AI citation indices rank the same publishers differently because they measure different things: what counts as a citation, which engines are sampled, which questions are asked, and how each cited domain is labelled. A controlled test of seven tracking tools on the same domain over 15 days found an 8.2x gap between the lowest and highest citation count. Source-role labelling adds a fifth layer: in the Machine Relations Index, the share of citations credited to editorial publications moved between 7.1% and 10.5% on the same data depending only on the classification rule.
None of these numbers is wrong. Each answers a different question, and a buyer who reads one without its definition is reading a number without a unit.
Last updated: October 15, 2026
What the major AI citation indices actually measure #
At least five distinct measurement systems publish AI citation data. Each defines the object of measurement differently.
Baden Bower AI Visibility Index (launched June 2026) tracked 12,040 citations across six engines: ChatGPT, Perplexity, Claude, Gemini, Google AI Overviews and Microsoft Copilot. It runs 20 buyer-intent questions, each repeated 10 times per engine (1,200 observations), and produces an AIO Citation Score and a citation share percentage. Forbes leads at 92 points and 13.8% share; Business Insider follows at 10.5%.
Foglift AI Search Citation Benchmark (Q2 2026) used 75 brand-neutral buyer-intent prompts across 25 verticals, generating 375 responses across five engines (ChatGPT, Claude, Gemini, Google AI Overview, Perplexity). It found 2,697 cited URLs from 1,119 distinct domains, with a cross-engine Jaccard similarity of 0.18: any two engines share less than one-fifth of their cited domains for the same queries.
Machine Relations Index (MRI) reports source-segment citation rates. A segment is one subject category paired with one buyer-question shape (best X, how to choose, is X worth it, problem first, top list, X vs Y). For each segment, the rate is the number of monitored answer runs that cited a domain divided by all observed runs in that segment. A segment publishes a rate only after it clears an evidence floor of 10 observed runs across 7 distinct run dates; thinner segments are marked collecting rather than scored. Each domain carries a confidence grade (A, B, C or collecting), a source role, the basis for that role, its engine breadth and its rank within its source role. The release of 2026-10-15 (methodology mri_score_v2.1) covers 31,063 cited domains and 193,069 source events from 25,826 monitored runs of 1,227 prompts on six engines (ChatGPT, Claude, Gemini, Perplexity, Google AI Mode and Google AI Overviews), observed from 2026-05-10 to 2026-10-15. Of its 157 segments, 117 publish a rate. The MRI glossary entry states the contract; the public file carries every figure.
EPR AI Labs Citation Share Index (Everything-PR) uses a five-factor framework: citation frequency (40%), cross-engine breadth (20%), query-type breadth (20%), extractability (15%) and crawl access (5%). Its September 2026 update reports third-party benchmark data and still lists EPR's own five-engine prompt-set data drop as scheduled.
BrightEdge AI Search Insights publishes weekly citation volatility reports on how AI search citations shift over time, focusing on temporal instability rather than static rankings. Platform surfaces have entered the same market: Ahrefs Brand Radar separates AI Overview, AI assistant and web visibility monitoring, and Otterly's AI Citations Report publishes citation-share analysis from a large prompt corpus. They widen the category without removing definition drift.
Why the same domain gets a different score from every index #
The kenimoto.dev comparison identified four structural causes for the divergence across seven tracking tools. Index data adds a fifth.
Definition variance is the largest factor #
Tools do not agree on what counts as a citation. Profound ($499/mo) counts only clickable source links. Peec AI (€89/mo) counts any brand mention, linked or unlinked. Otterly AI ($29/mo) counts cited URLs with daily deduplication. Bluefish AI calculates a share-of-voice rank. Semrush's AI Toolkit counts domain URLs in structured answer fields only. A DIY Python script tracking raw answer text found 54 citations where Otterly found 38 and Peec AI found 312, for the same domain, queries and period.
| Tool | Citations found (same domain, 15 days) | Multiple vs. baseline |
|---|---|---|
| Otterly AI | 38 | 1.0x |
| Self-built Python | 54 | 1.4x |
| Semrush AI Toolkit | 71 | 1.9x |
| Bluefish AI | 89 | 2.3x |
| Profound | 147 | 3.9x |
| Scrunch | 203 | 5.3x |
| Peec AI | 312 | 8.2x |
The Machine Relations Research review of 16 platforms' public measurement contracts found the same problem one level up: denominators, sampling and deduplication rules are often not stated at all.
