The Page Checks Most Sites Fail for AI Search, Ranked
We ran 20 common AI search checks on 3,691 pages. More than 80% of pages on company sites had no FAQ markup, and neither did 72.5% of pages AI already cites, so it can't be a requirement. Failures that can stop AI reading a page are rare, under 4%, and they're the ones to fix first. Here's the full ranking, with fixes.
The checks sites fail most often aren't the ones most likely to keep them out of AI answers. We ran 20 common AI search checks on 3,691 pages from company websites and pages AI already cites. More than 80% of company pages lacked each of these: a table, FAQ markup, a quotation and two links to other sites. But most pages AI cites fall short on those too. 72.5% of them have no FAQ markup, so it can't be a requirement for being cited.
Failures that can stop AI reading a page are rare, each on under 4% of the pages we could read, though we couldn't read 10.3% of the pages we tried, about half of those because they refused our crawler. If you fail one of these three, fix it first: good copy can't make up for a page an engine can't read. In the table, company pages come from the sites of companies in Ranqo, and pages AI cites turned up as sources in the AI answers we track.
20 page checks, ranked by how often they fail
Share of pages failing each check, ranked by company pages.
| Check | Company pages | Pages AI cites |
|---|---|---|
| 1. A tableNo proven effect | 86.4% | 63.7% |
| 2. Links to sourcesMay give AI less to quote | 86.2% | 62.5% |
| 3. FAQ markupNo proven effect | 84% | 72.5% |
| 4. A quotationMay give AI less to quote | 82.9% | 67.9% |
| 5. A figure up topMay give AI less to quote | 69.6% | 77.9% |
| 6. llms.txtNo proven effect | 48.3% | 52.5% |
| 7. Figures on the pageMay give AI less to quote | 46.9% | 36.9% |
| 8. A named authorNo proven effect | 42.7% | 22.5% |
| 9. A dateNo proven effect | 42.4% | 21% |
| 10. SubheadingsPage basics | 18.9% | 12.2% |
| 11. Any JSON-LDNo proven effect | 15.6% | 14.2% |
| 12. An H1Page basics | 11.5% | 4.1% |
| 13. A meta descriptionPage basics | 10.5% | 8% |
| 14. A sitemapPage basics | 9.1% | 4.6% |
| 15. A canonical tagPage basics | 8.7% | 6.2% |
| 16. Text in the HTMLCan stop AI reading the page | 3% | 1.1% |
| 17. No noindex or nosnippetCan stop AI reading the page | 1.1% | 0.9% |
| 18. Search crawlers allowedCan stop AI reading the page | 0.6% | 1.3% |
| 19. A titlePage basics | 0.5% | 0.2% |
| 20. A mobile viewportPage basics | 0.5% | 1.1% |
From 3,691 pages fetched in November 2026. Pass rules and fixes are in the sections below, and How We Measured says which pages each check ran on.
How to Read the Ranking
A high rank only tells you a check is commonly failed, not that it matters. So read each row with two questions in mind.
Do pages AI cites pass it?
If most cited pages fail a check, it can't be a requirement for being cited. That's true of 6 checks: figure up top, FAQ markup, quotation, table, links to sources and llms.txt. They can still help, but a page can get cited without them.
The reverse doesn't hold. Cited pages pass some checks far more often, and that doesn't prove the check earned the citation. Part of the gap is page type, since cited pages are often articles and company pages are often product pages. It doesn't all go away, though: on the smaller set of blog-style pages, 26.5% of company pages had no named author against 10.2% of cited pages. Publishers who add bylines may simply do everything else well too.
FAQ markup shows why this matters. Cited pages carry it more often, yet when Ahrefs tracked 1,885 pages that added JSON-LD against matched pages that didn't, they did no better: AI Overview citations fell 4.6% relative to the matched pages, with no measurable change in AI Mode or ChatGPT. That study pooled all schema types, and every page already had AI Overview citations, so it can't say what markup does for a page AI hasn't cited yet.
What's at stake?
- Can stop AI reading the page: an engine may not be able to fetch, read or show the page. Google, OpenAI and Anthropic document the crawl and index rules, and Vercel measured the JavaScript one.
- May give AI less to quote: the page can be read but offers less to quote. The lab evidence is mixed.
- Page basics: these help any reader, human or machine, tell what the page is. They fail less often here.
