Is llms.txt Dead in 2026? The SE Ranking 300K-Domain Study
10.13% of indexed domains publish llms.txt; 0% of the top 1000 do. SE Ranking May 2026 found no measurable citation lift. What moves the needle instead.
The honest update
We published What Is llms.txt and Why Every Site Needs One in 2026 on 2026-04-22. The post argued that llms.txt was the de facto AI-identity standard and would soon be universal. It quoted an early-N study suggesting sites with llms.txt were 1.6x more likely to be cited correctly by Perplexity.
SE Ranking ran the largest llms.txt analysis to date in May 2026 - roughly 300,000 indexed domains, controlled for site authority, schema markup density, and content recency. The findings do not support the strong version of that earlier claim. They do support a softer version. This post is the update.
The headline numbers from the SE Ranking dataset:
| Metric | May 2026 | Apr 2025 |
|---|---|---|
| llms.txt adoption, all domains | 10.13% | 0.4% |
| llms.txt adoption, top 1000 domains | 0% | 0% |
| Citation-rate lift attributable to llms.txt (controlled) | not statistically significant | (small-N) +60% reported |
| Optimal length per citation-rate regression | ≈ 800 chars | (recommended) 800-3000 chars |
Adoption grew ~25x in 13 months, which is genuine. The signal is real. The leverage on citations is not what the early-2025 narrative implied. The top of the web - the high-traffic publishers and platforms whose adoption would force AI engines to treat the file as authoritative - has not moved. The middle of the web has adopted heavily, which is why aggregate adoption looks like a hockey stick from the bottom and a flat line from the top.
What the early figures actually measured
The 1.6x-citation-accuracy figure widely cited from April 2025 was a small-N early study at the level of a few thousand domains, with no control for site authority, schema density, or content recency. SE Ranking's larger study replicated the methodology at ~300K-domain scale with the controls added. The lift attributable to llms.txt presence specifically - holding everything else equal - shrank to a non-significant effect.
The cleanest interpretation: the original 1.6x lift was a population-of-publishers confound. The publishers who adopted llms.txt early in 2025 were also the publishers doing every other AEO best-practice - FAQPage markup, content recency, structured Q&A, internal linking, the works. The citation lift came from the bundle of practices that early adopters happened to ship together, not from the llms.txt file itself.
This is the standard pattern when a new signal gets a strong early reading. The early adopters self-select; the lift looks attributable to the new signal; larger-N replication with controls dissolves the effect. It happened with PageRank-adjacent signals in the 2010s, with structured-data signals in the 2010s, and now with llms.txt in 2025-2026. The right response is the same in every case: update the model.
What the May 2026 data shows lifts citations
The positive-control set in SE Ranking's study held llms.txt presence constant and varied the schema markup density on the actual content. The results across the four major AI engines studied - ChatGPT search, Claude, Perplexity, and Google AI Mode - showed four clean signals.
FAQPage with real Q/A pairs
FAQPage with 4+ Question/Answer pairs scoped to questions users actually ask, with answers ≥300 chars each: +34% citation rate in Perplexity, +28% in ChatGPT search, controlling for content recency and domain authority.
The mechanism is mechanical. AI engines lift FAQPage-marked Q/A pairs almost verbatim into citation snippets when the question matches the user query. A page with 6 well-marked Q/A pairs that address 6 different user intents is, effectively, 6 citation opportunities competing for 6 different queries. A page with the same content in flowing paragraph form is 1 citation opportunity, and the AI engine has to do the extraction work itself, which it does less reliably.
We documented this pattern in FAQPage Schema for Citation Lift and the operating advice has not changed - 4-6 Q/A pairs per content page, each ≥300 chars, each addressing a distinct user question, each scored against the actual search queries the page targets.
ClaimReview on stat-dense paragraphs
ClaimReview with claimReviewed and reviewRating wrapped around verifiable statistical claims: +41% citation rate in AI Mode on the marked content, in the same study.
The mechanism is signal. Most stat-dense paragraphs on the web are unsourced; AI Mode penalizes them. A paragraph wrapped in ClaimReview tells the engine that the claim has been reviewed and rated, which lifts it above competing unsourced claims even when the underlying numbers are identical.
The implementation tax is real - every ClaimReview block requires claimReviewed (the proposition being reviewed), reviewRating (numeric, on a defined scale), author (the reviewer), datePublished, and itemReviewed.url (the source being reviewed). For a content team shipping 4 stat-dense posts per month, the per-post markup overhead is about 15 minutes. The lift on AI Mode citations more than pays for that time.
