The benchmark for AI search visibility across the web.
A quarterly index measuring how visible the world's websites are to ChatGPT, Perplexity, Claude and Google AI Overviews – published in the SearchScore SAVI Report.
Down from 41.4 in Q1
Cannot be reliably cited
Less than 1 in 500
And growing daily
A single benchmark for the state of AI search visibility.
Each site in the SearchScore dataset is given a GEO Score from 0 to 100, based on 130+ AI search visibility signals across 8 weighted categories. SAVI aggregates those GEO Scores into a single index value plus a set of headline statistics for each reporting period:
- The average GEO Score across the dataset
- The tier distribution – how many sites are Invisible, Low, Emerging, Strong, or AI-Ready
- The average score in each of the 8 SAVI categories
- The size of the gap between technical health and AI visibility
- Notable outliers – well-known brands scoring poorly, small businesses outperforming household names
SAVI sits alongside the SEO Score and CRO Score that SearchScore measures for each site. Together they form a single audit. SAVI specifically benchmarks the AI visibility layer – the GEO layer – at industry scale.
130+ signals. Eight weighted categories. One index.
Each site's GEO Score is built from 130+ signals across the categories below. SAVI aggregates the same way. The weights are calibrated against citation behaviour observed across ChatGPT, Perplexity, Claude and Google AI Overviews in the SearchScore benchmark.
| Weight | Category | What it measures |
|---|---|---|
| 25% | AI Citability | How directly content answers AI queries – quotable statistics, answer-first content, structured citations. |
| 20% | Brand Authority | External entity signals: Wikipedia, LinkedIn, social presence, third-party mentions. |
| 20% | Content Quality | E-E-A-T signals: author bios, bylines, contact info, sourced claims. |
| 15% | Technical Foundations | Crawlability, HTTPS, sitemaps, canonical tags. |
| 12% | AI Platform Readiness | IndexNow, Bing verification, Perplexity and ChatGPT crawler access. |
| 10% | On-Page Structure | JSON-LD schema markup – Organisation, Service, Article. |
| 10% | Topical Authority | Content hub depth, internal linking, structured headings. |
| 8% | User Experience | OpenGraph, Twitter Cards, RSS, video and multi-platform reach. |
The real web. No curation. Recomputed every edition.
Each reporting period, SAVI is computed from the GEO Scores of every site in the SearchScore benchmark at that point in time. No sites are excluded for being too small or too large. No curation is applied. The dataset is the real web as submitted to SearchScore by users running free audits.
This methodology produces a deliberately tough number. The Q2 2026 SAVI sits at 34/100, down from 41.4 in Q1 – not because the web got worse, but because the dataset grew from 350,000 to 850,000 sites, reaching deeper into the long tail where most AI-readiness work has never been done.
SAVI is recomputed every edition. The methodology is fixed. Year-on-year comparisons will be valid from Vol. 1 (Q1 2026) onwards.
Five tiers from invisible to AI-Ready.
Every site in the benchmark sits in one of five tiers based on its GEO Score. The tier distribution is one of SAVI's headline outputs each edition.
| Tier | GEO Score | What it means |
|---|---|---|
| AI-Ready | 81 – 100 | AI engines reach for this site first. <1% of the dataset. |
| Strong | 61 – 80 | Cited reliably for relevant queries. |
| Emerging | 41 – 60 | Cited occasionally; structurally improvable. |
| Low Visibility | 21 – 40 | AI engines rarely surface this site. |
| Invisible | 0 – 20 | Functionally absent from AI search. |
Across the Q2 2026 dataset, 74.2% of sites sit in Invisible or Low Visibility. Only 0.2% are AI-Ready.
Built by Ronnie Huss. First published on HackerNoon.
The SAVI methodology was developed by Ronnie Huss, founder of SearchScore, and first introduced in HackerNoon, 2026.
Each SAVI Report edition uses this citation format:
Press and researchers citing SAVI itself – not a specific report edition – should cite this page: