The GEO score is not a trade secret, it is methodological discipline

The myth of the trade secret

Several GEO agencies on the market communicate around a “proprietary score” presented as a black box. This gives an impression of technical complexity justifying the price. But in practice, this opacity creates three problems:

• The client does not know exactly what is measured, so cannot verify improvement.
• The agency cannot be challenged on methodological choices (and potentially makes structural errors that go undetected).
• The promise becomes unverifiable, which opens the door to commercial drifts.

PROEMA’s bet is the opposite: publish the score’s complete composition, explain weighting choices, and place IP not in the formula but in quarterly execution discipline.

The 12 indicators of the PROEMA score

Our overall GEO score is a weighted aggregate of 12 structural indicators, measured on the brand site and its environment.

#	Indicator	Weight	Measurement method
1	Schema.org @graph multi-types (≥5 types per hub page)	10%	Crawl + JSON-LD parsing
2	FAQPage Schema on eligible pages	8%	Crawl + editorial audit
3	Author Person + sameAs Wikidata	12%	Crawl + Wikidata validation
4	llms.txt present and compliant	6%	HTTP GET + format validation
5	Explicit dateModified on ≥80% of pages	5%	Crawl + parsing
6	Editorial density (words / page, H2-H3 structure)	7%	Manual editorial audit
7	Trilingual or more with valid hreflang	5%	Crawl + hreflang validation
8	Tier-1 press presence (≥3 verifiable mentions)	10%	Manual research
9	Wikidata brand entity and/or executive	8%	SPARQL Wikidata query
10	Direct Perplexity citation (rate on 50 target prompts)	10%	Manual measurement 6 LLMs
11	Direct ChatGPT/Claude/Gemini citation (same)	10%	Manual measurement
12	Absence of brand hallucinations (factual error rate)	9%	Manual qualitative audit

Maximum score: 100 points. Each indicator is measured on a 0-100 scale, then weighted. The result is a simple number, comparable quarter to quarter, comparable across brands.

Where is PROEMA’s IP then?

If the formula is public, what justifies the price and added value? Four things:

1. Target prompt curation. To measure indicators 10-11-12, a list of 20-50 target prompts per brand is needed. This list is built from market listening (Reddit, Quora, trade communities), competitive watch, and customer journey understanding. It is expert work taking 8-15 days per diagnostic. List quality determines score relevance.
2. Quarterly discipline. Measuring the score once is easy. Measuring it every quarter, with same rigor, on same prompts, tracking LLM evolution (ChatGPT changes, Perplexity evolves, Claude updates), that is what distinguishes usable measurement from cosmetic measurement.
3. Editorial arbitration. When a brand’s score drops, you must decide which project to prioritize. This arbitration rests on ten months of GEO Rocket portfolio observation, knowledge of LLM adoption dynamics, and mastery of editorial levers. It is senior expertise, not algorithmic.
4. Authority ecosystem. PROEMA can activate press push via Viviane Eeman (30 years press), knows francophone Wikipedia/Wikidata contributors, has relays on Reddit/Quora. These relays do not set up in 3 months, it is a relational asset.

Why publish openly

Three reasons why openness serves us:

• Credibility. A brand that measures can be challenged. A brand that measures and publishes its indicators gains authority. It is exactly the E-E-A-T mechanism LLMs value.
• Differentiation by transparency. On a market where most players cultivate opacity, publishing the grid becomes a positioning. PROEMA = the transparent GEO agency.
• Healthy competition. If other agencies adopt the grid, all the better, the market gains standardization. And our real IP (curation, discipline, ecosystem) remains defensible.

What we keep non-public

Three elements remain confidential:

• Per-client target prompt lists. These are the brand’s competitive angles. Publishing = giving these angles to competitors.
• Per-vertical specific weightings. The general score uses displayed weights. But by vertical (B2B, hospitality, F&B, tech), we adjust weightings. These adjustments are proprietary.
• Detailed comparative benchmarks. When we measure a brand’s score vs its 5 direct competitors, nominative detail stays with the client. Only aggregated averages are published (e.g., BE Barometer 2026).

The 12-indicator grid above is stable since PROEMA origin (October 2025). Any future evolution will be published here, with date and justification of change. A brand’s PROEMA GEO score measured in Q1 2026 will be comparable to the score measured in Q4 2027, same grid, same method, same rigor.

What this changes for a client

Concretely, a PROEMA client receives each quarter:

1. Their aggregate global score (out of 100), with evolution vs previous quarter.
2. Detail per indicator (each out of 100), with what moved.
3. Benchmark vs 3 direct competitors (named or anonymized per brief).
4. Prioritized action plan for the next quarter, based on indicators where marginal ROI is highest.

This transparency transforms the relationship: it is not an agency that “does GEO” and presents opaque results, it is a quarterly measurement discipline whose every movement the client understands. Trust builds at that place.

