RSL in 2027: Really Simple Licensing as a Standard for Link Publishers

TL;DR

RSL, Really Simple Licensing, is a machine-readable vocabulary that attaches terms and a price to individual paths on a site. Nearly every assessment asks the same question — will the AI companies honour it? — which grades the optional half of the standard.

The required half is a declaration: a dated, per-asset statement of what you would accept and from whom. Not a lock, not a preference signal. A standing offer — and the right question about one is not whether anyone has accepted, but whether it is the offer you want accepted.

For anyone whose content exists to be quoted, that has an uncomfortable edge. Most of a link-earning estate should be priced at zero with conditions attached, because its return arrives as corroboration rather than royalties — and the assets you would genuinely charge for are usually the ones nobody is taking. Price in three tiers, publish per path, and stop reading a licence file as a fence.

1. The question everyone is asking about RSL is the wrong one

Really Simple Licensing launched on 10 September 2025 with backing from Reddit, Yahoo, People Inc., Ziff Davis, Fastly, Quora and Medium; version 1.0 followed on 10 December 2025. By the end of that month Press Gazette reported roughly 1,500 media organisations endorsing it, and by April 2026 the RSL Collective told a WAN-IFRA audience it represented Wikipedia, Reddit, Yahoo, Torstar and USA Today, putting the prize at a $100 billion opportunity.

And almost every serious write-up ends in the same place. TechTarget, in November 2025, put it plainly: no legal enforcement mechanism, no law requiring an AI company to respect RSL terms. Speaking to Digiday that month, co-founder Doug Leeds did not dispute it — he assigned enforcement to other parties entirely: content delivery networks, lawsuits and government bodies. Eric Aledort of Arena Group conceded on the record that without a way to compel compliance, participation might not amount to much.

That candour is to everyone’s credit. It has also set the terms of the debate, and the terms are wrong. The industry has settled on one scoreboard — has a major lab signed up? — and it measures the part of the standard its own authors marked optional.

What is Really Simple Licensing?

RSL is an open, machine-readable vocabulary for attaching licensing terms, permitted uses and compensation models to specific paths on a website. It extends the yes/no logic of robots.txt into a language expressing who may use a page, for what purpose and at what price. It is free for any site to adopt.

Mechanically it is XML. A <content url> element scopes a licence to a URL or path, using the wildcard syntax from RFC 9309, the robots exclusion standard. Inside it, a <license> element carries <permits> and <prohibits> clauses for usage, user type and geography, plus a <payment> element naming the compensation model. Crawlers discover it via a License: line in robots.txt, a file at /license.xml, an HTTP Link header, or a link element in the head.

The required core, and the optional half

Read the 1.0 specification and the structure is explicit. The declarative XML vocabulary and the discovery mechanisms are the required foundation of any conformant implementation. Three further protocols sit alongside them, marked optional: the Open License Protocol, an OAuth 2.0 extension for acquiring and validating licence tokens; the Crawler Authorization Protocol, an HTTP authorisation scheme under RFC 7235 in which a crawler presents a token and receives 401 or 402 without one; and the Encrypted Media Standard, for locking non-public assets. The spec states that implementers may adopt RSL as a licensing and discovery framework without implementing any of them.

So the machinery everyone is grading — tokens, payment rails, the enforcement handshake — is explicitly severable. What remains is a way of saying, in a form a machine can parse, what you will accept for each part of your estate: smaller than a licensing regime, a different thing, and available now at no cost.

Why adoption is the wrong scoreboard

Nobody grades a price list by whether anyone has bought yet; its first job is to exist before the buyer arrives. Judging RSL on lab adoption in its first eighteen months is a category error of that shape, applied inconsistently — the alternative most often recommended in its place, hard blocking via robots.txt, runs on the same voluntary logic, and around 12.9% of bots ignored robots.txt as of Q1 2025, up from 3.3%.

The useful question is what the declaration layer does alone — which means being precise about what kind of object an RSL file is, because it is not the kind the fundamentals of how links and permissions interact would lead you to expect.