- No proven effect: we found no neutral test showing these change citations. Some are worth doing anyway.
Fix These First
These three rarely fail, but a single failure can matter a lot.
Checks that can stop AI reading a page
What each check looks for, and what to do if you fail it.
| Check | You pass if | The fix |
|---|---|---|
| Text in the HTML | You pass if: The page has 50+ words of text in its HTML before any script runs. Navigation and footer text count. | The fix: Render the main content on the server, or pre-render it. |
| No noindex or nosnippet | You pass if: No robots meta tag or X-Robots-Tag header sets noindex, none, nosnippet or max-snippet:0. | The fix: Remove the directive from any page you want quoted. |
| Search crawlers allowed | You pass if: robots.txt lets OAI-SearchBot, Claude-User, PerplexityBot, Googlebot and Bingbot fetch the page. | The fix: Remove the rule that blocks them. |
Text in the HTML
3% of company pages and 1.1% of cited pages had under 50 words before any script ran. In December 2024, Vercel's analysis found that “none of the major AI crawlers currently render JavaScript,” including OpenAI's, Anthropic's and Perplexity's, while Gemini uses Google's crawling, which does. That's nearly two years old, so test instead of assuming. Navigation counts toward our 50 words, so a page can pass with an empty body: open View Source and look for your main text, as our guide to what AI sees explains.
Search crawlers allowed
0.6% of company pages and 1.3% of cited pages were blocked in robots.txt for at least one of those crawlers. OpenAI says blocking OAI-SearchBot keeps a site out of ChatGPT's search answers, navigational links aside, and Anthropic says turning off Claude-User or Claude-SearchBot may reduce visibility. Our capture didn't cover Claude-SearchBot, so check that one yourself. Blocking a training crawler such as GPTBot is a separate choice, so we didn't count it.
This check reads robots.txt only, and a firewall that refuses bots won't show up here. Our guide to testing for blocks covers that case.
No noindex or nosnippet
1.1% of company pages and 0.9% of cited pages failed. Google says a page must be indexed and eligible for a snippet to appear in its AI features, and nosnippet also stops content being a direct input for AI Overviews and AI Mode. To test a page, search its source for both words, or use URL Inspection in Search Console.
Search Console's Search generative AI control also decides whether your site can appear in those features. It's on by default and a page check can't see it, so open it and confirm, including for any section set up as its own property.
What May Give AI Less to Quote
A page that passes those three is open to being read, as far as its own HTML and robots.txt go. These four are about how much it offers to quote, and the evidence is mixed. In a 2024 academic study (the GEO paper, at KDD), having a language model rewrite a source to cite sources, add quotations or add statistics raised its share of the answer by 30-40% on average, though top-ranked sources lost share and lower-ranked ones gained. A 2025 benchmark of the same rewrites, which looked at where a source was cited, found them largely ineffective and often harmful to its rank.
Neither tested a live page, and neither tested links or a figure up top as we define them. So do these because they make a page more useful and easier to trust, not because a check promises a citation. Use your own real sources and figures.
Checks that may give AI less to quote
What each check looks for, and what to do if you fail it.
| Check | You pass if | The fix |
|---|---|---|
| Links to sources | You pass if: Paragraphs or article lists link to other websites at least twice, leaving out social networks and big platforms such as Google and GitHub. | The fix: Link the claims that rest on someone else’s data. |
| A quotation | You pass if: The page has a blockquote of 8+ words or a quoted passage of 40+ characters. | The fix: Quote a named person saying something worth repeating. |
| A figure up top | You pass if: The first 100 words of the page’s text hold a percentage, price, multiple, or a number written with commas or as millions or billions. | The fix: Open with the number that answers the question, if there is one. |
| Figures on the page | You pass if: The page holds 2+ figures of that kind. | The fix: Replace “many” and “most” with numbers you can back. |
62.5% of cited pages have fewer than two links to other websites and 67.9% have no quotation, so neither is a gate, though both fail more often on company pages (86.2% and 82.9%).
Ranqo's research paper advises putting your facts up front, and a figure in the first 100 words is a plain test of that. Most cited pages fail it, 77.9% against 69.6% of company pages, though our rule counts only a narrow set of figures. Lead with a number when one answers the question, because it helps the reader. Just don't expect it to set you apart from pages AI already cites.
