For press and newsletters
One paragraph you can quote
A random sample of 1,000 homepages drawn from a public AI-tools directory was scored with 40 published checks on 2026-11-04: 620 returned a normal page and were scored, 287 never answered, and 93 answered with a non-200 status. The median score was 53/100, 367 of the 620 scored sites graded F and only 2 graded A; 48 of 620 disallow at least one AI crawler in robots.txt; 382 of 620 carry no Organization entity markup.
A separate hand-picked sample of 30 major sites (3 October 2026) found that all six news publishers in it disallow at least one tracked AI crawler - that figure belongs to the 30-site study, not to the 1,000-site benchmark.
Quoting is welcome with attribution to LLMention and a link to this page; the full CSV and JSON are below. Corrections - especially a domain that is no longer reachable - go to services@llmention-geo.com.
What we found
The median AI-tool homepage scores 53 out of 100. The distribution is not a bell: 2 sites earned an A and 367 earned an F, which is what a category looks like when nobody is checking the machine-readable layer of their own marketing page.
48 of the 620 scored sites do not admit at least one of GPTBot, ClaudeBot, PerplexityBot, OAI-SearchBot or Google-Extended at their root. That is a deliberate policy in robots.txt in most cases — a rule copied from a template, or a CDN default, or a decision to keep training data out — but the effect is the same: the engine that would cite the tool cannot read it. The 48 domains are listed here, with their scores.
382 of 620 scored homepages carry no Organization entity markup and 459 do not declare a single sameAs profile. A model reading those pages has no structured statement of who the company is, which is the difference between a page that can be quoted and a page that can be attributed.
The most common first failure is the least glamorous one. No Content-Signal is the first failing check on 297 homepages, no sitemap line in robots.txt on 139, and no robots.txt at all on 97. None of those is a content problem; all three are a five-minute fix.
What this sample is, and what it is not
The starting list is a public directory of AI tools. 1000 of its entries were drawn at random with a fixed seed (20261104) and deduplicated by hostname. It is not "the 1,000 best AI tools": it is a random sample of one directory, and directories over-represent tools that launched recently and under-represent the ones that never listed themselves.
Each row is the homepage only, fetched once, from one network location, by the same 40-check analyser used in the free scan. A homepage is the hardest page on a site to score well: it is short, navigational, and carries little of the evidence an answer needs. A low score here does not mean the site is invisible in AI answers; it means the page an engine lands on first gives it little to quote.
The results
Sorted by score. The ten strongest and ten weakest are listed first; the full 620 scored rows follow, then the 380 that could not be scored. Every domain links to a live scan of the same page. The strongest 50 are also on their own page.
The full ranking
620 scored homepages, highest first. This is the complete ranking, not a sample of it; the per-row flags (AI crawler access, entity markup, sameAs) are in the CSV and the JSON, so the full table stays readable instead of repeating six columns a thousand times.
What could not be scored
380 domains were sampled and could not be scored: 287 never answered a request on either scheme, 93 answered with a non-200 status, and 0 produced an error the scanner could not classify. They are listed rather than dropped, because a table that quietly deletes its missing rows is a table that overstates its sample. The first 200 are shown here; the JSON endpoint has all of them.
Method, in full
Sample: 1000 entries drawn without replacement from AIToolsList (public AI-tools directory) (https://github.com/lakey009/AI-Tools-List), seed 20261104, deduplicated by hostname; social networks, marketplaces and marketplace-hosted pages were removed before sampling because their homepage is a platform rather than the tool. One homepage per domain, fetched with an honest LLMentionBot user-agent, robots.txt read first but the homepage fetched regardless (the audit is about whether the page can be read at all, not whether the scanner obeys its own subject), one request per support file, one run per domain, on 2026-11-04.
Scored sites are those whose homepage returned HTTP 200 after redirects. A site that refuses a crawler-shaped request while serving a browser-shaped one is reported through the access checks rather than failed outright. The forty checks and their weights are published on the methodology page; the raw CSV is at /study/benchmark-2026.csv and the same data as JSON at /study/benchmark-2026/data/.
Limits a reader should hold on to: it is one directory's population, not the population of AI tools; one scan per site is one observation, and bot-management answers can differ between two requests seconds apart; the scanner sees the response its own address receives, not the response a real GPTBot would receive; and the score describes a homepage's machine-readability, not the quality, usefulness or popularity of the product on it. If a row about your domain is wrong, the fix is at /contact/ with the URL and the check id, and corrections are published with the next run.