The CRISP Methodology: How We Measure AI Answers

The CRISP Methodology

How a CRISP run is collected and scored: the five factors and their weights, the five AI engines, the collection conditions, where a company is placed, and what a score can and cannot claim.

CRISP is Third Hemisphere's proprietary technology and methodology, built upon the most cutting edge global standards and research.

The five factors

• Consensus. Do the engines agree about you? Your buyers do not all use the same engine. If only some of them name you, being found comes down to which one a buyer opens.
• Reputation. Do they speak in your favour? Being named is not being chosen. This is whether the engines put you forward when a buyer is deciding.
• Influence. Do trusted third parties back what is said about you? Engines repeat what credible third parties publish. When others carry your story, the engines tell it your way.
• Source. Can AI read and trust your own site? Engines check what they say against what they can read. A thin record, a contradicted one, or a site they cannot read counts against you.
• Presence. Do they mention you when buyers don't name you? Most buyers describe a need before they know a name. If the engines do not bring you up then, nothing else can happen.

The AI Authority Index

Third Hemisphere's AI Authority Index is the first published index to rank companies in AI answers across five engines. It is built on Third Hemisphere's proprietary CRISP technology and methodology, publishes each company's sample size and confidence interval, and ranks companies only where the difference between them is statistically significant.

Read the AI Authority Index

References, and how the method was built

CRISP was built from a review of the field it measures, begun as web research into how AI visibility and authority are measured. Third Hemisphere's index blueprints of 19 August 2026 record the review: they read the IAB's Measuring Visibility in the AI Era, reviewed the field's reference index and over 30 organisations publishing AI visibility indexes or benchmarks, and rejected each existing classification they reviewed as a comparison set. The research on generative engines, the web standards crawlers read, and the statistics of reporting a share with its uncertainty were reviewed with them. Every source the record names is listed here, with what CRISP took from it.

1. On 19 August 2026, the first blueprint read the field: a full teardown of the Semrush and Adobe AI Visibility Index 2026, a review of over 30 organisations publishing AI visibility indexes or benchmarks, sorted into five kinds, and the IAB's four layers of visibility.
2. The second and third blueprints, the same day, reviewed the classifications an index could borrow and the precedents for governing one: ANZSIC 2006, the IAB Tech Lab Content Taxonomy, the Google Product Taxonomy, GICS, the Australian energy schemes, SASB, NABERS, GRESB, the Global Reporting Initiative, and the Gartner Magic Quadrant.
3. They read the research on how far a ranking of brands from engine answers can be trusted, how a persona in the question changes the answer, how often engines agree, and whether a cited source supports the claim, and took the validation template for reading answers from a published entity-linking error taxonomy.
4. On 15 September 2026, the site audit was rebuilt from the research on generative engine optimisation. Every check in it and every piece of advice traces to a study, to the company's own measurement, or is labelled as unmeasured.
5. The research register was last reviewed on 15 September 2026 and is reviewed every 90 days, because the engines change fast enough to reorder its findings within two quarters.
6. On 22 September 2026, each of the four papers in the research register was opened and its stated finding compared with what the register says it found.
7. On 24 September 2026, the review read the industry's round-ups of answer-engine visibility tools, research on how consistently engines recommend brands, practitioners' writing on comparing scores between platforms, and a standards body's playbook for brand measurement inside AI platforms.

The review is Third Hemisphere's own. It covers the sources listed here, read on the dates given, and a source that is not listed was not part of it.