The Basics
Microsoft says titles, descriptions and H1s help AI systems read a page's purpose. They fail less often in this sample, but this sample is better kept than the web at large. In HTTP Archive's 2025 crawl of 16 million sites, a meta description, a canonical tag and a usable H1 were each missing on about a third of pages. If your site is smaller or older, expect to fail these more often than our ranking suggests.
This sample is better kept than the web
Share of pages missing each basic.
| Check | Company pages | Pages AI cites | Web-wide |
|---|---|---|---|
| An H1 | 11.5% | 4.1% | 34% |
| A meta description | 10.5% | 8% | 32.8% |
| A canonical tag | 8.7% | 6.2% | 33% |
| A mobile viewport | 0.5% | 1.1% | 4.8% |
| A title | 0.5% | 0.2% | 1.5% |
Web-wide figures are HTTP Archive's July 2025 crawl on mobile, subtracted from the presence rates it prints.
The basics
What each check looks for, and what to do if you fail it.
| Check | You pass if | The fix |
|---|---|---|
| Subheadings | You pass if: The page has 2+ H2 subheadings. | The fix: Break long text into titled sections. |
| An H1 | You pass if: The page has an H1 with text in it. | The fix: Make the page’s main heading an H1. |
| A meta description | You pass if: The page has a meta description with text in it. | The fix: Summarize the page in a sentence or two. |
| A sitemap | You pass if: The site has a sitemap at a standard path or listed in robots.txt. | The fix: Publish one and list it in robots.txt. |
| A canonical tag | You pass if: The page declares a canonical URL. | The fix: Point it at the version you want read. |
| A title | You pass if: The page has a title tag with text in it. | The fix: Write one that names what the page covers. |
| A mobile viewport | You pass if: The viewport tag sets device-width. | The fix: Add the standard viewport tag. |
Popular Checks With No Proven Effect
These turn up on many AI readiness checklists. We found no neutral test showing any of them changes citations, though some are worth doing for other reasons.
Checks with no proven effect
What each check looks for, and what to do if you fail it.
| Check | You pass if | The fix |
|---|---|---|
| A table | You pass if: The page has at least one HTML table. | The fix: Use one when you’re comparing things. |
| FAQ markup | You pass if: The page has FAQPage markup. | The fix: Skip the markup. Add a visible FAQ only if readers need one. |
| llms.txt | You pass if: The site serves a real llms.txt: text, with headings or links. | The fix: Optional. Don’t expect it to change Google’s AI features. |
| A named author | You pass if: The page shows a byline or carries an author tag or author markup. | The fix: Name a real person where a person wrote it, on the page as well as in the markup. |
| A date | You pass if: The page shows or declares a publish or update date. | The fix: Show the real date, and change it only when the page changes. |
| Any JSON-LD | You pass if: The page has at least one JSON-LD block. | The fix: Add it where it earns a rich result or keeps your data clean. |
FAQ markup and other JSON-LD
This is code in the page that readers don't see. Google says structured data isn't required for generative AI search and no special markup exists for it. Microsoft's guidance does recommend schema, but that's advice and not a test. Keep markup where it earns a rich result or keeps your data clean. We wouldn't add it to win citations, and our GEO checklist left FAQ markup off for this reason.
llms.txt
This is a text file that summarizes a site for AI tools. 48.3% of company sites and 52.5% of cited sites had none we could read, so cited sites were no more likely to have one. That's far higher adoption than the about 2.1% HTTP Archive found across the web in July 2025, though we can't say how much of it is deliberate: several files we spot-checked came from an SEO plugin, and HTTP Archive traced 39.6% of valid files to one CMS plugin. Google says its Search ignores the file.
A named author and a date
Cited pages pass these far more often: 42.7% of company pages had no named author against 22.5% of cited pages, and 42.4% had no date against 21%. Much of that is markup readers never see. Counting only what's visible on the page, 62.7% against 53.1% had no author, and 72% against 53.7% had no date.
The evidence on dates is thin and measures different things. A lab test moved rankings by changing dates, while a study of ChatGPT found that recency didn't reliably predict which of the pages it had retrieved it went on to cite. We found nothing citable on bylines. Show both anyway, since they're cheap and readers use them. Google encourages accurate bylines and warns against changing a date when the page hasn't really changed.
