How each index is defined, where the idea comes from, and how to read it. Version 1.1.0. Notation: O = observed count, E = expected count, pi = share of label i.
Question it answers. Two ideas A and B are each studied in a context C (a field, or the whole of OpenAlex). Are papers that combine A and B rarer than the popularity of A and B would lead us to expect?
| Symbol | Meaning (OpenAlex search counts) |
|---|---|
| n(C) | works matching the context C; all works in OpenAlex when no context is given |
| nA = n(A ∧ C) | works matching idea A and the context |
| nB = n(B ∧ C) | works matching idea B and the context |
| O = n(A ∧ B ∧ C) | works matching both ideas and the context (observed) |
Alternative spellings typed with ; are joined by OR inside each idea, e.g. ("drone" OR "UAV").
If a work in C mentioned A with probability pA = nA/n(C) and, independently, B with probability pB = nB/n(C), the number of works mentioning both would be Binomial(n(C), pApB) with mean
The ratio O/E is the classical observed-to-expected measure of co-occurrence: it is the "association strength" of co-word analysis[1],[25] and, for journal pairs, the "commonness" used to score the novelty of papers[9], commonness = Nijt Nt / (Nit Njt).
Because pApB is small, the binomial count is very close to Poisson(E): by Le Cam's inequality[8] the total-variation distance is below n(C)(pApB)2 = E2/n(C). With X ~ Poisson(E),
where plow = P(X ≤ O) is the exact one-sided p-value for H0: "A and B are combined at the independence rate" against H1: "they are combined less often". So G ≥ 95 means plow ≤ 0.05 and G ≥ 90 means plow ≤ 0.10.
| G | Reading shown by ScholarGauge |
|---|---|
| ≥ 90 | studied much less than its parts suggest: a likely gap (with O = 0: "not done, although its parts are common") |
| 60–90 | somewhat less studied than expected |
| 40–60 | about as studied as its parts suggest |
| 10–40 | somewhat more studied than expected |
| ≤ 10 | crowded: studied more than its parts suggest |
Counts are whole numbers, so when O equals E the index sits a little below 50 (O = E = 12 gives 42.4). When E < 1 and O = 0, G = 100(1 − e−E) < 64: a missing combination of two rare ideas is weak evidence of a gap, and ScholarGauge says so.
A = the chosen distribution, B = the chosen censoring scheme, each searched together with the chosen model; C = papers on the model (all of OpenAlex if no model is chosen); O = papers with all three. Each dictionary term is searched with its usual spellings joined by OR.
Comparing how often two things occur together with how often they would by chance is the basis of co-word analysis (Callon et al., 1983)[1] and of normalised co-occurrence measures (van Eck and Waltman, 2009)[25]. Novelty studies apply it to pairs of journals cited together: Uzzi et al. (2013)[24] compare observed pair frequencies with randomised citation networks (z-scores), Lee, Walsh and Wang (2015)[9] use the observed-to-expected ratio ("commonness"), and Wang, Veugelers and Stephan (2017)[28] score first-time pairs. The notion of a gap between two literatures that are each well studied but rarely linked goes back to Swanson's "undiscovered public knowledge" (1986)[23].
What is ScholarGauge's own: turning the observed-versus-expected comparison for two ideas into a Poisson tail probability on a 0–100 scale, computed from OpenAlex search counts. It is not a published standard index, so in a paper describe it by its formula (see How to cite).
The topic map reads the N most relevant OpenAlex papers for your query (100 or 200). For two keywords r and c among them, O = papers carrying both, nr and nc = papers carrying each, N = papers read that have keywords, and
with the same index G. A cell is outlined as a gap when G ≥ 90 and E ≥ 1. The reliability view does the same for distribution × censoring scheme among papers whose design is recognised.
The profile table measures how varied a researcher's papers are along one dimension at a time: OpenAlex topics, subfields, fields and domains; OpenAlex keywords; journals; co-authors; co-author countries; UN SDGs; and, when the reliability dictionary recognises them, distributions, censoring schemes, models and methods. The same indices are used in ecology for species and in scientometrics for disciplines[22].