Research on generative engines and AI search

1. Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., and Deshpande, A. (2024). GEO: Generative Engine Optimization. KDD 2024. arXiv:2311.09735. https://arxiv.org/abs/2311.09735 The measured effect of content changes on how often a page is quoted, which weights the site audit's advice per market, and the finding that repeating a keyword measured as a cost.
2. Yu, J., Yang, M., Ding, Y., and Sato, H. (2026). Structural Feature Engineering for Generative Engine Optimization: How Content Structure Shapes Citation Behavior. arXiv:2603.29979. https://arxiv.org/abs/2603.29979 The finding that document structure moves citation with the content held constant, and that the shape of the whole page outweighs sentence-level polish, which sets the weight of the audit's structural checks.
3. Zhang Kai, He Xinyue, and Yao Jingang. (2026). From Citation Selection to Citation Absorption: A Measurement Framework for Generative Engine Optimization Across AI Search Platforms. arXiv:2604.25707, version 2 of 29 April 2026. https://arxiv.org/abs/2604.25707 The difference between a page an engine cites and a page whose content reaches the answer, and the four kinds of extractable evidence such pages carry: definitions, numerical facts, comparisons, and procedural instructions.
4. Schulte, J., Bleeker, M., and Kaufmann, P. (2026). Don't Measure Once: Measuring Visibility in AI Search (GEO). arXiv:2604.07585. https://arxiv.org/abs/2604.07585 The finding that answers vary across runs, questions, and time, so visibility is a distribution: the reason every share carries an interval and a score is measured again.
5. Chen, M., Wang, X., Chen, K., and Koudas, N. (2025). Generative Engine Optimization: How to Dominate AI Search. arXiv:2509.08919. Preprint, submitted 10 September 2025. https://arxiv.org/abs/2509.08919 The finding that AI search shows a systematic and overwhelming bias towards earned media over brand-owned and social content, unlike Google's more balanced mix, and that engines differ in the domains they draw on: the Blueprint's case for earned coverage in the sources the engines cite.
6. Sharma, M., Tong, M., Korbak, T., Duvenaud, D., Askell, A., Bowman, S., Durmus, E., Hatfield-Dodds, Z., Johnston, S., Kravec, S., Maxwell, T., McCandlish, S., Ndousse, K., Rausch, O., Schiefer, N., Yan, D., Zhang, M., and Perez, E. (2024). Towards Understanding Sycophancy in Language Models. ICLR 2024. https://proceedings.iclr.cc/paper_files/paper/2024/hash/0105f7972202c1d4fb817da9f21a9663-Abstract-Conference.html Documented sycophancy in language models, recorded as an explanation to rule out before uniformly favourable answers are read as a finding.
7. Bardol, F. (2025). ChatGPT Reads Your Tone and Responds Accordingly -- Until It Does Not -- Emotional Framing Induces Bias in LLM Outputs. arXiv:2507.21083. https://arxiv.org/abs/2507.21083 A documented tone floor, where negative framing rebounds to neutral even when an engine is asked directly, recorded as an explanation to rule out when engines rarely speak against a company.
8. Żatuchin, D. (2026). Where Does the Noise Come From? A Variance-Components Decomposition of Non-Determinism in LLM Brand Answers. arXiv:2607.13304. Preprint, not peer-reviewed, submitted 14 July 2026. https://arxiv.org/abs/2607.13304 The low reliability of ranking brands from engine answers, which is why every share carries an interval and an order is claimed only where the intervals separate.
9. Jack, W., Lehman, N., Maloney, K., and Xu, S. (2026). Prominence-Stratified Failure Modes in Retrieval-Augmented Commercial Recommendation: A 37,000-Run Audit. arXiv:2605.27439. Preprint, submitted 22 May 2026. https://arxiv.org/abs/2605.27439 Evidence that long-tail and regional brands often never surface in any run, reviewed for what the index can say about small companies.
10. Wu, Y., Zhong, S., Kim, Y., and Xiong, C. (2025). What Generative Search Engines Like and How to Optimize Web Content Cooperatively. arXiv:2510.11438. Preprint, submitted 13 October 2025. https://arxiv.org/abs/2510.11438 An academic benchmark that learns what generative engines prefer in a page, reviewed beside GEO-bench as a check on the site audit's advice.
11. Tannenbaum, B. (2026). From Prompt to Recommendation: A Fitted Stage Model of Brand Visibility in AI Search. arXiv:2609.23162. Preprint, not peer-reviewed, submitted 19 September 2026. https://arxiv.org/abs/2609.23162 Evidence that a brand is named far more often when its own domain is in the retrieval path, which is why Source reads whether the company's own site can be found and quoted.
12. Kumar, P. (2026). Generative Engine Optimization at Scale: Measuring Brand Visibility Across AI Search Engines. arXiv:2606.20065. Preprint, submitted 18 June 2026. A vendor study. https://arxiv.org/abs/2606.20065 The finding that ranked best-of listicles are the most-cited content format, reviewed for the advice on which third-party pages carry a company.
13. Jack, W., Lehman, N., Maloney, K., and Xu, S. (2026). Persona Conditioning of Brand Recommendations in Retrieval-Augmented Commercial Chat: A Prominence-Stratified Cross-Provider Audit. arXiv:2605.30207. Preprint, submitted 28 May 2026. https://arxiv.org/abs/2605.30207 Evidence that who is asking changes the answer, which is why every question is written in one buyer's voice.
14. Żatuchin, D. (2026). Who Owns the AI Recommendation? A Multi-Industry Empirical Map of Brand Category Ownership Across Large Language Models. arXiv:2606.23057, version 2 of 23 September 2026. Preprint; its author discloses a commercial affiliation. https://arxiv.org/abs/2606.23057 A measured rate of full agreement between engines on a category's top brand, reviewed for Consensus.
15. Onweller, H., Lumer, E., Huber, A., Ramchandani, P., Subbiah, V. K., and Feld, C. (2026). Cited but Not Verified: Parsing and Evaluating Source Attribution in LLM Deep Research Agents. arXiv:2605.06635. Preprint, submitted 7 May 2026. https://arxiv.org/abs/2605.06635 Evidence that a valid link can point to a source that does not support the claim, reviewed for how CRISP reads what an answer cites.
16. Braşoveanu, A., Rizzo, G., Kuntschik, P., Weichselbraun, A., and Nixon, L. J. B. (2018). Framing Named Entity Linking Error Types. Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association. Anthology ID L18-1040. https://aclanthology.org/L18-1040/ The validation template for reading answers: two independent readers, their agreement measured, and a third reviewer for what they disagree on.
17. Netzer, O., Feldman, R., Goldenberg, J., and Fresko, M. (2012). Mine Your Own Business: Market-Structure Surveillance Through Text Mining. Marketing Science, 31(3), 521-543. https://doi.org/10.1287/mksc.1120.0713 The peer-reviewed lineage for deriving market structure from which brands are named together, behind Authority Sets.
18. Koskie, T. (2026): University of Sydney research on Australian media in AI-generated news summaries, 434 Copilot responses. Australian Journalism Review, DOI 10.1386/ajr_00183_1. Published 27 January 2026. Evidence that Australian sources are under-represented in AI answers, reviewed for the index's Australian scope.