For curious brands

If you want to test the grid on your brand without commitment, PROEMA’s free diagnostic (30 minutes, by email) uses a simplified version of the grid (6 main indicators). No score, but a qualitative reading. Email [email protected].

May 2026 update, the grid against external benchmarks

The PROEMA method has been tested against three external benchmarks published in Q1 2026. (1) The Princeton paper “GEO: Generative Engine Optimization” (Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan, Deshpande, arXiv:2311.09735, KDD 2024) remains the academic standard: it measures visibility lift up to 40% in generative-engine responses through nine optimisation strategies that overlap heavily with the PROEMA grid indicators. (2) The 5W Citation Source Audit Q1 2026 (synthesising 9 datasets from Similarweb, SEMrush, Profound, Peec AI, Ahrefs, Evertune, SE Ranking, Goodie, Passionfruit) confirms that LLM-citation concentration plays out on a couple of dozen highly authoritative domains, what the grid measures through the “press + Wikipedia + sameAs presence” axis.

(3) The July 2025 arXiv study by Kai-Cheng Yang (366,000 citations across 65,000 responses from Perplexity / OpenAI / Google AI Mode) analyses cross-LLM citation dynamics and validates three mechanisms the grid captures explicitly: long-form editorial coverage, identifiable human signature, coherent knowledge graph. None of the three benchmarks contradicts the PROEMA grid. Two methodological adjustments are being integrated into version 2.1 of the grid: (a) a “Reddit collapse resilience” sub-indicator (cf. September 2025 shift), (b) a “LinkedIn presence” sub-indicator (14.3% of ChatGPT Search citations). Sources: arXiv:2311.09735 (Princeton GEO), prnewswire.com (5W Citation Source Audit Q1 2026), arXiv (Yang 2025), semrush.com/blog/most-cited-domains-ai/.

Methodological deep-dive, benchmarking the PROEMA grid against 2026 academic references

The PROEMA GEO scoring methodology is a crossover between three public grids and portfolio feedback. The three grids: Princeton GEO (arXiv:2311.09735, KDD 2024), Kai-Cheng Yang 2025 (arXiv:2507.04881, analysis of 366 000 citations on 65 000 LLM responses), and 5W Citation Source Audit Q1 2026 (consolidation of 9 datasets Similarweb / SEMrush / Profound / Peec AI). Portfolio feedback comes from the three GEO Rocket sites (expertcafe.be, zeroproof.one, expertvin.be) and from the first PROEMA client audits.

“The GEO score is not a trade secret. It is a methodological discipline that becomes a commercial advantage when applied with rigor, at scale, without shortcuts.”

GEO Opportunity Score (SOG) v2, the formula

The PROEMA formula applied to a GEO audit is: SOG = (Intent × Citations × (1 − Brand Coverage) × Authority Multiplier × Strategic Fit) ÷ Effort. Each factor is measurable. Intent is query qualification (informational / transactional / navigational) weighted by estimated volume. Citations is the current number of LLM citations on the query, measured on 4 engines minimum. Brand Coverage is the percentage of citations already mentioning the client brand. Authority Multiplier reflects editorial robustness (Schema, Wikidata, press). Strategic Fit is the 1-5 strategic scoring (is the query strategic for the brand?). Effort is the estimated lift difficulty (1 = trivial, 10 = full project).

Comparison with the Princeton GEO grid

The Princeton paper (Aggarwal et al., KDD 2024) identifies 7 factors: citation sources, statistics-based content, quotation-based content, fluency optimization, easy-to-understand, authoritative content, technical terms. Four of these seven factors are directly incorporated in PROEMA SOG v2. The other three (fluency, easy-to-understand, technical terms) are editorial factors that PROEMA handles in the content production phase, not in scoring. Combined Princeton + SOG coverage reaches 9 factors, vs 5 to 6 in observed low-cost agency grids.

The 10 pattern axes

Beyond the formula, PROEMA measures 10 pattern axes characterizing a brand’s invisibility or visibility: 1. sectoral coverage, 2. nominative coverage (“brand X” queries), 3. advisory coverage (“best Y”), 4. comparative coverage (“X vs Y”), 5. transactional coverage (“buy X”), 6. place coverage (“X in Brussels”), 7. temporal coverage (“X 2026”), 8. procedural coverage (“how to do X”), 9. institutional coverage (“X and regulation”), 10. competitive coverage (comparative presence).

Double presentation, the PROEMA commercial standard

Each audit delivers two views: absolute view (X queries lost out of Y measured, business value associated if quantifiable) and 0-100 rating view (normalized score allowing cross-comparison within the client portfolio). The absolute view triggers commercial action (“you lose 142 queries/month on qualified intents”). The 0-100 view structures quarterly steering. Neither view alone is sufficient. The two together create the common language PROEMA + client.