How CiteGist measures what AI answers.
CiteGist compares captured AI answers, including the names, recommendations and sources they contain. Counts and percentages describe those observations; composite scores combine several signals.
WHAT WE MEASURE
Four things, measured on the same answers
| What | The question it answers | Unit |
|---|---|---|
| Engine baseline | Engine changes start a separate measurement baseline | Bing starts a separate baseline at the collection switch. Copilot remains active and keeps its separate history. Changes across the switch are not like-for-like growth; CiteRank weights remain 60/25/15. |
| A question page | What do the engines answer, and whom do they name, for this exact question? | Who was mentioned, at what position, and the share of answers that mention each one |
| A brand or tool page | Across every question it appears in, where does this brand stand? | Share of answers per question, over time |
| A voice | Which people do the engines reach for, unprompted, in a segment? | Share of Voice: the share of a Segment's answers that mention them |
| CiteRank | How visible is a person across our measured AI answers? | 0 to 100, from the three weighted signals below |
THE PIPELINE
From a question to a page
Ask
The same question goes to the active public engines. Sample counts vary by question and measurement series. Hebrew questions are asked from Israel.
Keep the answers
Every answer is stored word for word, with the sources it cited and the searches the engine ran before answering, so numbers can be recomputed from the stored answers.
Read the names, with their role
Each name is filed as a recommended option, a cited source, an employer, a platform, or a passing mention. Only recommended options are ranked. A publisher whose article the engines cite is a source, never a recommendation.
Count
A name's standing on a page uses the share of answers that mention it. A name that appears in two or more answers is confirmed.
Decide if the page is public
A page is published when at least three engines answered, at least three names are confirmed, and the answers carry substance and citations. Pages below that bar stay reachable by link but are not listed or indexed.
Title the page by what was answered
If the engines' own searches show they answered a narrower question than the one typed, the page says so in its title, and the exact prompt sent is shown underneath.
The five roles a name can be filed under
- RecommendedRANKED
- SourceNOT RANKED
- EmployerNOT RANKED
- PlatformNOT RANKED
- MentionNOT RANKED
SAMPLING
Why “more than once” matters
Ask an AI engine the same question twice and you get two lists. Research on brand answers finds that repeating the same prompt explains about a third of the variation, and that asking more engines and more wordings improves precision more than asking the same engine again and again. We ask the active public engines, and we tell you how share of answers that mention each name. A larger share is a stronger finding; a small share is a lead.
The variance-components study (arXiv:2607.13304)the Pebblous iteration reanalysis
MEASUREMENT NOISE
How much of a change is just noise?
Ask an AI engine the same question twice and the answers differ. Before we show any movement, we measured how much. We re-asked the same prompts on five engines, either twice on the same day or a week apart, and counted how far the results moved when nothing had really changed. These are measurements on our prompts, not on CiteRank itself.
Share of Voice counts every measured answer to a prompt or Segment, one per answer that mentions the entity. We do not publish a movement label when the available observations are too thin to support one.
A name that 40% or more of a prompt's answers agreed on was still at 40% or more on a re-ask in 75% of cases (163 prompts, same day, three samples per engine).
Rank place was not stable enough to show. On a re-ask with nothing changed, 67% of entities named by three or more engines moved two or more places, and 39% moved five or more (153 prompts). We do not show rank-place movement.
A week apart looked like the same day. The noise comes from asking again, not from time passing.
The CiteRank score itself: rescoring 118 people on all five engines moved their v6.4 Citation Rate by 2.7 points on average, and the two runs ranked them with a rank correlation of 0.60. For v7, scoring the previous and the latest run of 736 questions asked twice ranked the top 250 people with a rank correlation of 0.87 (28 days to September 24, 2026). What we have not measured yet is how the engines differ from one another, and we will not rank the engines before then.
Dates we treat as breaks
A score or rate whose two ends fall on either side of one of these dates is flagged, not trended.
2026-07-04: Citation Rate moves to source attribution.
2026-07-26: each probe asks for exactly 10 names, so rates before this date are not comparable.
2026-08-16: one engine (Grok) is removed from the set because the vendor cannot collect it.
2026-09-16: the pinned prompt series is re-cut; comparisons across versions are breaks.
2026-09-27: CiteRank moves to provisional v7, so v7 and v6.4 scores are not comparable.
Outside changes we watch, as reported by others: on 2026-05-07 ChatGPT changed how it shows links (reported by Profound and Qwairy); on 2026-01-27 Gemini 3 became the default behind Google AI Overviews (reported by SE Ranking).
CONFIRMATION
What confirmation looks like on a page
Confirmed names rank first. Names from a single answer are shown too, labelled, because where the engines disagree is exactly what you came to see.
MOVEMENT
The same questions, every Monday
A fixed set of questions is asked every Monday with the same setup. A change between Mondays is a change in what the engines say, not in how we asked. Single-week wiggles are weather; multi-week trends are climate.
VOICES
Voices and the index
A person becomes a voice when the engines name them unprompted in two or more answers from different engines, or across two questions. The Israel index is built from questions asked about Israel in Hebrew and in English; the two languages surface different people, and both belong.
CITERANK
CiteRank - the outcome score for a person
CiteRank measures who AI names unprompted across active listed prompts in a 28-day window. A prompt that names a voice never counts for that voice.
CiteRank is a relative, provisional 0–100 score. It is not a citation percentage or a guarantee of appearing in future answers.
