GEO Brand Citation Index — Scoring Methodology

1. Overview

The index was designed to answer a question existing brand tracking tools cannot: is your brand cited by AI systems because of current web presence, or because of historical training data? By comparing a retrieval-augmented platform (Perplexity, which searches the live web before answering) against a fixed-training-data platform (ChatGPT, which does not), the delta surfaces the gap between AI memory and AI retrieval.

This document is the complete public record of all scoring decisions. Any change to the methodology is versioned and recorded in the changelog. Scores from different methodology versions are not directly comparable without explicit adjustment notes in the monthly report.

2. All 54 queries

Query selection criteria

A query is included if it meets all three of the following:

1. Real evaluation intent. The query must reflect how a user actually compares tools — open-ended, category-level questions that prompt multi-brand responses. Single-brand navigational queries (“What is Semrush?”) are excluded.
2. Produces multiple brand mentions. Queries that consistently return single-brand answers yield no comparative data. Queries that generate list-format or comparison responses are preferred because they allow position-scoring across multiple brands per response.
3. Stable month-to-month. Queries referencing time-specific events or rapidly shifting contexts are excluded. The panel tests the same intent every month. Any modification requires a versioned methodology update.

The “best alternative to X” query pattern is an intentional exception to the no-brand-naming rule. It is a high-volume, real user query that generates rich multi-brand responses, and it specifically tests whether competing brands surface when a user is actively considering leaving a market leader.

Each query is sent identically to all three platforms in the same monthly session window with no paraphrasing or platform-specific modifications. The total per platform is 18 queries. Why fix the panel? If queries change between months, score movements become uninterpretable — any change could reflect a query shift, not a brand shift. Month-over-month comparability requires holding the query panel constant.

3. Brand panel — all 28 brands

SEO & Marketing Tools (12 brands): Semrush, Ahrefs, Google Search Console, Moz, Surfer SEO, Ubersuggest, Screaming Frog, Yoast SEO, Majestic, SE Ranking, Mangools, Rank Math

CRM & Sales (8 brands): Salesforce, HubSpot, Zoho CRM, Pipedrive, Freshsales, Insightly, SugarCRM, Keap

AI & LLM Tools (8 brands): ChatGPT, Copilot, Jasper, Claude, Notion AI, Writesonic, Copy.ai, Grammarly

Brand mentions are matched by name including common variants (e.g. “Ahrefs” and “ahrefs.com” both score for Ahrefs). URL citations within a response are tracked separately and contribute the +2 bonus to whichever brand’s domain is cited.

4. Scoring formula

Position weights

Rationale for position weighting

Position-weighted scoring reflects a well-established finding in information retrieval research: items presented first receive disproportionately more attention and are more likely to influence decisions. This effect — documented as primacy bias in cognitive psychology and position bias in information retrieval — has been replicated across web search (Joachims et al., 2005; Craswell et al., 2008) and shown to apply to long-context LLM outputs (Liu et al., 2023). A brand mentioned first in an AI response has meaningfully different visibility than the same brand mentioned fifth.

The weight values (5, 3, 2, 1) reflect steep initial advantage for first mention, a meaningful but smaller reward for second, diminishing returns for subsequent mentions, and a floor of 1 for any mention beyond third. The URL citation bonus reflects an additional quality signal: a cited domain URL indicates the platform actively retrieved content from that domain — a higher-fidelity signal than a training-data reference alone.

Worked example — scoring one query response

This process runs for all 6 queries in the vertical on that platform. Raw points are summed across all 6 responses per brand. The brand with the highest total raw points becomes the normalisation reference (100). All others are scaled proportionally per Section 5.

5. Normalisation

Normalisation is performed within verticals rather than globally because different verticals have structurally different response patterns. CRM queries produce shorter brand lists dominated by two or three names. AI tools queries produce longer, more varied lists. Global normalisation would systematically disadvantage CRM brands based on their vertical’s response structure, not their actual citation frequency within category.