Paper j carries a set Lj of labels in the dimension (one primary topic, but possibly several keywords, co-authors or SDGs). Its weight 1 is split equally among its labels, so a paper with many labels does not count more than one with a single label[27]:
n (column "Papers") is the number of papers with at least one label in that dimension.
| Column | Formula | Meaning | Source |
|---|---|---|---|
| Kinds | S = #{i : wi > 0} | richness: number of different labels | – |
| Effective kinds | N1 = eH, H = −Σ pi ln pi | Shannon entropy H turned into a number of kinds: the number of equally common labels that would give the same H | Shannon (1948)[20]; Hill (1973)[5]; Jost (2006)[7] |
| Simpson | D = 1 − Σ pi2 | Gini–Simpson index: chance that two papers drawn at random (with replacement) have different labels | Simpson (1949)[21] |
| Evenness | J = H / ln S (S ≥ 2) | Pielou evenness, 0–1; 1 when all labels are equally common | Pielou (1966)[14] |
| Most common | arg max pi | the commonest label and its share; ties are reported | – |
Richness, effective kinds and the inverse Simpson index are the Hill numbers of order 0, 1 and 2,
(1D is the limit as q → 1), so S ≥ eH ≥ 1/Σpi2 always, with equality only when all labels are equally common (Hill, 1973; Jost, 2006)[5],[7]. Effective numbers are easier to compare than entropies: doubling eH means twice as many equally common kinds.
Simpson's own estimator uses ni(ni − 1)/[n(n − 1)] to remove small-sample bias. With fractional counts that correction is not defined, so ScholarGauge uses the plug-in form; for a profile with few papers all indices are descriptive, not estimates of a population value.
Variety ignores how different the labels are: two topics in statistics count as much as statistics and history. Rao–Stirling diversity weights each pair of topics by their distance dij:
pi = share of the profile's papers whose primary OpenAlex topic is i. This is Rao's quadratic entropy[19] and Stirling's general diversity heuristic Σi≠j dijα(pipj)β with α = β = 1[22], the form used to measure interdisciplinarity in bibliometrics[16],[18],[10].
ScholarGauge measures distance on OpenAlex's four-level classification (4 domains › 26 fields › 252 subfields › 4,516 topics)[13], in equal steps up the tree, as in taxonomic distinctness in ecology (Warwick and Clarke, 1995)[29]:
| The two topics share | dij |
|---|---|
| the same subfield | 0.25 |
| the same field, different subfields | 0.50 |
| the same domain, different fields | 0.75 |
| different domains | 1.00 |
Bibliometric studies often use 1 − cosine similarity of citation profiles instead (Porter and Rafols, 2009); the tree distance needs no citation data and is the same for every field, which keeps profiles from different disciplines comparable.
Range. Δ = 0 when all papers share one topic. Because dij ≤ 1, Δ ≤ 1 − Σpi2 (the Gini–Simpson index), with equality when all topics lie in different domains; so Δ < 1. Some authors sum over i < j only, which gives half this value; ScholarGauge sums over all ordered pairs, as in Rao (1982).
The share of a researcher's later papers that open a topic new to them. Papers are grouped by year; the topics of the first year form the starting set T. For each later year y, in order, every paper of year y whose primary topic is not in T counts as new; then the topics of year y are added to T.
Papers of the same year do not count each other as earlier work. 0% means every later paper stayed in topics already worked on; 100% means every later paper opened a new topic. It is a simple descriptive rate in the spirit of studies of topic switching (Zeng et al., 2019)[30] and of tradition versus innovation in research strategies (Foster, Rzhetsky and Evans, 2015)[3], not the measure of either paper.