A table
86.4% of company pages had no table against 63.7% of cited pages, so most pages AI cites don't have one. Microsoft recommends comparison tables, but we found no neutral test of tables on their own. Use one when a reader is comparing things.
What Passing Won't Do
A page that passes all 20 still needs a reason to be chosen. In Ranqo's research paper, which measured five AI engines between March and May 2026, household names appeared in an average of 73% of answers to questions that didn't name them on their first tracking run, mid-market brands in 44% and niche brands in 11%. A brand's own site drew 2.9% of citations. So for most brands, what other sites say counts for a lot, which is why we'd pair these fixes with earning a place on the lists AI cites.
This ranking is also a snapshot. It shows what pages look like, and we didn't test whether fixing a check raises citations.
How We Measured
We fetched every page ourselves and read plain facts from its HTML, with no AI model involved:
| Pages AI cites | 2,641 pages read, one per website, from pages cited in the answers Ranqo tracks, most of them cited once. Forums, social networks, video sites, files such as PDFs and the sites of the brands being tracked are left out |
|---|---|
| Company pages | 1,050 pages read, a few per site: homepages and inner pages from the sites of companies in Ranqo, customers and prospects alike, mostly in business software and services and including many marketing agencies. Some of these pages may be cited by AI too |
| Fetching | November 2026, as RanqoBot, reading robots.txt first and skipping any page it disallowed. We couldn't read 422 of the 4,113 pages we tried: 200 refused us or disallowed us, and the rest errored, no longer existed or returned nothing. They're left out, not counted as failing, so pages that turn bots away are undercounted |
| Content checks | Links, quotation, both figure checks, subheadings, FAQ markup, JSON-LD, author, date and table ran only on English pages of 300+ words that aren't homepages: 450 company pages and 2,336 cited pages. Text in the HTML, crawlers, noindex, title, H1, meta description, canonical and viewport ran on every page read |
| Site checks | Sitemap and llms.txt count once per site |
| Left out | Page speed, because a page fetch can't measure it. HTTP Archive puts Core Web Vitals good on 48% of mobile sites, and we found no neutral study tying speed to AI citations. Readability, because the Flesch score was built for English and is a poor predictor of reading ease even there. Any check that needs a model's judgment |
How Ranqo Fits
The free AI Readiness Score runs a similar set of checks on any page in seconds, though its rules differ from the ones above. The free AI Crawler Inspector fetches a URL as each AI crawler and shows what came back.
In the dashboard, Page Audit scores a page on your site across six dimensions and lists what needs work under Top Recommendations, and you can re-audit after a fix and compare in History. Site Access fetches your site as each AI crawler against a browser, so it catches the refusals robots.txt doesn't show. Content Optimizer proposes rewrites for a page and tells you which facts to add, rather than inventing them. Some of these score items this post treats as unproven, so use the labels above to decide what comes first.
See what needs work on your pages
Page Audit shows what to fix on a page, and Ranqo's weekly tracking shows whether AI answers cite it. Try it free for 7 days, and read how to audit your site for AI readiness.
Start your free trialWritten by
Nisha Kumari
Nisha Kumari is Co-Founder at Ranqo, where she leads growth strategy and client acquisition. With a background in digital marketing and financial management, she specializes in SEO, Generative Engine Optimization, and helping brands build visibility across AI platforms.
Share this article
Related articles
How to Audit Your Website for AI Readiness
Most websites are optimized for Google but invisible to AI. A 6-dimension audit -- crawlability, content quality, page speed, AI readiness, citation potential, and authority -- reveals exactly where the gaps are.
The GEO Checklist: 30 Checks in the Order They Break
In our study of 149,912 AI citations, only 2.9% pointed to a brand's own site, yet most GEO checklists audit your own pages and add FAQ schema and llms.txt, which Google says its AI features don't need. AI visibility breaks in an order: access, the page, format, off-site sources, measurement. Here are 30 checks in that order, each with how to verify it.
What AI Actually Sees When It Crawls Your Site: A Live Walkthrough
Your site looks great in a browser. But AI crawlers see only raw HTML -- no JavaScript, no rendered components, no dynamic content. This is a live walkthrough of exactly what GPTBot, ClaudeBot, and PerplexityBot fetch when they visit, with verified data on every claim and a 6-method test you can run today.