6. Delta calculation

Perplexity is the live web benchmark because it is a retrieval-augmented generation (RAG) system that searches the current web before generating each response (Lewis et al., 2020). ChatGPT (GPT-4o mini, browsing disabled) is the training data baseline — it generates responses from a fixed corpus with a knowledge cutoff and no live retrieval. The two platforms represent opposite ends of the training-data versus live-retrieval spectrum. Gemini is tracked as a third data point but is not used in the primary delta; it occupies a hybrid position and is most useful as a triangulation signal.

Delta interpretation reference

7. Archetype assignment

8. Platform definitions

9. March 2026 — complete results

SEO & Marketing 12 brands · 6 queries per platform · 18 total query runs

CRM & Sales 8 brands · 6 queries per platform · 18 total query runs

AI & LLM Tools 8 brands · 6 queries per platform · 18 total query runs

10. Known limitations

11. References

The methodology draws on established research in information retrieval, position bias, retrieval-augmented generation, and temporal bias in language models.

• Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., … Kiela, D. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. Advances in Neural Information Processing Systems, 33, 9459–9474. arxiv.org/abs/2005.11401 — Foundational paper on RAG systems. Informs the distinction between Perplexity (retrieval-based) and ChatGPT (fixed training data) as opposite ends of the citation spectrum.
• Joachims, T., Granka, L., Pan, B., Hembrooke, H., & Gay, G. (2005). Accurately Interpreting Clickthrough Data as Implicit Feedback. Proceedings of SIGIR ’05, 154–161. — Seminal evidence for position bias in search results. Justifies the position-weighted scoring model: items presented first receive disproportionately more user attention and influence.
• Craswell, N., Zoeter, O., Taylor, M., & Ramsey, B. (2008). An Experimental Comparison of Click Position-Bias Models. Proceedings of WSDM ’08, 87–94. — Empirical validation of position bias across different result presentation formats. Supports the 5/3/2/1 weight decay applied in this index.
• Liu, N. F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., & Liang, P. (2023). Lost in the Middle: How Language Models Use Long Contexts. Transactions of the Association for Computational Linguistics, 12, 157–173. arxiv.org/abs/2307.03172 — Documents primacy and recency bias in LLM output generation. Supports elevated weighting for 1st and 2nd position citations in AI-generated lists.
• Gao, Y., Xiong, Y., Gao, X., Jia, K., Pan, J., Bi, Y., … Wang, H. (2023). Retrieval-Augmented Generation for Large Language Models: A Survey. arXiv preprint arXiv:2312.10997. arxiv.org/abs/2312.10997 — Comprehensive survey of RAG system architectures. Provides technical context for how Perplexity’s live retrieval produces fundamentally different brand citation behaviour from ChatGPT’s static inference.
• Metzler, D., Yi, T., Mahmoudi, G., & Najork, M. (2021). Rethinking Search: Making Domain Experts out of Dilettantes. ACM SIGIR Forum, 55(1). arxiv.org/abs/2105.02274 — Early framework for AI-native search and how LLMs change information retrieval. Contextualises why different AI platforms produce structurally different brand citation patterns.
• Nakano, R., Hilton, J., Balwit, A., Wu, J., Ouyang, L., Kim, C., … Schulman, J. (2021). WebGPT: Browser-assisted Question-Answering with Human Feedback. arXiv preprint arXiv:2112.09332. arxiv.org/abs/2112.09332 — OpenAI’s foundational work on web-grounded LM responses. Background on the technical distinction between training-data citation and retrieval-based citation.
• Lazaridou, A., Gribovskaya, E., Stokowiec, W., & Grigorev, N. (2022). Internet-Augmented Language Models Through Few-Shot Prompting for Open-Domain Question Answering. arXiv preprint arXiv:2203.05115. arxiv.org/abs/2203.05115 — Demonstrates how internet-augmented LLMs produce different factual outputs from fixed-training models. Supports the index’s central premise that Perplexity and ChatGPT produce structurally different brand citation distributions.

12. Methodology changelog

Every change to the query panel, scoring weights, normalisation method, platform selection, or archetype thresholds is documented here with a version number. Scores from different methodology versions are not directly comparable without explicit adjustment notes in the monthly report.