The third row of cards on a profile. All of them are computed from the papers ScholarGauge has read on the page, so load all papers ("Show more" on Scholar) before analysing; a profile with papers left unloaded gives smaller values than Scholar's own.
c(1) ≥ c(2) ≥ … ≥ c(n) are the citation counts of the profile's n papers in decreasing order (Google Scholar's counts on a Scholar profile, OpenAlex's in the Android app and on OpenAlex author profiles). Nj is the number of authors of paper j (OpenAlex), and the academic age is a = Y − y0 + 1, where Y is the current year and y0 the year of the first paper.
| Index | Definition | Meaning | Source |
|---|---|---|---|
| h-index | h papers have at least h citations each | Hirsch (2005)[6] | |
| g-index | the top g papers together have at least g² citations; g ≥ h, and g gives credit to very highly cited papers | Egghe (2006)[2] | |
| m-quotient | h gained per year of academic age; lets researchers at different career stages be compared | Hirsch (2005)[6] | |
| hI,norm | the h-index after sharing each paper's citations equally among its authors; corrects for field differences in team size | Harzing, Alakangas and Adams (2014)[4] | |
| hIa | individual annual h-index: hI,norm per year of academic age; corrects for both team size and career length | Harzing, Alakangas and Adams (2014)[4] |
The g-index is capped at the number of papers n (Egghe's paper also discusses padding with uncited "fictitious" papers; ScholarGauge does not). hI,norm and hIa use only the papers matched to OpenAlex, where the number of authors is known; the card says how many. Scholar's own h-index and i10-index are shown unchanged in the citations chart below the cards.
On a Google Scholar profile the chart and the totals (all years and "since" year) are read from Scholar's own sidebar, unchanged. Scholar draws no bar for a year without citations; ScholarGauge places each bar under the nearest year label, so such a year shows 0.
For each paper matched to OpenAlex, the profile owner's position in OpenAlex's author list (byline order) is classed as sole (one author), first, last or middle. The owner is recognised by the OpenAlex author ids that match the profile name, or by the name itself. The card gives the share of first-authored papers among papers where the owner was found, and the median number of authors per paper.
OpenAlex gives each work an open-access status[11], following the categories of Piwowar et al. (2018)[15] with "diamond" added:
| Status | Meaning |
|---|---|
| diamond | in a fully open-access journal that charges authors no fee |
| gold | in a fully open-access journal |
| hybrid | free under an open licence in a subscription journal |
| green | subscription only on the publisher's site, but a free copy is in a repository |
| bronze | free to read on the publisher's site, without an identifiable open licence |
| closed | everything else |
The card shows the share of matched papers that are not closed, and the count of each kind.
search results (quoted phrases joined with AND / OR);
per_page=1 returns only the total.Report the formula, the data source and the date of the counts, and cite the method papers. Suggested wording:
"Novelty was measured with the Poisson gap index G = 100·P(X > O), X ~ Poisson(E), where O is the number of OpenAlex works matching both ideas within the context and E = n(A∧C)·n(B∧C)/n(C) is the number expected under independence (cf. van Eck and Waltman, 2009; Lee, Walsh and Wang, 2015), computed with ScholarGauge 1.1.0 (KStats Lab, 2026) from OpenAlex counts retrieved on <date>."
"Research variety was summarised by the effective number of topics eH (Hill, 1973; Jost, 2006), the Gini–Simpson index (Simpson, 1949) and Pielou's evenness (Pielou, 1966) over OpenAlex topics with fractional counting, and interdisciplinarity by Rao–Stirling diversity (Rao, 1982; Stirling, 2007) with taxonomic distances on the OpenAlex topic hierarchy."
"Citation impact was summarised by the h-index (Hirsch, 2005), the g-index (Egghe, 2006), the m-quotient h/a and the author-normalised hI,norm and hIa (Harzing, Alakangas and Adams, 2014), with academic age a counted from the year of the first paper, computed with ScholarGauge 1.1.0 from Google Scholar citation counts retrieved on <date>."
Every reference was checked against its Crossref record (DOI) or, for web pages and the arXiv preprint, against the page itself: on 28 October 2026, and the references added in version 1.1 on 13 November 2026.