Veterinary (Taiwan) — AI Score methodology and version history
What this score measures, how it is computed, and what has changed.
1. What this score measures
How much AI crawlers visited this entity's IDA embassy page over the last 30 days.
2. The three axes
3. Parameters and how they were calibrated
| Parameter | Value | Calibration basis |
|---|---|---|
| Rolling window | 30 天 | Rolling, so earlier launches do not accumulate an unfair lead. |
| Per-bot daily cap | 30 | Daily visits measured at p99≈14 and max 750; 30 sits near p99.7, trimming full-site batch crawls. |
| Breadth cap | 30 | Indexing crawlers reach nearly 100% of pages, so breadth is an entry ticket, not a score. |
| Depth knee | 100 | Site-wide population P99 is 25 visits per 30 days; 4× that. The old value of 30 was cleared by a single batch crawl. |
| Activity full marks | 14 天 | Half the window, separating sustained interest from a one-off sweep. |
| Tenure full marks | 6 個月 | A qualifying month scores ≥ 55 on current performance. These 6 months are why a top score takes time. |
| Retrieval crawlers | OAI-SearchBot, PerplexityBot, bingbot, Googlebot | Signal of entering an answer engine's candidate pool; weighted above indexing crawlers. |
| Human referrals | ChatGPT-User, Claude-User, Perplexity-User | A human opened this page from inside an AI. Rare today, but ultimately the thing worth measuring. |
4. Tiers
| Tier | Meaning | Score band |
|---|---|---|
| Not detected | No visits in 30 days. Shown as “—”, never as zero. | — |
| Initial exposure | Picked up by indexing crawlers, around the population median of 3 visits. | 20–35 |
| Stable indexing | Several indexing crawlers return regularly; above the population P90 of 10 visits. | 35–50 |
| In the candidate pool | Retrieval crawlers visit, so answer engines can see the page (only 1.9% of the population). | 50–65 |
| Used by answer engines | Retrieval crawlers visit often and in volume. | 65–80 |
| Long-term authority | The above sustained for six months or more. No shortcut — only time. | 80–100 |
5. Honest disclosures
- Early baseline: crawler records span only 29 days so far, with no seasonal signal. Thresholds will be recalibrated as data accumulates.
- AI crawlers arrive in batches rather than continuously; a single batch entering or leaving the window moves the score. The persistence axis exists to dampen this.
- Human-referral signals remain very rare (only 8 of 39,547 pages have any), so they add points rather than dominate the score.
- Traffic generated by our own tooling is excluded at collection time and never counts.
6. Version history
Three axes: Breadth 30 + Depth 40 + Persistence 30. Time now counts: the 20-point tenure component requires six qualifying months, so 90 cannot be reached within half a year.
Changes- Added a persistence axis worth 30 (10 in-window activity + 20 tenure).
- Breadth is now tiered: indexing/training crawlers weigh less; retrieval and human-referral agents weigh more.
- Depth knee raised from 30 to 100 visits, so a single batch crawl no longer maxes the axis.
- Measured on 610 Taiwan dental clinics: median 74.6 → 43.9; top score 95.1 → 77.6.
- Scores ≥90 went from 31 to 0; scores ≥80 from 221 to 0.
- Median rank shift of 81 places. The ruler changed, not the clinics — previous scores stay visible for one month.
Depth switched to visits capped at 30 per bot per day (winsorized) so a full-site batch crawl cannot max the axis; breadth still uses raw counts.
Measured impact- Top score 100 → 95.1; perfect scores 8 → 0.
Breadth cap lowered from 60 to 40 with search-engine crawlers at half weight; depth raised from 40 to 60 (overflow segment 4 → 24).
Breadth became 10 points per crawler kind (half if it visited only once), replacing the old “distinct crawlers ÷ 6”. Traffic from our own tooling was excluded at collection time.
The two-segment depth curve was recalibrated against measured data: knee 20 → 30 visits, ceiling ln481 → ln971 (saturating near 1,000).
Depth became a two-segment log curve (36 + 4 for the overflow), so very high visit counts no longer tie with mid-range ones.
First release of the AI Score: breadth 60 (distinct crawler kinds ÷ 6) + depth 40 (log of visit count, capped at ln21), with the 30-day rolling window introduced here.
7. FAQ
- I changed nothing — why did my score move?
- A version change swaps the ruler; it is not a decline. Everyone's score moves at once, so the relative position is what matters. Previous scores stay visible for a period.
- Why is nobody at 100?
- Top marks are reserved for pages that are continuously retrieved by answer engines, opened by real people via AI, and sustain that for over six months. A score that saturates on day one carries no information.
- How often is this recalibrated?
- A full percentile re-run at 90 and 180 days of data. If the population distribution shifts materially, thresholds change and a new version is announced on this page.