AI is now the layer between your website and the people looking for it. They ask ChatGPT, Perplexity or Google's AI to compare options, and increasingly they send an agent to do the clicking. So we asked a simple question with a lot riding on it: is the live web actually ready for that, or is it still built purely for human eyes?
To find out we ran two separate analyses using our own free scanner tools, then looked at the aggregate. One measures whether AI can answer the questions your buyers ask about your site. The other measures whether an AI agent can act on your site at all. We looked at close to 10,000 websites. The web is failing both tests, in two very different ways.
The headline numbers:
97% of sites expose zero tools an AI agent can use. The agentic web is a near-empty field.AI scores 43 out of 100 answering "what does it cost", across more than 8,100 sites. Commercial facts are where it fails.Sites cluster at the extremes. You are either legible to AI or invisible to it, with little in between.
What we measured #
Two independent datasets, both from real websites submitted to our public tools between February and July 2026:
Answerability(9,846 unique websites, 52,158 graded answers). We trained a temporary AI assistant on each site, asked the questions a real buyer asks, and had a model grade every answer from 0 to 100 against what the site actually contained.Agent-readiness(216 websites). We scored each site 0 to 100 across four weighted categories: content discoverability (30%), AI agent tools (30%), content quality (25%) and technical readiness (15%).
Scores are model-assigned and are not a certification. Where an average could be dragged down by sites our crawler could barely read, we say so and exclude them.
Finding 01
The agentic web does not exist yet #
Across the agent-readiness scans, the average site scored 56 out of 100 overall, and only about 7% scored 75 or higher. But the overall number buries the real story, which lives entirely in one category.
Average category score across the agent-readiness scans
The average score for AI agent tools was 5 out of 100. That category checks whether a site exposes anything an AI agent can actually use: a WebMCP tool registration (the emerging in-browser standard, navigator.modelContext
), a live MCP endpoint at a discoverable path, or documented API access. 97.4% of sites had none of it. An agent that lands on them can read the page, but it cannot book, look up an order, check stock, or take a single action.
Now the contrast: the same sites averaged 98 out of 100 on technical basics like HTTPS, mobile rendering and speed. The web has diligently finished the last decade's checklist and has not written a line of the next one.
| Category | Avg score | Sites failing (under 50) |
|---|---|---|
| AI agent tools | 5 | 97.4% |
| Content quality | 65 | 18.8% |
| Content discoverability | 77 | 16.2% |
| Technical basics | 98 | 1.1% |
Being agent-actionable is not yet a baseline, it is a near-empty field. The first sites in any category to let an agent complete a task, not just read about one, are the ones agents will route their users to.
Finding 02
AI can describe you, but not price you #
The second dataset is bigger and hits closer to home. Across 9,846 sites and 52,158 graded answers, AI answered buyer questions at a mean of 67 out of 100 on sites it could read. Respectable, until you sort the questions by how well AI actually answered them.
| Buyer question | Avg answer score | Sites tested |
|---|---|---|
| Cancellation or refund terms | 33 | 1,340 |
| How do I subscribe / is there a newsletter | 38 | 3,300 |
| How much does it cost | 43 | 8,100+ |
| What integrations or APIs does it support | 45 | 375 |
| How often is new content published | 46 | 207 |
| Main benefits versus other products | 51 | 2,768 |
The pattern is brutal and consistent. The questions AI answers worst are the commercial ones: what it costs (43 out of 100, across more than 8,100 sites), whether you can cancel or refund (33), and how to actually sign up (38). These are the exact facts a buyer needs to say yes.
And the questions it answered best were the soft ones: how it works, step by step (83), who it is for (82), and what makes it different (81).
Read those two lists together. AI has completely absorbed your marketing story and cannot reliably state your price, your refund policy, or how to buy. Ask ChatGPT "how much is it and can I get a refund," and for roughly 8 in 10 sites the honest answer is a shrug.
Finding 03
You are either legible to AI, or you are not #
When we plotted every answerability score, the shape was not a bell curve. It was two walls. Roughly 37% of scans scored below 50 and 40% scored 75 or higher, with a thin valley between and a large cluster near zero: sites where AI, given the actual page content, could extract almost nothing useful.
| Score band | Sites |
|---|---|
| 0 to 9 | 1,735 |
| 10 to 39 | 1,649 |
| 40 to 59 | 1,384 |
| 60 to 79 | 2,670 |
| 80 to 100 | 4,078 |
There is very little "sort of readable." A site is either structured so AI can lift clean answers out of it, or it is a wall of design and JavaScript that AI slides off. Which side you land on is a property of how your content is built, not how good your product is.
The two gaps are separate problems #
It is tempting to assume these are the same failing. They are not. A site can score 85 on answerability, with clean crawlable content, and still score 0 on agent tools because it exposes no way for an agent to act. Getting AI to understand you and letting an agent use you are two different pieces of work. Most sites have done neither. A rare few have done the first. Almost none have done the second.
Why this happens #
None of this is a mystery once you look at how the pages are built. Pricing lives inside a JavaScript table that renders after load, so a crawler sees an empty container. Refund terms sit in a PDF, or a legal page nothing links to. There is no llms.txt
telling AI what the site is and where the important pages are, no structured data, and certainly no tool an agent could call. The information exists, for a human willing to click around. It is invisible to the AI now doing that clicking on their behalf.
How to fix it #
The good news is that all of this is fixable, and most of it is not hard. In rough order of effort versus payoff:
Put the commercial facts in plain, crawlable text. Pricing, refund and cancellation terms, and how to get started, as real HTML on the page, not trapped in a widget, an image, or a PDF.Publish an It tells AI what your site is and points it at the pages that matter. Generate one free with ourllms.txt
.llms.txt generator.Make your site agent-actionable. This is the 97% gap almost nobody has closed. ASiteSpeakAIchatbot with public MCP enabled registers WebMCP tools automatically, so an agent can look up your pricing, book, or hand off to a human instead of hitting a dead end.Check where you actually stand. Run the freeAI Score scannerthat produced this data and see your own site's numbers.Track whether it is working. Getting AI-readable is step one. Knowing whether AI assistants now mention and cite you is step two, which is what our sister toolMentionScoutmeasures across ChatGPT, Gemini, Perplexity and more.
What this means #
Every platform shift rewards the sites that move first. Mobile-friendly sites won the last one while everyone else waited. This one is bigger, because AI is not just reading your site, it is standing between you and your customer and deciding what to say on your behalf. Right now it is saying very little about your price and nothing to the agents showing up to buy. The sites that fix that, that become legible and then actionable, are the ones AI will send people to. The window where that is a differentiator instead of a baseline is open, and this data says it is wide open.
Method. Aggregates from SiteSpeakAI's public scanner tools, February to July 2026. Answerability: 9,846 unique sites across 52,158 model-graded answers (0 to 100); the reported mean excludes failed crawls scoring under 10, and the distribution is shown across all 11,516 scans. Agent-readiness: 216 unique sites (266 completed scans), weighted across content discoverability (30%), AI agent tools (30%), content quality (25%) and technical readiness (15%). Per-question scores are directional; the ranking is the finding. Scores are model-assigned and not a certification. A plain-text copy for AI crawlers is at /research/ai-readiness-report-2026.md.