GEO Ranking: Why There's No Position Number in AI Search

Why there's no position number to track

A ranked list requires a stable, ordered set of candidates that a system sorts the same way for every viewer of a given query — that's what "position four" means. Generative answers aren't sorted lists; they're one synthesized response assembled at the moment of the request, drawing on retrieval, training data, and (for some engines) live web results, then compressed into prose that may cite a couple of sources, several, or none. There's no "position four" in a paragraph. The instability is measurable: in SparkToro and Gumshoe's consistency study, volunteers ran the same prompts through ChatGPT, Claude, and Google's AI Overviews 2,961 times, and there was fewer than a 1 in 100 chance of getting the same list of brands twice — and roughly 1 in 1,000 of getting the same list in the same order. Vendors who sell a "GEO rank" number are applying a rank-tracking mental model to a system that doesn't produce ranks, which is exactly the confusion this term causes.

What actually gets measured instead

Citation frequency, share of voice, and presence per engine replace "rank" in GEO measurement, and each answers a different question:

• Citation frequency — out of a fixed, repeated set of prompts, how often does an engine mention your brand or site at all? This is a rate, not a position: "cited in some share of runs," never "ranked third."
• Share of voice — when your brand is mentioned, how does that frequency compare to named competitors across the same prompt set? A brand cited in a minority of runs while a rival appears in most of them has a real gap, even though neither has a "rank." This comparative framing is covered in detail in AI Share of Voice: How to Measure It vs. Competitors.
• Presence versus absence, per engine — ChatGPT, Perplexity, and AI Overviews draw on different retrieval systems and different indexes, so presence in one says nothing about presence in another. A brand can be cited reliably in Perplexity's web-grounded answers and be invisible in ChatGPT's, and both facts can be true at once.

None of these three is a substitute rank — each is a sampled rate that only means something next to the sample it came from.

Why one AI search is a bad test

A single spot-check tells you almost nothing, because the three biggest sources of variation in an AI answer — prompt phrasing, session/conversation state, and model version — are all invisible in a one-off test. Ask "best project management software for startups" and "what project management tool should a startup use" and you can get citation sets that barely overlap, even though a human would treat those as the same question. Run the same prompt in a fresh session versus one with prior conversation turns, and retrieval or context can shift the answer. And the sources engines cite churn on their own: Profound's volatility study re-ran identical prompts one month apart and found the cited domains changed by 40.5% on Perplexity up to 59.3% on Google AI Overviews. Anyone who checks a brand's visibility with one prompt in one chat window and concludes "we're not showing up in AI search" is drawing a rate-based conclusion from a sample size of one, which is the same error as judging a coin biased after a single flip.

A repeatable way to measure it yourself

How to measure your AI citation rate manually

1. 1

   Build a fixed prompt set of 15-25 realistic buyer questions

2. 2

   Pick your engines (ChatGPT, Perplexity, AI Overviews) and test separately

3. 3

   Run each prompt in a fresh session per engine and record citations

4. 4

   Tally citation frequency and rough share of voice per engine

5. 5

   Repeat the full set on a different day or week later

6. 6

   Log prompt text, date, engine, and model version for comparability

Before paying for any tool, run this manually — it takes an afternoon or two and gives you a real baseline instead of an anecdote.

1. Build a fixed prompt set, and make it bigger than feels necessary. Write dozens of realistic questions a buyer would actually ask, covering category questions ("best CRM for a small sales team"), comparison questions ("X vs Y for small business"), and direct brand questions ("is X good for enterprise"). Keep the exact wording — you'll reuse it every time. Distinct prompts are worth more than repeat runs of the same prompt: re-running one question produces answers that resemble each other, so each repeat adds less information than a new question does.
2. Pick your engines. At minimum, test ChatGPT and Perplexity, and check Google's AI Overviews where they appear. Treat each engine as a separate measurement, not one combined score.
3. Run each prompt in a fresh, logged-out or new session for each engine, so prior conversation history doesn't bias retrieval. Record whether your brand is mentioned, whether competitors are mentioned, and what source (if any) the engine cites for the claim.
4. Tally citation frequency per engine: mentions ÷ total prompts run, per engine. Do the same for each named competitor to get a rough share of voice.
5. Repeat the entire set a few weeks later before drawing any conclusion — and be careful what you read into the difference. With a small prompt set, a large swing between two rounds is exactly what random sampling produces even when nothing real has changed. The margin of error shrinks slowly: halving it takes roughly four times as many answers. Treat a change as real only when it holds across several rounds.
6. Log everything, including the exact prompt text, date, engine, and model version if the interface shows one, so a later re-run is actually comparable.

This method won't produce a rank. It produces a sampled citation rate with a documented margin of noise — which is a more honest and more useful number than a single screenshot of one ChatGPT answer, and it's the same basic logic behind the AI Selection Rate metric, which weights citations by whether they were actually used to support the answer rather than just name-dropped.