SameAs entity linkages
Organization schema with SameAs URLs pointing to Wikidata, LinkedIn, Crunchbase, GitHub, and (when applicable) Twitter and YouTube: +22% entity-disambiguation accuracy across the four engines studied.
The mechanism is verification. There are at least 4 companies named "Inite" across travel, fitness, and consulting. There are at least 12 "Apex" software products. When an AI engine resolves an entity reference in user query → entity → site URL, it cross-checks the SameAs targets to confirm it is mapping to the right entity. SameAs to Wikidata is the strongest single signal because Wikidata is the canonical entity graph the major engines defer to.
Three SameAs URLs - Wikidata + LinkedIn + GitHub - delivered roughly 80% of the entity-disambiguation lift in the SE Ranking data. The 4th and 5th targets added small marginal lifts. Past 5-6 targets the lift plateaued.
Speakable with cssSelector
Speakable schema with cssSelector targeting direct-answer blocks (.aeo-direct-answer or H1+H2 pairs): +18% answer-extraction accuracy in AI Mode, where the cited snippet matched the speakable-marked text rather than a different passage on the same page.
The mechanism is hint. The speakable property tells the engine "this passage is a clean answer to a question and is suitable for voice/answer extraction." AI Mode is the engine that listens to this signal most aggressively. ChatGPT and Claude listen less; Perplexity, in the SE Ranking study, did not show a measurable response to speakable specifically (though their citation logic does seem to favour passages that are short, declarative, and front-loaded - which is what speakable-marked passages tend to be).
We documented the pattern in Direct Answer Blocks for AEO and the inite.ai analyzer checks for it on every audit.
What llms.txt still does well
The contrarian frame should not overshoot. llms.txt remains the right surface for two specific jobs that the SE Ranking citation-rate metrics do not measure directly.
Entity disambiguation. If your brand name is non-unique, llms.txt is one of several inputs an AI engine uses to decide which entity you are. The SE Ranking study did not measure entity-disambiguation accuracy as a function of llms.txt presence; it measured citation rate. There is no published large-N study yet that isolates llms.txt's contribution to entity disambiguation. The mechanism is plausible - a clear H1 with the legal name, a one-line description, and an unambiguous SameAs anchor in the markdown helps the engine pin the entity - but the quantitative case is unmade.
Programmatic instruction routing. If you want AI agents to fetch from /docs/llm-overview.md rather than your sales landing page, llms.txt is where that routing is declared. As agentic browsing crosses 5-10% of inbound traffic in 2026, this matters more, not less. The SE Ranking dataset measured citation rates from search-flavour interfaces (ChatGPT search, Perplexity, AI Mode), not agentic-traffic routing - so this dimension is outside what was measured.
Neither of these maps to a clean citation-rate uplift in the May 2026 numbers. Both are real. The honest position in mid-2026: ship llms.txt, ship it well, but do not bill the work as a citation-lift project. The lift comes from the FAQPage / ClaimReview / SameAs / speakable substrate that goes around it.
Where the time should go
If a team has 8 hours of AEO time per week in mid-2026, the SE Ranking findings point to a clean allocation. We use this allocation across the /analyze tool and our own content team:
| Hours/week | Activity | Citation-lift basis |
|---|---|---|
| 4 | FAQPage + ClaimReview + speakable markup on new and existing content | +28-41% per engine |
| 2 | SameAs maintenance + content recency refresh on top 20 cited pages | +22% disambiguation, +12% staleness penalty avoided |
| 1 | Identity surface tier (llms.txt, ai.txt, robots-ai.txt, .well-known/agent-actions) | substrate; no isolated lift |
| 1 | Citation tracking + measurement (see Citation Tracking Metrics) | feedback loop |
This is the reverse of the April 2025 allocation, where teams shipped large llms.txt + ai.txt files first and added schema markup later. The May 2026 data flipped the priority.
What would change our mind
Two things need to be true for llms.txt to matter more in 2027 than it did in 2026.
Top-1000 adoption needs to cross 25-40%. As long as the high-traffic web does not publish llms.txt, the AI engines build their citation pipelines without it as a primary input. They cannot make a file authoritative when 0 of 1000 of the top sites publish it. The fastest path forward is a major-platform default - Shopify, Vercel, WordPress, or Squarespace shipping llms.txt by default for every customer site. None of those have happened. If one does, the calculus changes.