Each signal is scaled so the top person gets its full weight. The score is provisional: rescoring the same questions agrees closely at the top of the board, but the score has not yet reached our bar for agreeing with an outside measure of prominence.
| # | Signal | Weight | What it measures |
|---|---|---|---|
| 01 | Share of Voice (weighted by question demand) | 60 | Per Segment, the demand-weighted rate at which answers mention the Voice, compared with the Segment leader and averaged by each Segment's share of demand. |
| 02 | Breadth | 25 | The number of distinct prompt and engine cells in which the voice is named, log-scaled against the maximum. |
| 03 | Domain Reach | 15 | The number of distinct domains cited in answers that name the voice, log-scaled against the maximum. |
Inside Domain Reach
Domain Reach counts the distinct domains cited in answers that name the voice, then log-scales that count against the maximum. Those are the sources each answer cites, not necessarily sources that cite the person. Own-site attribution is tracked separately as a diagnostic with weight 0.
Eligible voices have an approved public profile and are named in at least 3 answers on at least 2 engines.
2026-09-27: CiteRank v7 - Share of Voice (weighted by question demand) 60, Breadth 25, Domain Reach 15.
How scores read
These labels describe a current score. They are not guaranteed bands, and a score can move between them as answers change.
Citability - will AI cite this piece of content?
Citability is a leading A–E grade on a single piece of content. It is reported beside CiteRank and is deliberately never summed into it.
| Component | Weight | What it measures |
|---|---|---|
| Information Gain | 40% | How much genuinely new signal the piece adds beyond what engines already have. |
| Original Framing | 25% | A distinct framing or claim engines can attribute to you, not a restatement of consensus. |
| Credibility | 20% | Verifiable sourcing, controlled domains, and a track record engines can trust. |
| Structure / Extractability | 15% | Clean structure engines can lift cleanly - headings, direct claims, quotable lines. |
Postability - is this draft ready to publish?
Postability is the craft companion to Citability, judged against the author's Voice DNA on a 0–100 scale. Advisory only; it never blocks publishing.
| Component | Weight | What it measures |
|---|---|---|
| Hook | 35% | Does the opener stop the scroll? |
| Voice | 40% | Does it sound like you, measured against your Voice DNA? |
| Format | 25% | Platform-fit and scannability - does the close land? |
HONEST LIMITS
What we don't claim
AI answers vary
AI answers vary by account, region, and model version - we fix our measurement setup so changes are comparable week to week.
A sample is a sample
Single-week wiggles are weather; multi-week trends are climate.
The plumbing changes
AI retrieval plumbing changes frequently - which is why every number on CiteGist carries the date it was computed.
Extraction is a judgement
A model reads each answer and decides what is a recommendation and what is a citation. A daily check re-reads a sample of pages and pulls any page where that judgement failed.
ANTI-GAMING
What CiteRank deliberately ignores
We don't count follower counts
Reach on a social platform is not citation. A 500-follower researcher can outrank a million-follower account.
We don't score your page wording
CiteRank reads what engines name and cite in their answers, not the keywords on your own pages.
We don't sell rank
A subscription unlocks your own diagnostics and actions. It cannot buy a position - the answers decide that.
REFERENCES
The evidence this rests on
The sampling design and the weights are grounded in external, verifiable research - not house opinion.
The foundational framework for optimizing content toward generative-engine citation.
Repeating the same prompt explains roughly a third of the variation in brand answers.
More engines and more wordings improve precision more than repeating one engine.
Which platforms LLMs cite most, by engine, at scale.
Prompt-level analysis of which sources surface across generative engines.
Measured lift in AI citations following structural and distribution changes.
QUESTIONS
The details
What is a sample?
One sample is one captured answer from one engine. Sample counts vary by question and measurement series.
What does Share of Voice mean?
It is the share of answers to a question, or to all the questions in a Segment, that mention it. An answer counts once however often it repeats the name, and one answer can mention several, so shares do not add up to 100%.
Why is a brand I saw in an answer not on the page?
Because the answer cited it as a source or named it in passing rather than recommending it. Sources are listed in the citations panel.
Why did a page disappear?
It fell below the publish bar after a sweep. It is still reachable by its link and returns when it clears the bar.
Are Hebrew and English rankings the same?
No. The same question in the two languages surfaces largely different names, so each has its own page, and both feed the Israel index.
What is CiteRank?
CiteRank measures a person's visibility in AI answers. The score out of 100 combines Share of Voice (weighted by question demand) (60 points), Breadth (25 points), Domain Reach (15 points). The measurement covers a 28-day window. Questions that name the person are excluded. Own-site attribution is tracked separately.
How is the leaderboard ranked?
By provisional CiteRank v7, computed every six hours over the latest 28 days of answers, minus each person's measurement noise. The noise is measured on every refresh by scoring the previous and the latest run of each question separately, and it is wider for people named in fewer answers. People whose result rounds to the same number share a rank, and tied voices show the same rank number. To be ranked, a person must be named in at least 3 answers on at least 2 engines. Sorting by Mentions ranks every eligible person.
What is Citability?
Citability is a leading A–E grade on a single piece of content that estimates whether AI will cite it. It weights Information Gain 40%, Original Framing 25%, Credibility 20% and Structure/Extractability 15%, with bands A ≥ 85, B ≥ 70, C ≥ 55, D ≥ 40 and E below 40. It is reported beside CiteRank and is never summed into it.
What is Postability?
Postability is the craft companion to Citability: a 0–100 grade judged against the author's Voice DNA that asks whether a draft is ready to publish. It weights Hook 35%, Voice 40% and Format 25%. It is advisory only and never blocks publishing.
See whether AI is citing you.
Check whether you have a measured CiteRank.