AI SEO Statistics You Can Actually Trust (and the Ones to Stop Quoting)

Most numbers in circulation trace back to vendor blog posts, lab studies, or a single snapshot. A few survive a check against the primary document. These are the ones worth quoting, with the limit each one carries.

Notice the pattern. The figures that hold up are narrow. Each one answers a specific question, and none of them promises an outcome. If you want the wider picture of what the research supports, see what AI SEO research actually shows.

AI answers pass through four stages, and most statistics only measure one of them.

• Parametric: what the model knows with no web access, frozen at a training cutoff.
• Access: whether a crawler may fetch and store your page.
• Retrieval: whether your page enters the model's context for a given query.
• Generation: which retrieved pages the model uses and cites.

Nearly every GEO technique with a published effect size was measured at the generation stage, on documents already placed inside the model's context. A July 2026 critical survey of 45 GEO studies (arXiv 2607.14035) concluded that already-retrieved content can change its citation, but no reviewed technique shows a stable, cross-platform causal effect on discoverability. That gap explains why lab numbers rarely survive contact with live engines. Our comparison of lab GEO studies and live engines covers what transfers.

The second problem is brand size. Big brands get cited more, rank higher, have more reviews, and update pages more often. Almost every correlational study in this field cannot separate those things. When a study says pages with feature X get cited more, the honest reading is often that large brands have feature X. The guide to correlation vs causation in AI search studies shows how to spot this.

The third problem is that citation is not the same as use. OpenAI's documentation says the number of sources a model consults is often greater than the number it cites. Any citation count undercounts influence, and not randomly.

Ahrefs found that 37.9% of URLs cited in Google AI Overviews also appeared within the first 10 results, in a study published 2 March 2026. The organic-only figure was 37.1%. Roughly 31% of citations came from positions 11 to 100 and roughly 31% from beyond position 100.

Three things limit how you can use it:

• It is not a trend. Ahrefs earlier reported around 76%, and many posts present the drop as a measured decline. Ahrefs changed its parsing method between the two studies, so the endpoints came from different instruments.
• It is URL-level. A large site cited through a page that doesn't itself rank counts as a miss, which penalises big sites.
• It conflicts with another figure. BrightEdge reported only 16.7% of citations from the top 10 in September 2025, with no published sample size and an undefined overlap measure. These two numbers do not corroborate each other. One method is wrong or they measure different things.

What it means: organic ranking is associated with AI Overview citations, but most cited URLs sit outside the top 10. Google states there are no additional requirements or special optimizations needed to appear in AI Overviews or AI Mode. Keep doing sound SEO, and read SEO for AI search for where that effort actually lands.

Ahrefs tracked 1,885 pages that added schema markup and compared each with three matched control pages. It found no major uplift on any platform. AI Overviews citations fell 4.6%, a statistically significant drop. AI Mode rose 2.4% and ChatGPT rose 2.2%, neither of which Ahrefs treated as a meaningful lift.

This lines up with Google's own guidance that structured data isn't required for generative AI search. FAQ rich results were also deprecated in May 2026, which removes the most common reason agencies gave for adding FAQ markup.

One boundary matters. Every treated page already had more than 100 AI Overview citations before schema went on. So the study says nothing about whether schema helps a page that has never been cited get in. That question is untested, not settled.

The practical line: schema is not a lever for AI answers. Google says existing markup does no harm, so removing it isn't supported either. Spend the time on access checks instead.

The strongest statistic in the evidence base is about rendering. Vercel found that ChatGPT's crawler requested JavaScript files on 11.50% of requests and Claude's on 23.84%, but that neither executed them. In Vercel's words, they can't read client-side rendered content.

An independent retest in June 2026 found that GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, Claude-SearchBot, PerplexityBot, Meta-ExternalAgent and Bytespider did not render JavaScript. Googlebot and AppleBot are the exceptions, so this matters for ChatGPT, Claude and Perplexity, not for Google's AI features.

Carry the caveats. Vercel's data is a one-month sample from nextjs.org, Microsoft Copilot was excluded, and Vercel sells server-side rendering. A companion figure about AI crawlers hitting 404 pages comes from the same single-network snapshot and has never been replicated, so leave it out.

One more access number is real but rarely relevant. Google said in March 2026 that Googlebot fetches up to 2MB per URL and ignores bytes past that point. Google also says the vast majority of sites will never hit it.

What to do: open your key pages with JavaScript disabled and check that prices, product names and core claims are in the raw HTML.

These figures circulate widely. Each one breaks on inspection.