2. Not a lock, not a signal: RSL is a standing offer

Four vocabularies now say roughly the same thing

Declaring a preference about AI use is no longer scarce. Cloudflare published its Content Signals Policy on 24 September 2025 under CC0, defining three signals — search, ai-input and ai-train — written into robots.txt with a preamble framing them as a condition of access. In July 2026 it added a fourth, use, prepending use=reference to the managed robots.txt of every customer on its defaults. The W3C has TDMRep, a reservation manifest at /.well-known/tdmrep.json. HTML meta tags and X-Robots-Tag headers carry noai. The IETF has chartered an AI Preferences working group to standardise the vocabulary properly.

The one thing only RSL says

Every vocabulary in that list answers the same question: may you? RSL answers a different one: at what price, paid to whom, on what terms? It is the only machine-readable format in general use that converts a permission question into a commercial proposition, because it is the only one carrying a <payment> element and a named payee.

That difference has a consequence people keep skipping. A preference cannot be complied with in a way that creates an agreement — Cloudflare’s own text says silence neither grants nor restricts permission. An offer can be accepted. Whether an acceptance forms an enforceable contract is contested, and section 8 gives it the space it deserves; the structural point stands either way. A preference has no acceptance state. An offer does.

So RSL is neither of the two things the industry keeps calling it. Not a lock, because the locking protocols are optional and unimplemented at scale. Not a signal, because signals were already free and RSL costs more than they do. It is a standing offer: published, dated, specific, addressed to the world, open until withdrawn.

What an unaccepted offer still does

Four things, none of which require a counterparty to say yes.

It sets a reference price. A publisher with a dated, published number opens negotiations there; one without opens wherever the buyer starts. Anchoring is not a legal mechanism. It is a commercial one, and free.
It creates notice, on a date, in machine-readable form. Under Article 4(3) of the EU Copyright in the Digital Single Market Directive, the text and data mining exception applies only where rights have not been expressly reserved in a machine-readable manner, and Article 53(1)(c) of the EU AI Act requires general-purpose model providers to run a copyright policy respecting those reservations. There the declaration is not a wish addressed to a company but a trigger condition in a statute — and the compliance obligations the AI Act places on content use give an otherwise inert file somewhere to bite.
It delegates enforcement to infrastructure. Cloudflare, Akamai and Fastly have all signed up to use RSL to validate crawlers, per the RSL Collective’s Nick Boyden at WAN-IFRA in April 2026 — which matters more than lab adoption, because it moves the check from the counterparty’s conscience to your own edge.
It is the artefact any collective body needs to act for you. A rights organisation cannot pool what has not been declared. Whatever one thinks of the ASCAP analogy, pooling needs per-asset terms expressed identically across members.

Key takeaway

RSL keeps getting graded as a failure because it is being graded as enforcement, which is the optional half. The required half is a priced, dated, per-path declaration — useful with zero counterparty adoption, and emitted by no other vocabulary in the field.

3. The Standing-Offer Map: pricing an estate that contains two businesses

Once the file is an offer rather than a fence, the operational question changes. It stops being “should we allow this?” and becomes “what would we accept, for which pages, from whom?” That cannot be answered once for a whole domain, because a site built to earn links holds assets with opposite economics in the same file system. A guide written to be cited and a dataset sold by the seat are not the same kind of property: one is an advertisement whose return is corroboration, the other inventory whose return is revenue. Almost every strategic approach to earning links depends on the first being consumed as widely as possible, and almost every licensing instinct pushes toward restricting it.

Tier 1 — The Shop Window — priced at zero, on purpose

What sits here: Guides, glossaries, methodology explainers, definitional pages, free tools, statistics round-ups — anything built to be quoted.

The offer: Free. Attribution required. search and ai-summarize both permitted. Training usually permitted, since a page defining your category benefits from being in the weights.

Why: The return here arrives as corroboration, not royalties. Restricting summarisation is charging admission to your own advertisement.

Tier 2 — The Counter — free to read, priced to reuse

What sits here: Original research, proprietary benchmarks, survey data, interactive calculators and other assets built to attract links at volume — anything with real acquisition cost you still publish openly.

The offer: Free for search and reference with attribution required; a named price for training and inference. This is the tier where a number belongs.

Why: You want it cited and you want the citation to name you, but bulk reuse without attribution turns an asset you paid to build into a commodity input. A price distinguishes the two.