What this looked like on my own site

Numbers first, because that is the whole point of this article. KinetixSEO tracks its own domain with exactly this setup: a fixed set of thirteen prompts, spread over the topics I want kinetixseo.com to be found for, checked per engine. The most recent round, on 13 September 2026, came back like this.

Perplexity cited us for three of the ten prompts it answered. Two of those were about GEO audit tools, the third was "How do I check if ChatGPT can see my website?" Every other engine cited us for none of them. Not ChatGPT, not Gemini, not Claude (with or without web search), not Google's AI Overviews, not Copilot.

Is that a good result? No. Is it the full picture? Also no, and this is exactly where the rest of this article earns its keep. Thirteen prompts checked once is a directional read, nothing more. The next round can look quite different without a single change on the site, so I will run the same set again in a few weeks before I believe any trend. And the engines are not measured the same way: for ChatGPT, Gemini and Claude without web search, the tracker asks the model's API whether it knows us, which tests what the model remembers from training. That is not what someone sees in the consumer app after a live web search, and the two often cite different sources. A "not cited" there is a warning sign, not proof.

What the round does show clearly is the per-engine point from earlier. Perplexity, which searches the live web for every answer, is the only engine citing us, and only for the topic closest to what the product actually does. Presence in one engine told me nothing about the others.

My honest status: visible in one engine, for one topic, on a sample too small to call a trend. Not a rank. Lesson learned: the number you screenshot is not the number you have.

Treat these numbers as sampled, not tracked

Rank tracking vs. GEO citation sampling

	Keyword rank tracking	GEO citation sampling
Underlying system	Deterministic, fixed results page	Probabilistic, generated per request
What's reported	A position number (e.g. #4)	A sampled rate (e.g. 8/20 prompts)
Stability day to day	Consistent, polled repeatedly	Can shift with phrasing, session, model version
Comparable across tools?	Yes, same methodology industry-wide	Only within the same prompt set and date range

Say this plainly to anyone reading a GEO report: a citation rate from a small prompt set run once is not comparable to a keyword holding position four for six months, and reporting it as if it carries that kind of stability is misleading. A rank-tracking number describes a deterministic system polled repeatedly with consistent results. A GEO citation rate describes a probabilistic system sampled a limited number of times, subject to prompt wording, session state, and silent model updates — all three of which can shift the number without anything on your website changing. The right way to present a citation rate is with its sample size and date range attached ("cited in N of M prompts, run between [dates], via ChatGPT web") and paired with a re-run a few weeks later to show whether it's stable or noisy. Anyone presenting a single-session citation percentage with the confidence of a search rank — no sample size, no re-test, no engine breakdown — is either misunderstanding the measurement or hoping you won't ask how it was produced. If your brand's numbers look thin, the more useful next step is diagnosing why, which is covered in Why Is My Brand Not Showing Up in ChatGPT?

Frequently asked questions

What is a good GEO ranking?

There's no "good ranking" to hit because GEO doesn't produce a rank — the closer question is what citation frequency or share of voice looks healthy for your category, and that depends entirely on how many competitors are being cited across the same prompt set. A brand cited in half the runs while no competitor gets close is in a strong comparative position; the same rate means little if a rival shows up in nearly every answer. Judge the number against competitors in the same sample, not against an absolute benchmark, since no universal "good score" exists across engines or categories.

Can I track my GEO ranking daily like a keyword rank?

Not meaningfully, because daily sampling captures noise, not trend. Since AI answers vary by prompt phrasing, session state, and undisclosed model updates, a single day's citation count can swing widely without any real change in visibility. A more reliable cadence is running the same fixed prompt set every few weeks, on a fresh session each time, and looking at the trend across several rounds rather than any one day's number.

Why does my brand show up in ChatGPT but not Perplexity?

Because the two engines use different retrieval and indexing systems, so presence in one doesn't imply presence in the other. Perplexity leans heavily on live web retrieval and tends to cite recent, well-structured pages directly; ChatGPT's citation behavior depends on which mode is active and draws on a different mix of training data and browsing. Testing each engine separately, as described above, is the only way to know where the gap actually is.

Do GEO tools give an exact ranking number?

Some tools display a single score, but that number is still built from sampled citation rates across a prompt set, not a polled position on a fixed results page. If a tool presents its output as a precise rank without disclosing sample size, prompt set, or re-test interval, treat that precision as manufactured rather than measured — the underlying data is probabilistic no matter how the dashboard rounds it.

How many prompts do I need to test to trust the result?

More than most people expect. After running their consistency study, SparkToro concluded that visibility measured across dozens to hundreds of prompts, run multiple times, is a reasonable metric; a set of a couple of dozen prompts gives a directional read at best. Grow the number of distinct prompts before adding repeat runs, because repeats of the same question cluster together, and only trust a trend that holds across several rounds spread over weeks.