Industry reviews and standards for AI visibility measurement

1. Search Engine Land (2026): a round-up of answer-engine visibility platforms. Read 24 September 2026. The size and shape of the tool category: industry round-ups through 2026 review between 19 and 24 answer-engine visibility platforms.
2. tryprofound.com (2026): a round-up of answer-engine visibility platforms. Read 24 September 2026. A second count of the same category, read beside the first so the size of the category does not rest on one round-up.
3. SparkToro: research on how inconsistently AI engines recommend brands. Read 24 September 2026. Evidence that engines recommend brands inconsistently, so a single reading is a sample of one: the reason CRISP asks many questions of every engine in its roster and prints an interval.
4. authoritytech.io: on why AI visibility scores cannot be compared between platforms. Read 24 September 2026. Practitioners' observation that each platform defines its own score, so scores cannot be compared between platforms: the reason CRISP fixes one instrument, one set of engines, and one scoring profile for every company in an edition.
5. IAB: a playbook for brand measurement inside AI platforms, as reported by Marketing Dive. Read 24 September 2026. Evidence that a standards body has moved to bring order to brand measurement inside AI platforms, which CRISP meets with a published instrument and a sample and an interval on every figure.
6. IAB (Interactive Advertising Bureau). (2026). Measuring Visibility in the AI Era. Published 3 August 2026. https://www.iab.com/guidelines/measuring-visibility-in-the-ai-era/ Four layers of AI visibility: Presence, Prominence, Portrayal, and Persuasion. CRISP measures past Presence into how engines describe and recommend a company.
7. Semrush and Adobe. (2026). AI Visibility Index 2026. Published 26 June 2026. ai-visibility-index.semrush.com. Reviewed in full as the field's reference index. CRISP publishes its instrument, fixes its engines, folds every spelling of a company into one, and prints an interval on every figure.
8. Digital Nomads HQ. (2026). State of AI Search for Australian SMEs 2026. Captured 11 to 12 May 2026. digitalnomadshq.com.au/state-of-ai-search. An Australian measurement of AI answers in search results that prints a Wilson interval on every share, the closest precedent for how CRISP reports.
9. Somantra (2026): quarterly analysis of Australian general insurance in AI answers, 2,437,107 citation records across 28,725 domains over seven months. Published 17 August 2026. Evidence that a source's persistence across months outweighs how often it is cited. Its published figures moved between editions, so none is quoted.
10. Mi3, AdNews, Campaign Brief, and company announcements: the Australian competitive scan, 1 July to 19 August 2026. Confirmation that no cross-sector, brand-level Australian AI visibility index launched in that window.
11. Muck Rack. (2025). What Is AI Reading? December 2025. A vendor study of the sources ChatGPT, Claude, and Gemini cite, released 3 December 2025. https://martechseries.com/predictive-ai/ai-platforms-machine-learning/earned-media-still-drives-generative-ai-citations-as-press-release-visibility-grows/ The Blueprint's case for earned media: in the citations it studied, earned media was the large majority and paid placement a small minority, and the journalists most pitched barely overlap with those the engines cite for a brand. Its figures are quoted only beside this citation.