Tier 3 — The Vault — no free tier at all

What sits here: Subscriber content, paywalled databases, client deliverables, raw data files — anything already behind authentication.

The offer: Licence required, no free use. Encryption or authentication enforced at the server, with the licence describing terms rather than substituting for the lock.

Why: The only tier where a licence does the job licences were designed for, and the only one where refusal costs nothing in visibility.

Why the split has to be per path

This is the mechanical reason RSL is worth the effort over the simpler vocabularies. Cloudflare’s content signals attach to a robots.txt user-agent group; TDMRep attaches to a manifest. Neither expresses two commercial positions for two parts of one domain with any precision. RSL’s <content url> element does, using ordinary path patterns: /guides/* gets the shop-window licence, /research/* the counter licence, /data/* the vault licence, in one file.

Uniform pricing across non-uniform assets is the central error, and it runs both ways. A site-wide restrictive licence prices the shop window like the vault and quietly withdraws material you spent years promoting. A site-wide permissive licence gives away the only thing anyone might have paid for. The granularity is the product.

What most sites are about to publish instead

The tooling defaults point the wrong way for anyone whose model depends on being repeated. Plugins implementing RSL and content signals ship with restriction recommended and applied site-wide, because that is the only setting a plugin can safely guess at — a licence written by a template, published across an estate the template has never seen, then forgotten. Doing it properly takes an afternoon of ordinary crawl and header work.

4. The vocabulary, and the two words that decide everything

Those decisions are expressed in the standard’s own terms, and RSL’s usage vocabulary is more granular than most people realise. The values a licence can permit or prohibit include all, train-ai, train-genai, ai-use, ai-summarize and search.

Usage value	What it governs	Who should care most
search	Indexing and retrieval for a search result or link.	Everyone. This is the value almost nobody should restrict.
ai-summarize	Composing a summary or answer from the page.	Anyone whose citations come from answer engines. The decisive one.
ai-use	General use of the content by an AI system at run time.	Broader than summarisation; covers agent reading and tool use.
train-ai	Inclusion in training data for machine learning generally.	Anyone worried about substitution rather than attribution.
train-genai	Inclusion in training data for generative models specifically.	The narrower training carve-out most publishers actually mean.

Search yes, summarise no: the setting that quietly costs you citations

The pairing that matters is search against ai-summarize, because they are performed by the same fetch and separated only by intent. A page can permit indexing and prohibit summarisation: it returns 200 to every crawler you test, appears in every index you check, passes every crawlability audit — and the sentence naming you is still off-limits to the system composing the answer.

That is a use condition, and no technical audit detects use conditions, because there is nothing to detect at fetch time. It is why Cloudflare shipped its use signal in July 2026: the original three-way split assumed reading a page for search and reading it to compose a reference were the same act. Anyone who has worked on getting content extracted into answer boxes and snippets knows how far the gap between indexed and quoted can run. RSL makes it a term of licence rather than an accident of ranking.

The payment types, and what pay-per-inference is really asking for

What is pay-per-inference? RSL supports free, attribution, subscription, purchase, royalty, pay-per-crawl and pay-per-inference. Pay-per-inference pays a publisher each time an AI system uses their content to generate a response, rather than each time it fetches the page: pay-per-crawl bills the access, pay-per-inference bills the use.

The distinction matters because crawling and using have decoupled. More than half of AI crawler requests re-fetch pages that have not changed, and Cloudflare Radar reported in June 2026 that automated traffic had reached 57.5% of HTML requests against 42.5% human. Billing the fetch bills noise; billing the use bills value — which is where the economics of agentic browsing and what a machine visit is worth keeps landing too.

There is a catch no publisher should paper over. Pay-per-inference asks to be paid on a metric only the counterparty can observe: you cannot see which paragraph shaped which answer, the operator can, and reports it to you. That is a governance dependency — a royalty settled on the other side’s telemetry, with no independent audit trail. An analysis of AI content-access infrastructure published on arXiv in 2026 makes exactly this criticism: RSL addresses policy declaration and provides no audit ledger, no provenance proofs and no way to verify a given output used your content.

Key takeaway

Set search and ai-summarize deliberately and separately on every path. Permitting the first while prohibiting the second is invisible to every technical check you run and is the single cheapest way to disappear from answers while appearing perfectly healthy in your crawl reports.

