The AI Score Algorithm

The AI Score is the grade for one probe: one prompt, sent to one engine, in one region. Every other number on the panel is an average of these.

Scope

One probe

Range

0–100

Graded by

Gemini 2.5 Flash

Rubric chosen by

Prompt intent

There are two rubrics, and the prompt’s intent decides which one applies. A prompt that never names your brand measures earned visibility. A prompt that names it can’t — so it measures the verdict the answer reaches instead.

Discovery rubric

Applies to category and scenario prompts, plus any prompt with no recorded intent.

AI Score = Position (0–85) + Sentiment (0–15) + Frequency bonus (0–8) Clipped to [0, 100].

Mention check

Before anything is scored, the grader decides whether the brand appears at all — by name, alias, URL, or a pronoun clearly referring to it. If it does not, every score for that probe is 0 and no verdict is recorded.

The check is the model’s judgement, not a string match. That is deliberate: a regex can’t tell your brand from a same-named company in a different industry, can’t tell a brand name from the same word used ordinarily (“Medium”, “Apple”), and can’t tell a mention in the answer from a hit inside an echoed search URL. All three errors overstate visibility, which is the worst direction to be wrong in.

Position Score (0–85)

Discovery prompts only. Branded prompts score 0 here.

Position answers: how much of the answer is actually about you? The grader first decides whether the answer is an ordered list (ranked or numbered options) or unordered narrative prose, then scores accordingly.

Ordered answers — scored by rank:

Placement	Score
Sole item in the list	85
First	80
Middle	70
Last	65

Unordered answers — scored by role:

Role	Score
Primary — central subject, appears early, described at length	75
Secondary — one of several options offered	65
Side mention — a brief example in passing	50

On top of that, clear emphasis lifts the score: if the brand is strongly recommended, highlighted, or reappears in the intro or conclusion, an ordered placement may rise to 82–85 and a primary role to 78–82.

The panel displays position rescaled to a 0–100 axis (raw ÷ 85 × 100) so it reads on the same scale as the other cards.

Sentiment Score (0–15)

Applies to every intent. Sentiment answers: when the AI talks about you, how does it sound?

Tone	Score
Very positive	15
Positive	10
Neutral	7
Slightly negative	4
Negative	0

Two things worth knowing:

• Only mentioned probes count toward the panel average. An answer that never brings your brand up has no tone toward it, so it is excluded rather than counted as a zero. Otherwise low visibility would masquerade as bad reputation.
• The panel shows sentiment rescaled to 0–100 (raw ÷ 15 × 100).

Mention Frequency Bonus (0–8)

Discovery prompts only. Being referenced repeatedly, or across several sections of an answer, signals the model treats you as a load-bearing option rather than filler.

Occurrences	Bonus
3+ mentions, or appears in multiple sections/categories	6–8
2 mentions	3–5
1 mention	0

The bonus is folded into the AI Score and is not displayed as its own card.

Branded rubric

Applies to comparison and trust prompts — the ones whose text already contains your brand name. Mention is guaranteed, so it earns nothing. Position and frequency are set to 0 and the score is derived from a discrete verdict.

Comparison

The question is who the answer picks when you’re placed next to a named rival.

Verdict	Meaning	Band
win	Clearly recommends or favors you over the alternatives	80–100
tie	Genuinely balanced — “depends on the use case”, or you win some scenarios and a rival wins others	45–60
lose	Favors a competitor, or is dismissive about you	0–40

Within lose, the grader lands near 0 when the brand is dismissed or judged clearly worse, and around 30–40 when it is simply the weaker option. If the brand turns out not to be discussed at all despite being named in the prompt, the verdict is lose and the score falls in 0–20.

Trust

The question is whether the answer treats you as legitimate.

Verdict	Meaning	Band
legit	Endorsed as legitimate, trustworthy, positive	80–100 (65–80 with notable caveats)
mixed	Neutral or mixed — some concerns, or inconclusive	45–60
scam	Flagged as a scam or illegitimate, or advises against you	0–40

Sentiment is still scored on branded prompts, on the same 0–15 scale, and still feeds the Sentiment card. Only position and frequency are dropped.

Why branded prompts aren’t just “extra mentions”

If comparison and trust prompts were scored by the discovery rubric, every one of them would count as a mention with a good position — you’re in the prompt, so you’re in the answer. Mention Rate would drift toward 100% and stop measuring anything. Splitting the rubrics keeps discovery honest and turns branded prompts into their own signal: a win rate and a reputation verdict.

Validation

The grader returns JSON. Nothing it returns is trusted unchecked:

Field	Enforced range	On a bad value
position_score	0–100	Clamped
sentiment_score	0–15	Clamped
frequency_bonus	0–8	Clamped
final_score	0–100	Clamped
mention_count	≥ 0, integer	Rounded and floored at 0

An unparseable response is not scored as 0 — it is retried, and if retries are exhausted the probe is flagged not analyzed and excluded from every metric. See averages & denominators.

What else the grading pass produces

The same analysis call that scores the answer also extracts:

• Competitors — every business named in the answer that competes with you, with the reason it came up. These are resolved to parent brands before aggregation; see competitors.
• Recommendations — specific, actionable suggestions for improving visibility on that prompt.
• Citation categories — each cited URL classified as a third-party blog, a brand website, or a community discussion. See citations & sources.

Free-text fields come back in the same language as the answer, so a Doubao answer in Chinese yields Chinese reasoning rather than a translation.

Reliability guardrails

A run is a few hundred network calls to third-party services, so the pipeline is built to fail loudly rather than hang:

Guardrail	Value
Retries per probe	Up to 3, exponential backoff
Query call timeout	120 s
Analysis call timeout	90 s
Web-scrape task budget	180 s
Total budget for one probe	300 s, including all retries

Retries are best-effort within the budget: once a probe passes its deadline no new attempt starts. A retry re-runs only that probe’s query and analysis — it never re-runs the batch, and it never consumes extra allowance.

Final Score → How these per-probe scores become the headline number. Averages & denominators → Which probes each panel average actually counts.