Engine coverage varies #
Peec AI samples five engines (ChatGPT, Claude, Gemini, Perplexity, Copilot). Baden Bower samples six including Copilot; the MRI samples six including Google AI Mode and AI Overviews but not Copilot. The surfaces are not uniform: Google documents AI features as Search surfaces that link to supporting web pages, Microsoft describes Copilot Search as a cited search experience, Perplexity exposes ranked search results through its Search API, and Anthropic documents web search citations for Claude. Because each engine has distinct preferences (Foglift found Perplexity cites YouTube in 41.3% of responses while ChatGPT and Claude cite it in 0%), changing the engine set changes every domain's total and relative rank. The MRI's release manifest publishes each engine's observed days and runs, so a reader can see when one engine's coverage of the window is thinner than another's.
Sampling frequency and deduplication create drift #
Otterly deduplicates within 24-hour windows. Peec AI counts each separate mention across its measurement period. A weekly tracker that deduplicates by URL reports a fraction of what a daily tracker counting text mentions reports. The MRI counts a domain once per monitored run, and excludes from the denominator any engine-day on which that engine returned no citation at all (81 engine-days and 948 runs in the 2026-10-15 release, each listed in the file), so runs that could not have cited anything do not depress a rate.
Query construction and segmentation decide what gets found #
Baden Bower uses 20 buyer-intent questions repeated 10 times. Foglift uses 75 brand-neutral prompts across 25 verticals. The MRI uses 1,227 prompts across 24 subject categories and six question shapes, and reports rates per segment rather than one global number. Segmentation matters as much as sample size. In the 2026-10-15 release Forbes is the most-cited of 1,027 classified editorial publications, cited in 1,025 of 25,826 monitored runs (3.97%) on all six engines with confidence grade A. Across the 72 published segments where it is cited at all, its rate runs from 0.45% (1 of 224 runs, consumer products, is X worth it) to 35.0% (77 of 220 runs, consumer finance, best X). An index that reports Forbes as one number is averaging across a 78-fold spread.
Source-role labelling changes class shares #
Indices that summarise citations by source class (editorial publications, vendor sites, databases, community platforms) depend on how each domain gets its label. The MRI records the basis for every label: a reviewed decision, the domain's own site evidence, a registry suffix such as .gov, or a word in the domain name. On the 2026-10-14 release, editorial publications held 10.5% of domain citations when a word in the domain name could assign the label and 7.1% when only review or site evidence could, and five of 20 categories changed their leading source class with the rule. No citation moved; only the label did.
From the 2026-10-15 release, the MRI no longer labels a domain an editorial publication on a name pattern alone. Editorial publications hold 9.0% of the release's 170,065 domain citations and vendor-owned sites 10.3%; 64.6% go to domains with no specific role, reported as Other observed source. Name-pattern labels remain only for market and company databases (755 domains, 2.0% of citations).
Where the indices agree #
Despite the divergence in counts, several findings hold across measurement systems.
Forbes is among the most-cited publications in every system that ranks publications. Baden Bower ranks Forbes first (92 score, 13.8% citation share). Foglift finds Forbes leads in Gemini (21.3%) and Google AI Overview (20%). The MRI ranks Forbes fourth of 31,063 cited domains and first among editorial publications.
Cross-engine overlap is low. Foglift's Jaccard similarity of 0.18 means engines do not cite the same sources. Of the 81 unique domains in any engine's top 25, only healthline.com appeared across all five engines. An index that tracks fewer than four engines cannot see this divergence.
