Last reviewed 19 October 2026.
Getting crawled by an AI system is the easy half. The harder question is whether there is anything on your page a model can lift cleanly and attribute. We scored 243 of the world’s most-visited websites on exactly that. The median score was 45 out of 100.
Nearly two-thirds of the top of the web is structurally difficult for an AI to quote, not because the content is bad, but because of how it is built.
How many top websites can AI quote cleanly?
Fig. 1Extractability bands across 243 scored homepages
3.3%
| Band | Sites | Share | Meaning |
|---|---|---|---|
| Strong (80+) | 8 | 3.3% | Liftable answer, clear boundaries, attributable |
| Workable (55–79) | 78 | 32.1% | Extractable in parts, one or two fixes short |
| Weak (below 55) | 157 | 64.6% | A model would struggle to quote without distorting |
Only eight sites out of 243 scored above 80. These are among the most resourced websites in existence.
Where do top websites fail?
Fig. 2Pass rate by check, n = 243
| Check | Weight | Pass rate |
|---|---|---|
| Author and dates declared | 10 | 5.8% |
| Question-form headings | 10 | 16.9% |
| Direct answer near the top | 20 | 20.6% |
| Cites two or more external sources | 10 | 44% |
| Lists or tables present | 15 | 44.9% |
| Exactly one H1 | 10 | 46.1% |
| Structured data present | 10 | 53.1% |
| Three or more H2 sections | 15 | 58.4% |
| No paragraph over 120 words | 10 | 98.8% |
Three findings worth sitting with
1. The median opening paragraph is 18 words
Against a target of 25 to 90: long enough to contain an answer, short enough to lift whole. Sixty-nine sites had no substantial opening paragraph at all. This is the highest-weighted check in our model, and 79% of top sites fail it.
The pattern is familiar: a headline, a hero image, a short punchy line, then navigation. Beautiful for a human who chose to be there. Useless to a system scanning for a self-contained response.
2. Fewer than half have exactly one H1
Fig. 3H1 count per homepage, n = 243
| H1 count | Sites |
|---|---|
| Zero | 105 |
| Exactly one | 112 |
| Two | 11 |
| Three | 3 |
| Four or more | 12 |
A twenty-year-old fundamental, failed by 54% of the biggest sites on the web. 105 have no H1 whatsoever.
3. Only 5.8% declare both an author and dates
The weakest result in the study. Retrieval systems that cite prefer content they can attribute and date. Almost none of the top of the web makes that easy, which suggests the gap is not competitive difficulty: it is that nobody has bothered.
What should a site fix first?
If you have been told your AI visibility problem is authority, that may be true. But the data suggests a large share of pages are failing before authority is even assessed: there is simply no clean passage to extract.
The fixes are unglamorous and cheap, and 96.7% of the top of the web has not made them:
- Open with a real answer of 25 to 90 words.
- Use one H1.
- Break content into labelled H2 sections.
- Add a list or a table.
- Declare who wrote it and when.
- Link to your sources.
None of that requires a budget. Our guide to why websites fail the AI extractability test walks through each fix.
Methodology
Domains were drawn from the Tranco top-1M list (26 July 2026), the same sample frame as our AI crawler access study.
- Sample
- Tranco top-1M, 26 July 2026. Infrastructure, CDN and DNS domains excluded by pattern.
- Yield
- 506 domains attempted; 243 returned a fetchable homepage and were scored (48%).
- Fetch
- Each homepage fetched with a standard browser user-agent. Navigation, header, footer, script and form markup stripped before analysis.
- Checks
- Direct answer near the top (20), lists or tables (15), three or more H2 sections (15), exactly one H1 (10), question-form headings (10), no paragraph over 120 words (10), structured data present (10), author and dates declared (10), cites at least two external domains (10).
- Score
- 110 raw points, normalised to 100. Strong 80 and above, workable 55 to 79, weak below 55.
Limitations
Found during the study and corrected
- We caught a normalisation error mid-analysis. The checks sum to 110, not 100, and our first pass reported raw scores as percentages. Every figure here is normalised. Corrected.
- These are homepages, not articles. A homepage is not primarily an answer page, so these scores understate what the same sites achieve on editorial content.
- The yield rate was 48%. Homepages block automated fetches far more aggressively than robots.txt does, so the scored set skews toward sites with lighter bot protection.
- The weights are ours. They reflect our judgement about what matters for extraction, not a published standard. The raw per-check data is in the dataset so you can weight it differently.
Can I download the data?
Yes. Published openly: 506 rows with per-check results, paragraph lengths, heading counts and external link counts.
answer-extractability-study-2026-07-26.csv506 rows, per-check resultsHow do I score my own page?
The same nine checks are available as a free tool. It returns the specific fix for each failure and shows your heading outline as a machine sees it. Run the AI Answer Extractability Checker.
This is the second study in a series. The first looked at whether AI crawlers can reach you at all; the third carried the same checks to 148 local businesses in 20 cities. Together they cover access, and whether there is anything worth taking once you are in.
See all studies, datasets and the methodology on the MoxSEO Research page, or see how we apply this in our AI SEO service.
Questions about this study
What is AI answer extractability?
AI answer extractability is how easily an AI system can lift a clean, self-contained answer from a page and credit it. We score it with nine weighted structural checks, such as a direct answer near the top, exactly one H1, labelled sections, lists or tables, a declared author and dates, and cited sources.
How many top websites are structurally quotable by AI?
Only 8 of the 243 homepages we scored, or 3.3%, reached the strong band of 80 or more. 78 (32.1%) were workable at 55 to 79, and 157 (64.6%) were weak, below 55. The median score was 45 out of 100.
Which check do most top websites fail?
Author and dates declared: only 5.8% of the 243 sites pass it. Question-form headings pass on 16.9% and a direct answer near the top on 20.6%. The direct answer is the highest-weighted check, and 79% of top sites fail it.
What counts as a direct answer near the top?
An opening paragraph of 25 to 90 words: long enough to contain an answer, short enough for a model to lift whole. The median opening paragraph across the sample was 18 words, and 69 sites had no substantial opening paragraph at all.
Why did the study score homepages instead of articles?
The question was whether an AI asked about a company can extract a clean answer from its front door. A homepage is not primarily an answer page, so these scores likely understate what the same sites achieve on editorial content.
How is the extractability score calculated?
Nine checks carry 110 raw points: direct answer near the top 20, lists or tables 15, three or more H2 sections 15, and 10 each for exactly one H1, question-form headings, no paragraph over 120 words, structured data, author and dates, and citing two or more external domains. Scores are normalised to 100.
Can I score my own page?
Yes. The free AI Answer Extractability Checker runs the same nine checks, returns the specific fix for each failure and shows your heading outline as a machine sees it.