5. The link publisher’s RSL file, line by line

Here is the three-tier map as a working file. Each element is a commercial decision in technical clothing; the annotations identify which.

<!– TIER 1: the shop window –>

<permits type=”usage”>search,ai-summarize,ai-use,train-ai</permits>

<legal type=”warranty”>ownership</legal>

</license>

</content>

<!– TIER 2: the counter –>

<permits type=”usage”>search,ai-summarize</permits>

</license>

<permits type=”usage”>train-genai</permits>

</license>

</content>

<!– TIER 3: the vault –>

</license>

</content>

</rsl>

The path patterns are the strategy. Three <content> blocks, three prices. A file with one block covering / is not a pricing decision, it is a mood.
Tier 1 permits training on purpose. A page whose job is to establish what your category means, and who defines it, is worth more inside a model than outside one — the same logic behind treating inclusion in training corpora as a deliberate content strategy rather than a leak to be plugged.
Tier 2 carries two licences on one path. Free for reading and answering, priced for training. Multiple <license> elements under one <content> element let you sell one use while giving away another.
The number in the training licence does the real work. It is almost certainly wrong. It is also the only thing in the file a buyer can respond to, and writing it down forces a conversation most publishers have avoided for two years.
Tier 3 prohibits everything and relies on the server. The licence describes the terms; authentication enforces them. Never let a declaration stand in for a lock.

Where the file goes, and how crawlers find it

Publish it in more than one place, because a crawler ignoring one route may honour another. Put the document at /license.xml. Add a License: line to robots.txt — an RSL extension rather than part of RFC 9309, so it is inert to anything not taught to read it. Emit an HTTP Link header with rel=”license” and type=”application/rsl+xml”. A morning of work, and most auditing tools will not check any of it for you yet.

What not to put in it

Rights you do not hold. Guest contributions, licensed imagery, syndicated feeds and user comments are frequently outside your grant. Reserving rights over them is at best noise, at worst a representation you cannot support.
A price you would not honour. If a buyer accepts your number tomorrow and you feel sick, the number was theatre.
Prohibitions on pages you are actively promoting. The most common self-inflicted wound in the exercise, usually inherited from a plugin default rather than chosen.

6. Your backlink profile is a portfolio of other people’s licences

Everything above concerns pages you own — for anyone building links, the smaller half of the problem. Ahrefs, studying 75,000 brands, found mentions correlating with AI visibility at 0.664 against 0.218 for backlinks; Lantern found roughly 85% of brand mentions in AI answers grounded in third-party sources; Muck Rack’s May 2026 data put earned coverage at 84% of cited sources against 0.3% for paid. The corroboration doing the work is not on your domain.

So the terms governing it are not yours either. Every page mentioning you is licensed by its owner, under a file you never saw. A backlink profile has quietly become a portfolio of third-party licences, and the standard competitive backlink analysis workflow has no column for it.

The screen a crawlability check cannot see

Consider a placement in a well-regarded trade title. It returns 200, it is indexed, your backlink tool counts it and your link velocity reporting logs it. And the host’s licence permits search and prohibits ai-summarize — so the paragraph naming you sits on a page an answer engine may index and may not quote.

No fetchability audit catches this, because nothing fails. The distinction between reachable and usable is new, it lives in a file most SEO tools do not parse, and it governs the pages carrying your most valuable corroboration.

What this changes in digital PR

Not the target list, but the arithmetic behind it. A placement now has a terms component alongside its authority component, and the two are uncorrelated. BuzzStream’s early-2026 audit found 79% of top news sites blocking AI training bots and 71% also blocking retrieval crawlers — so the ad-funded prestige tier, the traditional prize of a digital PR programme, is the most restricted, while institutional hosts, trade bodies and the kinds of sponsorship placements that sit on non-commercial domains are the least.

The response is not to abandon national coverage, which still earns referral traffic, brand demand and secondary pickup. It is to stop assuming a placement’s authority transfers automatically into an answer. Two placements with identical domain metrics can have opposite citation value, and the variable separating them is a licence file — as true of platform placements as editorial ones. A listicle placement built for AI citation sits under its host’s terms; a link earned on a community platform like Hacker News sits under that platform’s.

