, stored with their domain and their order, and flagged when they point at your domain or a competitor's. This column is where the editorial work is, because it names the page that has to change.
• Sentiment and competitor overlap. How the framing read, and which rival names appeared beside yours. Being named dismissively next to three competitors is a different result from being named first, and a mention count cannot tell them apart.

Reading those four per model, side by side, is the whole workflow. Collapsing them into one score is the common failure: the number moves, and nothing in it says which assistant moved or why.

§Why the same prompt returns different brands on each model

Ask four assistants the same question in the same hour and you will get four different shortlists. That is not a bug in the measurement — it is the thing being measured, and it happens because the four are built on different constraints. Perplexity, in its CEO's words, is “restricted to sourcing information directly from the web, eschewing any reliance on pre-existing knowledge within the model”, so a brand it cannot find a live source for simply does not appear. Another assistant answering from training memory has no such restriction and will happily name you from something written a year ago. A third leans on what the Google index currently trusts. Same question, three different definitions of “who counts”.

Two ways to use it. First, pick your diagnostic per assistant: where position is stable, a drop is a signal worth investigating; where list order shifts between runs anyway, the same drop is noise. Second, treat the citation row as the work queue — one assistant inlines whatever it has, one wants editorial sources, one wants fresh ones, one mirrors what search already trusts, and those are four different content jobs.

The per-engine detail lives on its own page, because each assistant deserves more than a column: the ChatGPT rank tracker, the Claude rank tracker, the Perplexity rank tracker and the Gemini rank tracker. Google AI Overviews sits on the AI Overviews tracker page as research rather than a fifth scan stream, because it behaves more like a search result than like an assistant.

§How to track AI rankings week to week

Ad-hoc checks feel productive and prove nothing. Type a question into an assistant twice in an afternoon and the shortlist can change; do it a month later, from a different account, after a model update, and you are comparing two different experiments. A standardised prompt set on a schedule is what converts that into a measurement. The discipline is deliberately boring:

• The prompt set stays still. The same questions run every week, so the only thing changing between scans is the answers. Prompt allowances are set by plan — 50 prompts on the Starter plan, up to 300 prompts on Advanced — and the set is yours to edit when your market moves.
• The instrument is stamped on every scan. Each session records the model versions, the sampling settings and the geography it ran under, snapshotted when the scan is created rather than read live — so an account setting edited mid-run cannot split one scan across two instruments.
• Benchmark versions are carried, and changes are annotated. Every scan carries a benchmark version alongside the model version. When we change how the instrument works, the version increments and the trend chart marks the boundary, so a step in the line reads as “the instrument moved” rather than “we lost visibility”.
• Re-runs cannot inflate a trend. Multiple scans on the same day collapse to one canonical session per brand per day per benchmark version for charting, so re-scanning to check something never shows up as growth.
• All four assistants run the same day. Same questions, same window, four panels — which is the only way the disagreement between them is readable.

Without a fixed set, stamped versions and per-model panels, you are producing a number rather than a measurement. The number will still move. You just will not know why.

What the assistants reward (and what they ignore)

Want to look before you sign up for anything? What is genuinely free in AI rank tracking compares the free AI ranking checkers and limited plans, each with the date we last checked its terms.

§How to evaluate any AI rank tracker

Every vendor in this category will show you a dashboard with a number going up. These are the questions that tell you whether the number means anything. They work on any tool, including this one — take them into your next demo and ask them in this order.

1. Which models, and which versions of them? Ask for the model identifier, not the brand name. "ChatGPT" is a product; the thing that answers your prompt is one specific model version. A vendor who cannot name it also cannot tell you when it changed underneath your chart.
2. How often does the benchmark run, and is the cadence fixed? Ad-hoc checks produce anecdotes. A standing cadence produces a trend. Ask whether every brand in the account runs on the same schedule, or whether some drift and quietly break the comparison.
3. Is the prompt set fixed, or editable? You want both — a set that stays still so weeks are comparable, and the ability to change it when your market does. Then ask the follow-up: when you edit a prompt, does the history reset, or does the old trend silently absorb a new question?
4. Do they store the raw response text? A score with no answer behind it cannot be audited. Being able to read the sentence that named a competitor instead of you is what turns a dashboard into a content brief. A number on its own does not.
5. How do they handle model nondeterminism? Ask the same assistant the same question twice and you can get two different shortlists. Ask how the vendor holds the sampling settings still, and whether they say plainly that a single week-on-week wobble is noise.
6. What happens to the trend when a provider ships a new model? Worth asking first, because it is the question that decides whether the chart is readable at all. If the instrument changes and the line does not say so, every dip looks like a loss of visibility.
7. Is position measured, or only presence? Being named eighth in a list is not the same as being named first, and a yes/no mention flag hides the difference. Ask how position is derived, in words you can repeat back to your team.
8. Is location part of the instrument, and can you export the rows? Assistants answer differently by region, so ask whether geography is set per brand or once per account. Then ask what export exists, and on which plan — before you need it at quarter-end.

How the tools compare on those questions

The cross-model category is crowded, and the vendors below all advertise multi-engine coverage with different trade-offs on cadence, version stamping and how citations surface. Cells reflect what each vendor publishes on its own product pages — not independent testing — so follow the link in each row before treating any of it as a verdict.

Where we land, plainly: SimplyRank is a four-model, weekly, pinned-version, location-aware tracker that treats citation context as a first-class metric. That is a fit decision, not a feature-count one. Daily cadence suits high-volume content programmes running experiments on AI surfaces; weekly suits editorial programmes moving at the speed the answers actually change. Pick against your own calendar, not against a table.

§What an AI rank tracker cannot tell you

This category is new enough that the marketing has run ahead of the measurement. Four limits are worth knowing before you buy anything — ours included.

§When the models are not recommending you

When mentions are thin across several assistants at once, the answer is rarely “more content”. It is usually structural: the pages the models trust in your category are not your pages. Three places to look first, in this order:

1. Third-party comparison and alternatives pages. Every assistant leans on these — review sites, versus pages, round-ups, forum threads. Read the ones your citation column already names and check that you appear, framed correctly and currently. Claim and update outdated review-site entries; publish a clear, specific comparison page of your own.
2. Buyer-fit specificity. “For businesses of all sizes” gives a model nothing to anchor to. Pages that name an industry, a team shape or a job to be done give it a reason to raise you on the questions that match — and those are the questions worth tracking.
3. The shape of your proof. The assistants weight evidence differently: some want editorial coverage on trusted publications, some want a live citable source, some want whatever search already ranks. Your citation column tells you which kind of proof is missing, per model. That is the difference between a content plan and guesswork.