The spec needs to standardize on a small machine-readable schema. The current spec is freeform markdown, which means every AI engine ends up parsing it differently and treating it as a hint rather than a contract. The IETF draft work on AI Identity - informally tracked alongside the Web Bot Auth HTTP Message Signatures spec - points toward a JSON-LD-compatible identity profile that engines could rely on as a contract. If a small, strict schema lands and the major engines commit to honoring it, the file becomes high-leverage again.
Neither change is imminent. We are revisiting in October 2026 with the H2 SE Ranking refresh.
What changed in the analyzer
The /analyze tool's citation-readiness score reweighting, motivated directly by the SE Ranking findings:
| Score component | Old weight | New weight |
|---|---|---|
| Identity surface (llms.txt, ai.txt, robots-ai.txt, .well-known/agent-actions) | 35% | 15% |
| Schema patterns (FAQPage, ClaimReview, SameAs, speakable, Organization) | 30% | 50% |
| Content recency + direct-answer density | 25% | 25% |
| Robots policy + crawler allowlist hygiene | 10% | 10% |
The new score better reflects where the empirical lift actually lives in mid-2026. Findings in the analyzer report now carry a per-finding citation-lift estimate keyed to the SE Ranking dataset - so a team shipping FAQPage with 6 Q/A pairs sees an explicit "+34% Perplexity citation rate (SE Ranking 2026 estimate)" attached to that finding.
The other analyzer tools - citation tracking, ai-crawler allowlist, browser-agent readiness - all sit in the same product. The reweighting is one update, not a relaunch.
The single-sentence summary
llms.txt is not dead and it is not useless, but it is not the citation-lever the early-2025 narrative implied; the SE Ranking May 2026 study tells us where the lift actually lives in mid-2026, which is the Schema.org substrate - FAQPage, ClaimReview, SameAs, speakable - and that is where teams with 8 hours of AEO time per week should spend it.
Update the model when the larger-N data tells you to. Adjust the allocation. Keep shipping.
01If 10.13% of domains have llms.txt and it does not lift citations, why publish it at all?+
Because the absence of a measurable citation lift in the aggregate is not the same as the file having no purpose. llms.txt remains the right surface for two specific jobs that are not measured by citation-rate metrics. (1) Entity disambiguation - if your brand name is non-unique (there are at least 4 companies named 'Inite' across travel, fitness, and consulting), llms.txt is one of the inputs an AI engine uses to decide which entity you are. The SE Ranking study did not measure entity-disambiguation accuracy directly; it measured citation rate. (2) Programmatic instruction routing - if you want AI agents to fetch from /docs/llm-overview.md rather than your sales landing page, llms.txt is where that routing is declared. Neither of these maps cleanly to a citation-rate uplift, but both are real and both matter once your traffic from agentic browsing crosses 5% of total. The honest position in mid-2026: ship llms.txt, ship it well, but do not bill the work as a citation-lift project. The lift comes from the FAQPage / ClaimReview / SameAs / speakable substrate that goes around it.
02What did the April 2025 narrative get wrong about llms.txt?+
Three things. (1) The 1.6x-citation-accuracy figure widely quoted from a small-N early study did not survive larger-N replication. SE Ranking's ~300K-domain analysis at scale showed no statistically significant lift in citation rate attributable to llms.txt presence, once controlled for the confounders. The original finding looks like a population-of-publishers effect: the publishers who adopted llms.txt early were also doing every other AEO best-practice, and the citation lift came from the bundle, not from the file. (2) Adoption was projected to follow robots.txt's trajectory toward universal coverage. In practice it has stalled in the mid-tier and not penetrated the top-1000 high-traffic web at all. The major content platforms, the major retailers, the major media houses - none of them have llms.txt as of May 2026, which means the AI engines cannot rely on it as a primary signal even if they wanted to. (3) The contract was over-specified - early templates suggested 3-5 KB markdown files with detailed product trees. Most AI engines, when they fetch llms.txt at all, treat it as a hint, not a contract. The 144-line llms.txt vs the 39-line llms.txt produces no measurable difference in either entity-disambiguation accuracy or citation rate. The optimal length per the SE Ranking data is closer to 800 chars than 3 KB.
