Version 1.0.0 · effective 2026-09-01
The Greenbug AI Recommendation Score methodology
Published in full, including the parts that make the score harder to earn and the things it cannot tell you. This page renders from the same constants the scoring engine uses, so the published methodology and the computed score cannot drift apart.
1. Governing principle
The Recommendation Score measures observed AI outcomes only.
Completing technical SEO, adding schema, fixing citations or publishing content does not raise the Recommendation Score. Those are inputs we control, and they are measured separately by the Entity Authority and AI Readiness diagnostics.
The practical consequence is that a month of genuine work can end with a flat score. That is reported as a flat score.
2. Formula and weights
Recommendation Score =
(Presence Rate × 0.40)
+ (Top-1 Recommendation Rate × 0.20)
+ (Top-3 Recommendation Rate × 0.10)
+ (Citation Rate × 0.15)
+ (Accuracy Rate × 0.10)
+ (Positive Recommendation Rate × 0.05)
× 100
Clamped to 0–100 and rounded to one decimal place. Weights sum to exactly 1 and are identical for every business measured.
| Component | Weight |
|---|---|
| Presence Rate | 40% |
| Top-1 Recommendation Rate | 20% |
| Top-3 Recommendation Rate | 10% |
| Citation Rate | 15% |
| Accuracy Rate | 10% |
| Positive Recommendation Rate | 5% |
| Total | 100% |
3. Component definitions and denominators
Denominators are where scoring methods quietly differ from one another. Ours are fixed and stated.
- Presence Rate
- Of all observations in the period, the share in which the business was mentioned at all.
- Top-1 Recommendation Rate
- Of all observations, the share in which the business was the first business recommended.
- Top-3 Recommendation Rate
- Of all observations, the share in which the business appeared in the first three businesses recommended.
- Citation Rate
- Of all observations, the share in which the AI answer cited a source the business controls or is named on.
- Accuracy Rate
- Of mentions that contained checkable factual claims, the share in which those claims were correct.
- Positive Recommendation Rate
- Of mentions, the share framed positively or strongly positively rather than neutrally or negatively.
Presence, Top-1, Top-3 and Citation all divide by all observations in the period. An answer that never named the business is a loss, not an exclusion.
Citation is recorded independently of mention. An AI can cite a business’s own page while recommending somebody else, and that is a distinct and diagnostically useful state rather than a contradiction.
Accuracy divides by mentions that contained checkable factual claims. An answer that names a business without asserting anything about it cannot be right or wrong, and is excluded from both numerator and denominator. The denominator is disclosed on every report rather than hidden inside the percentage.
Positive Recommendation Rate divides by all mentions. A neutral mention counts in the denominator and not the numerator — being named is not the same as being recommended.
4. Score bands
| Range | Band | Interpretation |
|---|---|---|
| 0–19 | Minimal | Rarely appears in tested AI recommendations. |
| 20–39 | Emerging | Appears inconsistently. |
| 40–59 | Established | Meaningful visibility but not dominant. |
| 60–74 | Strong | Frequently recommended. |
| 75–89 | Leading | One of the strongest businesses tested. |
| 90–100 | Dominant | Exceptional recommendation visibility across platforms and query types. |
Bands are published with integer bounds but scores carry one decimal. A band owns everything from its minimum up to the next band’s minimum, so 39.9 is Emerging.
5. Testing methodology
24 fixed commercial-intent prompts × 5 platforms × 2 runs = 240 observations per business per cycle.
Every prompt runs more than once because AI answers are non-deterministic; a single answer is an anecdote, and a single answer moving is not a trend.
Benchmark platforms
- ChatGPT
- Gemini
- Perplexity
- Claude
- Google AI Mode
Prompt intent distribution
Every client library uses this identical distribution, so libraries are comparable across businesses and industries. A library weighted towards brand-validation prompts would flatter an established name; a fixed distribution prevents that.
| Intent | Prompts |
|---|---|
| Discovery | 3 |
| Direct Recommendation | 3 |
| Local | 3 |
| Trust | 2 |
| Comparison | 2 |
| Problem / Need | 3 |
| Expertise | 2 |
| Value | 2 |
| Brand Validation | 2 |
| Alternatives | 2 |
| Total | 24 |
Testing context
Tests run from a consistent unauthenticated context with an explicitly stated geography. Location is never inherited from the device of whoever is running the test — a benchmark that moved because an operator was working from a different city would be worthless.
6. Prompt freezing
Once written, benchmark prompt text is never changed. Changing prompts between cycles would destroy historical comparability — and would let an agency manufacture improvement simply by asking easier questions.
This is enforced in the database by a trigger rather than left to policy: an update to a benchmark prompt’s text raises an exception. To retire a prompt we deactivate it and create a new prompt code; both remain in the record.