Precedents for classification, governance, and independence

1. ISSB. (2024). Background paper, agenda reference 6A, May 2024, on why SASB built its own industry classification; and the SASB industry list of 10 October 2025. The precedent for building a classification into the instrument, which is why CRISP's market taxonomy is its own.
2. NABERS. nabers.gov.au. The precedent for accredited assessors, verifiable primary data, and a protected mark, reviewed for how the index is governed.
3. GRESB. gresb.com. The precedent for separating who governs a standard from who sells the assessment, the closest one for a firm that publishes the index and sells the audit.
4. Global Reporting Initiative: standards free to download. The precedent for free access making a classification the default.
5. Gartner Magic Quadrant, and ZL Technologies v Gartner (2009). The precedents that the market definition and the inclusion rules are the asset, so they are published, and that a ranking framed as assessed opinion survived appeal.
6. Australian Bureau of Statistics: the 2022 ANZSIC public consultation, which ran nine weeks and drew 62 submissions. The precedent for a public consultation before a taxonomy is called consulted.
7. Digital News Report Australia 2026, University of Canberra with Reuters Institute. Published 16 June 2026. The precedent for a university as co-author of a published measure.

Web standards the site audit reads

1. Schema.org: the structured data vocabulary, read as JSON-LD. https://schema.org The structured data the audit reads to tell whose page it is, including the sameAs links that resolve a name to one company.

Statistical method

1. Wilson, E. B. (1927). Probable inference, the law of succession, and statistical inference. Journal of the American Statistical Association, 22(158), 209-212. The interval on every share CRISP prints, which stays inside the scale near 0 and 100.
2. Newcombe, R. G. (1998). Interval estimation for the difference between independent proportions: comparison of eleven methods. Statistics in Medicine, 17(8), 873-890. The interval on a difference of two shares, which is what Reputation weighs.
3. Cohen, J. (1960). A coefficient of agreement for nominal scales. Educational and Psychological Measurement, 20(1), 37-46. The chance-corrected agreement between the two independent readings of every answer, which must reproduce before an edition publishes.

Classification and population data

1. Australian Bureau of Statistics. Counts of Australian Businesses, including Entries and Exits. Data at 30 June 2025. The count of actively trading Australian businesses by employment size, which sets the population an engine can plausibly name.
2. Australian Bureau of Statistics. ANZSIC 2006, the official Australian industry classification. abs.gov.au. The industry divisions each CRISP industry is tied to, with every departure recorded. Reviewed and rejected as a comparison set, because it has no class for whole categories of today's sellers.
3. IAB Tech Lab. Content Taxonomy 3.1. github.com/InteractiveAdvertisingBureau/Taxonomies. Reviewed and rejected as a comparison set: it classifies what a web page is about, so it has no concept of a supplier or a competitive set.
4. Google Product Taxonomy. Reviewed and rejected as a comparison set: it classifies products, and a company that designs, sells, and installs has no place in it.
5. GICS: 11 sectors, 25 industry groups, 74 industries, and 163 sub-industries. Reviewed and rejected as a comparison set: it gives too few classes to the sellers the index measures.
6. Australian energy scheme classifications: Victorian Energy Upgrades (esc.vic.gov.au), the NSW Energy Savings Scheme, the Clean Energy Council products program, and NABERS ratings. Reviewed and rejected for a national index: each classifies an activity, a method, a product, or a building, and none classifies a business by the job a customer hired it to do.

• Free AI Authority Check
• AI visibility audit
• GEO audit
• GEO vs SEO
• Check your AI visibility
• Improve your AI visibility
• AI Authority Blueprint: AUD 700
• The AI Authority Index
• The CRISP methodology
• AI Authority & Visibility