97% of llms.txt files are never read — and 4 other AEO assumptions the 2026 data just killed

Four rigorous studies, cross-checked against each other

This isn't a single new study — it's a cross-check of the four most rigorous, large-scale pieces of AEO research published in the past year, so the conclusions don't rest on one methodology or one company's dataset.

• Ahrefs' primary research: a crawl-log analysis of 137,000 domains' llms.txt adoption, a correlation study across 75,000 brands, and a citation-source analysis of 863,000 keywords and 4M AI Overview URLs (2025–2026).
• Zyppy's AI Citation Ranking Factors analysis: an evidence-weighted meta-analysis by Cyrus Shepard synthesizing 54 experiments, patents, and case studies into a 0–10 score per factor (May 2026).
• Semrush's 2026 AI Visibility Index: 126 million U.S. AI search prompts across ChatGPT, Gemini, Google AI Mode, and AI Overviews (April 2026).
• The founding academic study, "GEO: Generative Engine Optimization" (Princeton, Georgia Tech, IIT Delhi), the first controlled experiment on which content changes move AI visibility (ACM KDD 2024).

llms.txt does almost nothing right now

Ahrefs studied server logs and live traffic across 137,000 domains and found that only about 3% of the sites hosting an llms.txt file ever had it fetched by anything at all — and "fetched" doesn't mean "used." Of that already-tiny slice, named AI tools like GPTBot accounted for roughly a fifth of requests; a meaningful chunk of the rest came from the GEO tooling industry auditing itself, not from AI platforms consuming the file.

Zyppy's independent meta-analysis, working from a completely different set of studies, landed in the same place: llms.txt scored 2.0 out of 10 for evidence of impact on AI citations — the lowest score of the 23 factors it evaluated.

Two unrelated research efforts, two different methods, one conclusion: llms.txt is largely theater right now. It's cheap to add, but it shouldn't be anywhere near the top of a startup's AEO priority list.

Schema markup is oversold as a growth lever

Zyppy scored structured data at 5.6/10 and flagged it as contested — LLMs aren't trained on schema markup directly, so there's no clean mechanism for why it would help. Ahrefs' own AI Overview studies found adding schema markup showed no meaningful positive effect on citation rates, and a slight dip in one analysis.

Some studies still find a positive correlation, which is why it's not scored at the bottom like llms.txt — but the evidence doesn't support treating it as a growth lever.

Your brand's footprint off your own site

Ahrefs' correlation study across 75,000 brands is the most useful single dataset here. Off-site brand signals dominate: brand web mentions correlate at 0.664 with AI Overview visibility, brand anchor text at 0.527 — both roughly two to three times stronger than backlinks, which sit at just 0.218. The single strongest individual signal was YouTube mentions, at 0.737, ahead of every traditional SEO metric tested.

Semrush's much larger 126-million-prompt study points the same direction from a different angle: companies running AI visibility as part of an integrated SEO/content/brand program saw meaningfully better outcomes than those treating it as a separate initiative — 81% reported traffic or lead gains, versus 36% for teams running it in a silo.

The mechanism is intuitive once you see the numbers: AI models absorb signal about your brand from everywhere it gets discussed, not just from your own domain. A decade of SEO trained marketers to optimize their own site above all else. That instinct under-serves AI visibility.

Specificity and sourcing genuinely help

The Princeton/Georgia Tech/IIT Delhi paper that coined "Generative Engine Optimization" ran a controlled experiment — not a correlation study — across roughly 10,000 queries, testing nine content strategies. Adding statistics and citing external sources produced the largest, most consistent gains, on the order of a 30–40%+ visibility lift depending on the metric. Keyword stuffing, by contrast, produced little to no effect.

One honest caveat: this study used GPT-3.5-turbo in a controlled benchmark, not the current retrieval-augmented systems powering ChatGPT, Gemini, or Perplexity today. The direction of the finding (specific, sourced, well-attributed content gets cited more) is echoed in the newer 2026 studies above — but the exact magnitude shouldn't be treated as current gospel.

Ranking well the old-fashioned way still matters most

Zyppy's top five evidence-weighted factors were URL accessibility, traditional search rank, "fan-out" rank (how a page performs across the many sub-queries an AI system generates from one prompt), preview controls, and query-answer match — all scored above 9.2/10. Ahrefs' citation-source study backs this up: even as AI Overviews increasingly pull from beyond the top 10 organic results (down to 38% of citations from the top 10, from 76% roughly a year earlier), ranking well in traditional search is still the single strongest, best-evidenced lever available.

Translation: most of what makes a page rankable in Google still makes it citable in AI search. The industry's newest advice is mostly a reweighting of fundamentals that were already true, not a wholesale replacement of them.

AI citations don't stay put

Independent GEO-agency research (not from Ahrefs, Semrush, or Zyppy — treat with more caution given the commercial incentive to make the space look complex) has tracked the same queries across multiple AI platforms over several weeks and found citation churn is high: in one study tracking over 1,100 URLs across three waves, roughly a third of cited sources persisted from one measurement to the next, with the rest replaced entirely.

We can't independently verify the exact figures, but the general shape — that AI citations are less stable than Google rankings — is directionally consistent with how these systems work (fresh retrieval on every query, rather than a cached ranking). Worth knowing, not worth building a strategy around a specific churn percentage from a single vendor study.

What this means for a funded B2B startup right now

The evidence points to a reweighting, not a revolution. Five moves follow directly from the data:

• Deprioritize llms.txt and schema markup as growth bets. Add them if they're cheap (they are), but don't spend a sprint on them or expect them to move a metric.
• Invest in being talked about off your own domain — press, YouTube, Reddit/community presence, and genuine mentions elsewhere carry more weight than anything purely on-site.
• Write with real specificity. Named studies, real numbers, and cited sources outperform confident-sounding generalities — one of the few findings with controlled-experiment backing, not just correlation.
• Don't abandon traditional SEO for a separate "AEO strategy." The evidence says they're mostly the same task right now: rank well, get crawled, be extractable.
• Measure AI citations separately from rankings, but hold the numbers loosely. The tooling is young and inconsistent between vendors; treat any single AI-visibility score as directional, not precise.