The standing-offer test

Run this on every path before you publish anything, and on any placement you are about to pay for.

1. What do you want to happen to this page inside an AI answer? Three honest options: be repeated, be repeated with attribution, or not be repeated. Most estates contain all three and are about to publish one.

2. Would you accept nothing for that outcome? If yes, your price is zero and the file should say so loudly, with attribution as the condition. Silence is not the same as free, and it is read differently.

3. If no, name the number and the unit — per crawl, per inference, per year. If you cannot name the number and the unit, you do not have a price. You have a grievance.

7. Worked example: pricing an estate that exists to be repeated

Harptree Analytics is an invented but deliberately ordinary case: a Bristol research publisher covering UK commercial insurance, £6.8m revenue, 45 staff, three asset classes. A 900-page free editorial estate. The Harptree Premium Index, a quarterly benchmark of commercial premium movement, published free and responsible for 340 referring domains in three years. And a subscription database at £4,200 per seat, 610 seats.

The trigger is a January 2027 board paper. Marketing can show the Index number being restated in AI answers, correctly, with no attribution. The board’s instinct: put a pay-per-inference licence on the Index and make them pay for it.

What the four-week audit found

The most-extracted thing they own is the least protected thing they own. The Index number is a fact. UK law gives no copyright in a fact; the database right under the Copyright and Rights in Databases Regulations 1997 protects substantial investment in obtaining, verifying and presenting a database’s contents, not the figure. A licence claims expression and access, and does not reach the number — which is precisely what travels.
Their own policy was set by two parties, neither of them Harptree. Their robots.txt carried a managed CDN ruleset, and a plugin installed in 2025 by a departed contractor was emitting a site-wide noai tag on every page — including the Index landing page they spend £180,000 a year promoting.
Their earned estate ran on six regimes at once. Of the 60 referring domains contributing the most tracked answer citations, 14 published RSL or content-signal terms: five permitted search and prohibited summarisation, six were free with attribution, three required a licence for any AI use. A further 23 were ad-funded pages behind a CDN default due to flip in September.

What they changed

Three moves, £112,000 reallocated, no new headcount.

A three-tier RSL file, published in week three. Shop window on /guides/* and /glossary/*: free, attribution required, summarisation and training permitted. Counter on /research/* and /index/*: free for search and summarisation with attribution, £14,000 a year named as the standing training price. Vault on /data/*: licence required. Roughly £6,000 of developer time, and the blanket noai tag removed the same week.
A canonical fact page for the Index. One stable URL, one number per quarter, dated, methodology beside it. The reasoning was explicit: if you cannot licence a fact, be the cheapest and most obviously authoritative place to obtain it.
£70,000 of the £310,000 digital PR budget reallocated. Two prestige national targets moved out of the citation-supply line into the brand line, where their referral value is real; four trade and institutional hosts whose terms permit summarisation moved in. Nothing was cancelled — the accounting changed, and with it the pitch priority.

Month nine

Royalty revenue from RSL: nil. No operator accepted the standing offer — the expected outcome, written into the board paper as such.
One inbound, from an unexpected direction. A data-licensing intermediary found Harptree’s terms in month five while crawling published RSL files and opened a conversation about the subscription database. It closed in month eight at £31,000 for a one-year training licence — more than twice the number in the file, which Harptree only learned was low because the file had forced them to write one down.
Citation share moved, but not because of the licence. Tracked answers naming Harptree as the source for commercial premium movement rose from 9% to 24% of measured mentions, traced to the canonical fact page and the reallocated placements. Referring domains rose 41, of which 29 came from the four new institutional hosts. The licence contributed nothing measurable to any of it.

The line the finance director used

“The licence earned us nothing and taught us our own price. We had spent two years arguing about whether to charge without ever writing down a number.”

8. Where this argument is weakest

The strongest objection is not that RSL is unenforceable. It is that this article has dressed impotence up as strategy, and it runs like this.

“A standing offer nobody can accept is not an offer, it is a wish with XML around it. Contract formation requires acceptance and consideration; a crawler that ignores your file has accepted nothing, and no court has held that fetching a page under browsewrap-style conditions forms a licence. Meanwhile the leverage that has actually moved money came from the opposite behaviour — litigation and hard refusal. The Associated Press and the New York Times did not reach their positions by publishing price lists. And pricing your shop window at zero is precisely the conclusion the AI companies would write for you: unilateral disarmament with a spreadsheet attached.”