• "GEO delivers up to 40% more visibility." The GEO paper (Aggarwal et al., KDD 2024) measured this in an engine the authors built, with sources supplied in context. Its main metric counts words the intervention itself added. The paper reports a 30 to 40% band across three methods, not the itemised figures usually attributed to it. The same paper found keyword stuffing lowered its own metric (17.8 against a 19.5 baseline).
• "ChatGPT is 92% of AI traffic." The figure is 92.4% of trackable referral sessions on a self-selected agency panel of 166 properties. ChatGPT tags referral links with utm_source=chatgpt.com, most rivals don't, and Google AI Overviews send no referral at all. It measures tagging, not usage.
• "Anthropic crawls 38,000 pages per referral." Cloudflare's figure is a July 2025 snapshot. 2026 ratios are roughly an order of magnitude lower, and a training crawler has no referral surface by design.
• "75% of cited pages were updated within a year." Seer's study had no control group of uncited pages and read dates that publishers declare themselves. Without a base rate for the whole web, it says nothing about what engines reward.
• "Review profiles raise citation rates from 1% to 53.5%." The Seer study never defines citation rate and its tiers have no thresholds. Treat it as unusable.
• Any "AI rank" or position in Google AI. Google's Search Console report for AI features shows impressions only, with no clicks, CTR or position. No ground truth exists to check a rank against.

You cannot control what a model learned in training, which queries it runs, or which retrieved page it cites. You can control whether your pages are reachable, readable and eligible. Work in this order.

1. Check robots.txt per bot. As of October 2026, GPTBot governs OpenAI training only, and OAI-SearchBot governs ChatGPT search. OpenAI says each setting is independent. Also check PerplexityBot and Claude-SearchBot.
2. Check the firewall and CDN. A CDN can return 403 to a crawler that robots.txt allows. Cloudflare began blocking Training and Agent crawlers by default on ad-displaying pages for new domains and free-tier customers from 15 September 2026. The AI Crawler Checker requests a page as each AI crawler and shows which ones get turned away.
3. Check raw HTML. Make sure the facts that matter appear before JavaScript runs.
4. Check suppression controls. Look for stray noindex, nosnippet, max-snippet:0 and data-nosnippet. Review Google Search Console and Bing Webmaster Tools separately.
5. State specific facts plainly. A price, a number or a one-line definition on a fetchable page gives an engine a sentence it can quote.
6. Measure each engine on its own. In Otterly's 2026 experiment, Microsoft Copilot produced 2,718 of 3,934 tracked citations for one site while four Google and OpenAI surfaces barely cited it. A blended score would have hidden that. The free checker on this site gives a per-engine snapshot for up to three buyer questions.

• Treating vendor research as neutral. Ahrefs, Vercel, Otterly and Seer all sell something related to their findings, and so does 99Visibility. That doesn't make the data wrong. It means you should look for replication.
• Comparing numbers from different instruments. Estimates of how often ChatGPT searches the web range from about 18% (Profound) to 34.5% (Semrush) to 80.5% (cloro). They measure different things on different populations. None is the true rate.
• Trusting prompt-based counts as ground truth. ChatGPT rewrites queries before searching, so tools only sample the prompts they chose. See whether AI visibility tools can see real prompts.
• Believing scores that don't exist. Google says E-E-A-T isn't a specific ranking factor, so there is no E-E-A-T score to improve. Domain Authority is a Moz metric no engine computes.
• Accepting guaranteed outcomes. No statistic supports a promise of citations. Read why AI citation guarantees fail before signing anything that offers one.

Did AI Overview citations from top-10 pages really drop from 76% to 38%? No measured decline exists. Ahrefs reported around 76% in an earlier study and 37.9% in March 2026, but it changed its parsing method between the two. The endpoints came from different instruments, so you cannot read the gap as a trend. The 37.9% figure stands on its own as a March 2026 measurement of URL-level overlap between AI Overview citations and Google's first 10 results.

Does adding schema markup increase AI citations? The best available test says no. Ahrefs compared 1,885 pages that added schema against matched controls and found no major uplift, with a significant 4.6% drop on AI Overviews. Google says structured data isn't required for generative AI search. The study only covered pages already heavily cited, so whether schema helps a new page enter answers is untested. Existing markup does no harm.

Where does the 40% GEO uplift statistic come from? It comes from the 2023 GEO paper by Aggarwal et al., published at KDD 2024. The authors built their own answer engine and supplied sources directly in its context, so the test never touched live search. The main metric counted words the technique added. The paper reports a 30 to 40% band across three methods. It does not show that live engines cite you more.

Is ChatGPT really 92% of AI referral traffic? That figure is 92.4% of trackable referral sessions on a self-selected panel of 166 properties. ChatGPT adds utm_source=chatgpt.com to referral links, which makes it easy to count, while Google AI Overviews send no referral at all. Even ChatGPT's tagging is incomplete on the free tier and mobile app. Treat the number as a measure of tracking, not of market share.

Which AI SEO statistic should I act on first? Start with the crawler rendering data. Vercel found ChatGPT and Claude crawlers fetch JavaScript but don't execute it, and a June 2026 retest found the same across the main AI crawlers except Googlebot and AppleBot. If your prices or product facts load only after JavaScript runs, those crawlers may not see them. Fixing that is in your control and checkable.