SAVI Report · Vol. 3 · Q3 2026
When someone asks ChatGPT, Gemini or Google's AI to recommend a business, it reads websites and puts a few names forward. We have analysed more than a million of them. 216 are ready to be picked. None are unambiguous enough to be picked without hesitation, and the highest score anyone has recorded is 86.6.
An average tells you where the middle of the web sits. A ceiling tells you what is achievable with the way sites are built today, and this one falls short of the bar. The gap between 86.6 and 90 does not look large. It is the difference between a site an AI can cite and a site an AI cannot misread.
Every site gets a score from 0 to 100 for how ready it is to be found and recommended by AI. To reach the top AI-Ready band you need 80. Just over half of all sites land in Emerging, the 40 to 59 band where a site is legible to a machine but gives it no reason to choose that site over any other.
A third of the web sits below 40 and is rarely usable by an AI at all. When someone asks an AI to recommend a business, it names two or three, not ten. When almost every site looks the same to a machine, the few that stand out win far more than their share.
The score is built from eight weighted areas. Sites do well on the plumbing an AI needs to reach them and on letting the crawlers in. They fail on everything an AI needs to trust and cite them: the machine-readable facts, the external corroboration, the evidence that a qualified human wrote the page.
Note which category carries the most weight. E-E-A-T Content is 24% of the score, the heaviest of the eight, and it averages 38.5. And note what is not the top weight: AI Citability scores 70.4 because its components are largely permissions a site already grants. Doing well on the category named after citation does not make a site citable.
The same thing keeps happening at every scale we measure. Sites are reachable, the crawlers are let in, the page loads properly, and then nobody adds the small labels that let an AI work out who published this and why it should be believed.
Only 7.8% of sites disallow GPTBot, and the picture is similar for the other major crawlers. Permission is close to universal. But 54.3% emit no JSON-LD whatsoever, 67.6% declare no Organisation schema, and 72.4% publish no llms.txt. The crawler arrives, is welcomed, and finds prose written for a human skim-reader with no structure attached. It leaves with nothing it can safely quote.
Not one site in the 20,000-site sample is missing HTTPS. That is a standard adopted to saturation, invisible because it works. Against it: 96.6% publish no FAQ schema, 91.3% no Person schema, 86.1% no IndexNow, 83.9% no author bio. Technical SEO had two decades and a clear feedback loop, fix it and watch the ranking move. AI citability has neither. Nothing on an analytics dashboard says that ChatGPT read your page and could not work out who wrote it.
Structured Data averages 26.7 out of 100 and Brand Authority 23.7, the weakest of the eight areas. Almost no site publishes the facts an AI needs to lift, or the external corroboration it weighs before repeating them.
Organisation schema is about fifteen lines of JSON. An author bio is a paragraph and a link. The reason they are missing is not cost or difficulty. It is that nobody has told most site owners these signals are now load-bearing.
The Verified band at 90 and above has been open since July and nothing has reached it. 216 sites have done enough to be selected. None have done enough to be selected without hesitation. That is the whole opportunity, stated as a number.
The most important change in this edition is not a finding. It is the instrument. Averaging a growing corpus each quarter does not measure the web, it measures who joined the pool. When the dataset expands into the long tail the average falls, and none of that fall is a real decline.
From this edition the index runs on a fixed panel of 102,873 domains, frozen in July 2026 and stratified by tier so that weighted aggregates reconstruct the whole population. The same domains are re-audited every wave, on the same method and the same scorer version. Wave 1, dated 1 August 2026, is the baseline at a weighted mean GEO of 45.24. It has nothing to be compared against yet, which is what a baseline is for.
Those readings were produced on an earlier scoring engine, which ran a different set of checks and scored roughly three points more generously at the top of the range, and on a corpus that more than doubled between the two editions. Both effects move an average without anything changing on the web. Publishing the comparison anyway would have produced a headline about the average rising, which would be an artefact in both directions. We have reset the series here instead. The AI-Ready figure, 216 of 1,000,000+, is measured consistently and is the number to track. Earlier editions keep their published figures.
Ordered by how much score they move against how long they take. The first four are a working week. The fifth is the one that takes months, which is why it matters most.
216 of the more than 1,000,000 websites analysed score 80 or above on the GEO scale, which is 0.022%, or about 1 in 4,600. No website has reached the Verified tier at 90 or above, and the highest score recorded anywhere is 86.6.
Verified sits at GEO 90 and above: the band where a site has not merely satisfied the signals but left nothing for a model to guess at. It is a strict subset of AI-Ready, so it changes no counts. As of Q3 2026 it is empty; the ceiling recorded across the whole corpus is 86.6.
No. Only 7.8% of websites disallow GPTBot in robots.txt, and the picture is similar for the other major AI crawlers. Access is close to universal. The binding constraint is that 54.3% publish no structured data at all, so the crawler is welcomed in and finds nothing it can safely quote.
Brand Authority, at an average of 23.7 out of 100, followed by Structured Data at 26.7. Technical foundations average 70.8 by contrast. E-E-A-T Content carries the heaviest weight in the index at 24% and averages 38.5, so the category that counts most is also the one the web is worst at.
Those readings were produced on an earlier scoring engine and on a corpus that more than doubled between them. Both effects move an average without anything changing on the web, so a like-for-like comparison would measure the method rather than the market. From Q3 2026 the index runs on a fixed panel of 102,873 domains with a wave 1 baseline of 45.24 weighted mean GEO.
Check your inbox in the next few minutes for the full report. If it is not there, take a look in your spam folder.
How we did this. We check each live website the same way an AI reader would, looking at more than 250 things that affect whether AI can find, understand and recommend it, and turn that into a score out of 100. The headline figures cover more than 1,000,000 websites analysed to Q3 2026, entering the corpus organically through the free SearchScore audit rather than by curation. Category averages and signal rates come from a 20,000-site sample of the most recent audits scored on the current engine. The index baseline comes from a fixed panel of 102,873 domains, frozen in July 2026, stratified by tier and re-audited each wave on the same method and scorer version. This edition does not compare its average against Q1 or Q2: those readings used an earlier scoring engine on a corpus that more than doubled between them, so the difference would not reflect any change on the web. Earlier editions keep their published figures.