That is a fair statement of the case, and the first half is largely correct. Four bounds, in descending order of confidence.

The contract route is weak; the statutory route is not. Browsewrap enforceability against a crawler is untested and may fail. The EU pathway needs no contract: Article 4(3) of the DSM Directive makes a machine-readable reservation the condition on which the mining exception fails, and the AI Act puts the obligation to respect it on the model provider. That is regulation, not agreement. The UK’s position is narrower — section 29A of the Copyright, Designs and Patents Act 1988 covers non-commercial research — so a UK publisher’s reservation does most of its work against operators serving EU users, which matters to anyone weighing how content rules differ across European markets.
Litigation leverage requires a litigable corpus. AP and the NYT hold decades of distinctive reporting and a legal budget — perhaps a few hundred organisations worldwide. Not a 900-page marketing estate whose questions a competent model could answer from twenty other sources. For substitutable estates the litigation route is unavailable at any price.
Disarmament assumes extraction is a loss. On the shop window, extraction is the product: the page exists to be repeated, and pricing repetition at zero is correct pricing of a promotional asset, not surrender. The disarmament error is pricing the vault at zero too — exactly what a site-wide permissive licence does, and why the per-path split is the whole argument.
RSL lacks what a working royalty regime needs. The arXiv analysis cited earlier is right that there is no audit ledger, no provenance tracking and no way to prove a given output drew on your content. That caps pay-per-inference today, and it is why the CDN-validated path is more realistic than the token protocols. Publishing a price does not make the metering exist.

What would prove this framing wrong?

Two things, pointing in opposite directions. If a court in the UK, EU or US holds that a machine-readable licence presented at fetch time forms an enforceable agreement with a crawler, RSL stops being an offer and becomes a lock — and the advice inverts, because a permissive shop-window licence would then bind rather than promote. Conversely, if the IETF’s AI Preferences work lands and the major operators implement that vocabulary instead, RSL’s declaration layer becomes redundant, only the Collective’s pooling function survives, and the advice shrinks to: join the pool, skip the XML.

9. What to do on Monday

Pull your robots.txt, any /.well-known/tdmrep.json, and the head of five representative pages. Find out what you are declaring and who decided it. Assume a plugin or a managed CDN ruleset until proven otherwise.
Sort your estate into the three tiers by path pattern. An afternoon with a sitemap is enough for most sites.
For every path in the counter and vault tiers, write a number and a unit. If you cannot, it belongs in the shop window until you can.
Publish the file at /license.xml, add the License: line to robots.txt, emit the Link header, and verify ai-summarize is permitted on every page you are actively promoting.
Check your 30 highest-value referring domains for RSL or content-signal terms and record which permit summarisation. That column belongs in your prospecting sheet permanently, beside vertical fit and audience relevance.
Identify the one fact your organisation is best known for and build the canonical page for it — stable URL, dated figure, visible methodology. You cannot licence a fact, so own its provenance.
Diary the file for annual review and log the publication date. The date is part of the instrument.

The part that outlives RSL

RSL may not be the standard that wins. The IETF process may supersede it, the Collective may never assemble the bargaining power it is aiming at, the token protocols may never see real traffic. None of that touches the underlying practice, older than any of these vocabularies and likely to outlast all of them: every machine-readable statement you publish about your own content is read by more parties than the one you wrote it for. Robots.txt was written in 1994 to stop servers falling over; inside a decade its main use was commercial.

The durable behaviour follows. Keep a published, dated, per-asset price, revise it once a year, and read the absence of buyers as information about your price rather than about the market’s morality — the discipline every serious measurement habit in this field eventually forces on you.

The version for anyone whose income depends on being cited is sharper. The part of your estate that earns your links should be the cheapest thing you own, deliberately — priced at zero, conditioned on attribution, permitted for summarisation, promoted like the advertisement it is. Charge for the vault. Publish the window. The failure mode is not generosity or restriction; it is one price applied to an estate holding two businesses. Nobody has ever been paid for content they refused to price.