03Then what does move the citation needle in mid-2026?+
Schema.org markup on the actual content that needs to be cited, not declarative-identity files at the root. The SE Ranking dataset's positive-control set - sites with the same llms.txt presence but varying schema density - showed clean linear correlation between schema-marker count and citation rate. Four specific markup patterns produced the cleanest lifts. (1) FAQPage with 4+ Question/Answer pairs scoped to the question users actually ask: +34% Perplexity, +28% ChatGPT search. The mechanism is mechanical - AI engines lift these directly into their citation snippets. (2) ClaimReview with claimReviewed + reviewRating wrapped around verifiable statistical claims: +41% AI Mode citation rate on the marked content. The mechanism is signal - the schema tells the engine the claim has been reviewed, which lifts it above unsourced competing claims. (3) Organization with SameAs URLs pointing to Wikidata, LinkedIn, Crunchbase, GitHub, and (when applicable) Twitter and YouTube: +22% entity-disambiguation accuracy. The mechanism is verification - AI engines cross-check the SameAs entries to confirm they are talking about the right entity. (4) Speakable with cssSelector targeting direct-answer blocks: +18% AI Mode answer extraction accuracy. The mechanism is hint - the schema tells the engine 'this passage is a clean answer to a question', which lifts it above competing passages on the same page. Stack all four on a page and the marginal lifts compound, modestly. None of the four require llms.txt to work.
04How should a team in mid-2026 actually invest its AEO time?+
Three-tier priority. Tier 1 - ship the four schema patterns above on every content page that should be cited. FAQPage with real, distinct, useful Q/A pairs ≥300 chars each. ClaimReview on stat-dense paragraphs. Organization SameAs on the site root. Speakable on the direct-answer blocks. This is where the 18-41% citation-rate lifts live. Tier 2 - get the entity-identity surface right. Publish llms.txt (small, ~800 chars, tight), ai.txt with the SEMrush-grade detail patterns shown in [the original llms.txt comparison post](/en/blog/llms-txt-vs-ai-txt-vs-robots-txt), .well-known/agent-actions for agentic checkout endpoints if you have them, and a clean robots.txt that allowlists GPTBot/ClaudeBot/PerplexityBot/Google-Extended explicitly. This is the substrate that lets every other signal land. The lift is not measurable in isolation but the absence of this tier silently caps the others. Tier 3 - the harder, slower work that produces compounding lift. The [direct-answer block pattern](/en/blog/direct-answer-blocks-aeo) - one 300-700 character paragraph per page that answers the page's primary question cleanly. The [citation-tracking metrics](/en/blog/citation-tracking-aeo-metrics) that let you measure your own lift week-over-week. The content-recency discipline that gets you re-fetched. The internal linking pattern that consolidates topical authority. This is 80% of the citation-rate ceiling. Spending Tier 3 time on a perfect llms.txt is misallocated capital.
05What does this mean for the inite.ai AEO analyzer product specifically?+
The analyzer at [/analyze](/en/analyze) checks llms.txt, ai.txt, .well-known/agent-actions, robots.txt, and the full Schema.org markup graph - same nine-step pipeline it has always had. What is changing in the H2 2026 product iteration, motivated directly by the SE Ranking findings: the citation-readiness score that the analyzer returns is being re-weighted. The current weighting was substrate-heavy (llms.txt + ai.txt + robots-ai.txt accounted for 35% of the score). The mid-2026 reweighting pushes that to 15% and lifts the schema-pattern weighting from 30% to 50%. FAQPage / ClaimReview / SameAs / speakable presence and quality are now the dominant axis of the score, with content recency and direct-answer density at 25%, and identity surface at 15%. The full analyzer report now ships with a per-finding citation-lift estimate keyed to the SE Ranking dataset - so a team that ships FAQPage with 6 Q/A pairs sees an explicit '+34% Perplexity citation rate (SE Ranking 2026 estimate)' on that finding. The job of the product is to tell teams where the lift actually lives in mid-2026, not where the lift used to live in mid-2025.
06Will llms.txt come back? What needs to be true for it to matter more?+
Two things need to change. (1) Top-1000 adoption needs to cross some adoption threshold - probably 25-40% - for AI engines to treat the absence of llms.txt as a missing signal. As long as the top of the web does not publish it, the engines build their pipelines without it as a primary input, which means publishing it stays low-leverage. The fastest path to top-1000 adoption is a major-platform mandate - Shopify, Vercel, WordPress, or Squarespace shipping llms.txt by default for every customer site - which has not happened. (2) The spec needs to standardize on a small, machine-readable schema rather than freeform markdown. The IETF draft work on AI Identity (informally tracked alongside the Web Bot Auth HTTP Message Signatures spec) points in that direction. A standardized identity profile - probably JSON-LD compatible, probably a strict superset of Schema.org Organization - would be the version of llms.txt that AI engines could rely on as a contract rather than a hint. If both happen, the file matters more in 2027. If neither happens, llms.txt stays a useful but secondary surface, exactly as it is now.