Exploratory prompts live in a separate experimental set and never contribute to a published score. Quarterly expansions may add prompt volume, competitors, cities, services or secondary platforms — the fixed benchmark stays unchanged so the historical series stays valid.
7. Benchmark coverage and the Platform Score
The Greenbug AI Recommendation Score requires all 5 benchmark platforms. A measurement covering fewer yields a Platform Score — a real, defensible measurement of one platform’s behaviour, computed with the same weights and denominators, which is not the Greenbug AI Recommendation Score and must never be presented as one.
This is structural rather than a matter of wording: a Platform Score is a distinct type in the codebase, and the database refuses to publish a partial benchmark as the official score. Every surface showing a partial measurement displays platform-by-platform coverage.
8. Supporting diagnostic scores
Two diagnostics run alongside the Recommendation Score. Neither is ever folded into it. They exist so that work we control can be tracked honestly without contaminating the measurement of what AI actually did.
Entity Authority Score
How consistently and authoritatively the business exists across the sources an AI system could draw on.
| Item | Weight |
|---|---|
| Business identity consistency (NAP) | 15% |
| Authoritative third-party references | 15% |
| Local citations | 12% |
| Industry citations | 10% |
| Media mentions | 8% |
| Professional directories | 8% |
| Reviews and reputation | 12% |
| Structured business information | 7% |
| Topical authority | 8% |
| Geographic authority | 5% |
AI Readiness Score
How technically and structurally prepared the business’s own site is to be found, read and understood.
| Item | Weight |
|---|---|
| Crawlability | 7% |
| Indexability | 7% |
| robots.txt | 5% |
| Sitemap | 5% |
| AI crawler access | 9% |
| Schema / structured data | 9% |
| Service page structure | 8% |
| Location page structure | 8% |
| FAQ content | 6% |
| Internal linking | 5% |
| Entity clarity | 7% |
| Author / expert information | 5% |
| Original proof | 6% |
| Case studies | 5% |
| Reviews / testimonials on site | 4% |
| Page speed / technical health | 4% |
9. What is recorded for every observation
This is the audit trail. Greenbug must always be able to show how a score was calculated and which answers produced it.
- Client
- Test date
- Platform
- Prompt ID
- Prompt code
- Exact prompt text
- Geography
- Intent
- Repetition
- Mentioned
- Position
- Top-1
- Top-3
- Cited
- Accurate
- Recommendation strength
- Competitors named
- Source domains
- Source URLs
- Full response capture
- Snapshot URL
- Tester
- Notes
- Methodology version
10. Limitations
Published because a methodology that does not state its limits is marketing, not method.
- AI systems are non-deterministic. Two identical prompts can return different businesses minutes apart, which is why every prompt is run twice on every platform and why single-observation movements are never reported as trend.
- Results are sampled, not exhaustive. A 240-observation monthly sample estimates recommendation behaviour; it does not enumerate every answer a consumer could receive.
- Platform personalisation, account history and location inference can change answers. Tests are run from a consistent unauthenticated context with an explicit stated geography, which is not identical to any one consumer's context.
- Model updates shipped by platform vendors can move scores independently of any work performed. Known platform changes are logged in the measurement record for the period.
- Citation detection depends on the platform surfacing sources. Platforms that cite sparsely will show lower citation rates for every business tested, client and competitor alike.
- Greenbug does not control any AI model. We improve the information and authority signals models can draw on; we cannot guarantee any specific recommendation outcome.
11. Conflicts policy
- Editorial rankings and index placement cannot be purchased, and being a Greenbug client cannot raise a score or a ranking.
- Scores are computed from recorded observations by a fixed formula. No manual adjustment is applied to a client score.
- Where Greenbug has a commercial relationship with a business appearing in public research, that relationship is disclosed on the page.
- Greenbug.ai measures itself using the same methodology and publishes its own score, including when it is unflattering.
- Any sponsored placement, if ever introduced, must be visually and structurally separated from editorial rankings and labelled as sponsored.
12. Version history
Every stored score records the methodology version it was computed under. Historical scores are not retroactively rewritten when the methodology changes; both versions are shown.
| Version | Effective | Change |
|---|---|---|
| 1.0.0 | 2026-09-01 | Initial published methodology: six weighted components, 24 frozen benchmark prompts, five platforms, two repetitions, six score bands. |
Reviewing this and want the underlying detail? The canonical document lives in the repository as METHODOLOGY.md, and this page renders from the same constants the scoring engine imports. If you find a discrepancy between them, it is a bug and we want to hear about it — tell us.
Measured the same way, published the same way
If you want your business assessed under this methodology, the baseline is where it starts.