Recency matters, but not uniformly. Baden Bower found coverage published within 12 months was cited 2.9 times more often than older content. The MRI reports how many observed days each domain was cited (Forbes: 130 of 152), a stability signal rather than a recency filter.
Editorial coverage outperforms distribution. Baden Bower's data shows authored editorial content cited 2.3 times more often than press releases. In the 2026-10-15 MRI release, editorial publications hold 9.0% of domain citations against 0.4% for wire and press-release distribution.
Which measurement answers which decision #
The right method depends on the decision the measurement informs. The examples name who publishes each kind of data, not a recommendation among them.
| Decision | Question the data must answer | Method that answers it | Published examples |
|---|---|---|---|
| Media placement priorities | Which publications are cited for buyer queries? | Publication-level citation share or per-domain rate on buyer-intent prompts | Baden Bower; MRI category pages |
| Own visibility over time | Is our domain cited more or less than last month? | One fixed definition tracked daily or weekly | Otterly; a custom script |
| Category source landscape | Which source types dominate answers in our category? | Segment rates with a published source-role basis | MRI; Foglift by vertical |
| Content strategy | Which content types and structures get cited? | Structural classification of cited pages | Foglift |
| Brand presence in answer text | Is our name in the answer, linked or not? | Mention counting, not source citation | Peec AI and other mention trackers |
No single index answers every question. The reference on AI search visibility measurement tools separates Google Search Console AI reports, citation indices and third-party monitoring platforms into the three layers operators combine.
Five tests for any AI citation measure #
- Definition. Does it say whether it counts linked sources, text mentions or share-of-voice positions? The 8.2x gap exists because tools define a citation differently, and most do not say precisely how.
- Engine coverage. Which engines, and for how much of the window? Six-engine designs (Baden Bower, MRI) cover more of the retrieval surface than five-engine designs, but they are not the same six.
- Denominator and segment. Is the number a share of all runs, of runs in one category and question shape, or of runs that cited anything? A rate without its denominator cannot be compared.
- Evidence and time. Does it publish how many runs and dates stand behind a figure, and withhold a figure when there are too few? A single-day snapshot, a daily trend and an evidence-floored rate answer different questions.
- Source-role basis. If it reports shares by source type, does it publish each domain's label and the reason for it? Without the basis, a difference of three or four points between two studies can be entirely a difference of taxonomy.
FAQ #
Why do AI citation tracking tools disagree so much? #
Mostly because they define a citation differently: some count only clickable links, others any text mention. A comparative test found an 8.2x gap between tools for the same domain. Engine coverage, deduplication, question segmentation and source-role labelling compound the difference.
Which AI citation index is most accurate? #
None is accurate in general, because each measures a different quantity. The useful question is whether an index publishes its definition, denominator, engine set, evidence counts and labelling rules, so its figures can be checked and compared over time.
What is source role in AI citation measurement? #
Source role is the label an index gives a cited domain, such as editorial publication, vendor-owned site, market database or community platform. It does not change how often a domain is cited; it changes how citations are summed into classes. In the Machine Relations Index the label rule moved the editorial share between 7.1% and 10.5% on the same data, so a source-type share is only comparable when the rule behind it is published.
How does the Machine Relations Index measure citations? #
It divides the monitored answer runs that cited a domain by all observed runs in the same subject category and question shape, and publishes the rate only once a segment has 10 observed runs across 7 run dates. The 2026-10-15 release covers 31,063 domains across 25,826 runs on six engines, with a confidence grade, source role and role basis for each domain. It assigns no points and does not combine signals into a composite.
How many engines should an AI citation index track? #
At least four. Foglift's cross-engine Jaccard similarity of 0.18 means engines share less than one-fifth of cited domains for the same queries, so omitting a major engine creates systematic blind spots in any ranking.
Updated October 15, 2026: the Machine Relations Index figures and method description now follow release mri_score_v2.1+2026-10-15; the June edition described the v1 composite retired on 2026-07-05 (release manifest).