TL;DR
In 2026 the strategic question for content owners flipped from “how do I keep AI off my site” to “how do I become a source the models can’t do without.” Reddit turned its forum archive into roughly $130m a year in AI licensing revenue — about 10% of its top line — by being uniquely structured, fresh and hard to replace.
Stack Overflow is the cautionary half of the model: the same Q&A corpus that made it priceless to train on now answers the questions that built it, and monthly question volume has collapsed from ~200,000 to under 50,000. Being a great training source can quietly cannibalise the thing that made you valuable.
This is a UK-focused ai training source strategy: a 7-factor Training-Source Value Scorecard you can run on your own content this week, the four ingredients every preferred source shares, the new licensing plumbing (RSL, pay-per-crawl, llms.txt), and what the UK’s abandonment of the copyright “opt-out” means for British publishers and brands.
For two years the defensive instinct ruled. Publishers reached for robots.txt, flipped the “block AI scrapers” toggle their CDN helpfully added, and hoped the models would starve. They did not starve. They got richer, and the sites that blocked everything mostly succeeded in making themselves invisible in the one place discovery is now growing: AI answers. The smarter players noticed something else. The companies building frontier models were no longer trying to vacuum the whole web for free. They were paying — selectively, and a lot — for a small number of sources that were structured, current, clean to licence and genuinely hard to reproduce.
That is the shift this guide is about. An ai training source strategy treats your content not only as a thing humans read, but as a supply input into a data economy with real prices attached. Reddit and Stack Overflow are the two canonical case studies — one a triumph, one a warning — and between them they teach almost everything a UK content owner needs to know about how to become a source the models prefer, and how to avoid licensing away your own future. If you are new to why any of this still touches links and rankings, our primer on what backlinks actually are and how authority flows is the foundation the rest of this builds on.
What you’ll take away:
- The data-supply economy in numbers — who is paying what, and why the deals moved from “training dumps” to live feeds.
- The Training-Source Value Scorecard: a 7-factor rubric (scored 0–14) you can run on your own corpus before lunch.
- What Reddit got right — and the Stack Overflow paradox that should scare anyone licensing their only moat.
- The four ingredients of a preferred source, scaled down for UK SMEs and publishers without 100 billion comments.
- The UK legal reality in 2026: the opt-out is dead, licensing is the default, and what that means for your leverage.
- The mechanics — RSL, pay-per-crawl (HTTP 402), llms.txt and the AI-crawler controls — with a copy-ready config block.
- A 30/60/90-day plan, and how all of this compounds with ordinary link earning.
1. The data-supply economy, in numbers
Start with the figure that reframed the whole debate. On the same day it filed to go public in February 2024, Reddit announced a content-licensing deal with Google reportedly worth $60m a year; a few months later it signed OpenAI for an estimated $70m. By early 2025 the company’s COO confirmed those AI deals made up about 10% of Reddit’s ~$1.3bn revenue — roughly $130m a year for letting models drink from a forum its users wrote for free. Bloomberg later put Reddit’s total disclosed licensing contract value north of $200m.
Reddit is not an outlier; it is the leading edge of a market. Through 2024–26, AI labs signed dozens of publishers at price points ranging from roughly $5m to $60m a year — Axel Springer, the Associated Press, the Financial Times, Dotdash Meredith, Informa, Shutterstock. News Corp’s OpenAI agreement was reported at more than $250m over five years. And the price of not licensing got a number too: Anthropic’s $1.5bn copyright settlement in September 2025 — about $3,000 per book across roughly 7 million titles — is now the reference benchmark every legal team quotes.
The deals moved from “buy once” to “rent continuously”
The most important structural change is what the money now buys. Early deals were weighted toward training — a one-off licence to ingest a corpus. The newer wave is weighted toward real-time access: continuously refreshed feeds the model grounds its live answers on, with attribution back to the source. In one analysis of 2025 contracts, around 77% were real-time or attribution-type rather than pure training licences; The Guardian’s and The Washington Post’s 2025 OpenAI deals notably emphasised summaries, quotes and links over training rights. This matters for strategy: a corpus you can keep fresh is an annuity; a static archive is a one-time cheque.
The headline numbers, in one place:
| Source / deal | Reported value | Structure | Year |
| Reddit — Google | ~$60m / year | Data-API access | 2024 |
| Reddit — OpenAI | ~$70m / year (est.) | Real-time feed | 2024 |
| News Corp — OpenAI | >$250m / 5 years | Training + access | 2024 |
| Industry band (AP, FT, Axel Springer, etc.) | ~$5m–$60m / year | Mixed | 2024–26 |
| Anthropic — authors (settlement) | ~$1.5bn (~$3k/book) | Cost of NOT licensing | 2025 |
Figures are as publicly reported; many deal terms are undisclosed or estimated. The pattern, not any single number, is the point.
Why quality and uniqueness command the premium
There is a supply-side reason the cheques got bigger. The first generation of models was trained on whatever could be scraped — Common Crawl, the open web, a lot of low-grade duplication. That well is now largely tapped, and labs have run into what researchers call the “data wall”: the highest-value text has already been ingested, and scraping more of the same web adds little. Synthetic data helps with volume but tends to amplify a model’s existing blind spots rather than teach it anything new. What moves a model forward now is data it cannot get anywhere else — real human conversation, current events, expert Q&A, proprietary records, niche local detail. That is precisely the data a handful of sources hold, and precisely why those sources, not the generic web, are the ones getting paid. Your job is to be on the short side of that scarcity.
One more number frames the urgency for publishers weighing licensing against traffic: AI summaries are estimated to have cut publisher click-throughs by roughly 80%. The old web bargain — content in exchange for referral traffic — is breaking. Licensing revenue is, for many sites, the replacement bargain. For the full picture of how AI visibility now interacts with rankings, see what to do when you stop getting cited by AI.
2. The Training-Source Value Scorecard
Before any licensing tactics, answer one question honestly: is your content actually valuable to a model? Most isn’t — not because it’s bad, but because it’s replaceable. AI labs pay for data that improves the model or its answers in ways cheaper data can’t. Score your corpus across seven factors, 0–2 each, for a maximum of 14. Run it before you spend a penny on licensing infrastructure.
| Factor | Score 2 if… | Max |
| 1. Corpus uniqueness | The data exists nowhere else at this scale — first-party experience, proprietary records, or community knowledge no competitor holds. | 0–2 |
| 2. Structure & parseability | Content is chunked, labelled and consistent (Q&A pairs, tables, schema, clean headings) so a model can extract answers cheaply. | 0–2 |
| 3. Freshness / refresh rate | You produce new, dated material continuously — a live feed is worth more than a frozen archive. | 0–2 |
| 4. Signal-to-noise | Moderation, voting or editorial review means the average item is trustworthy; low spam, low duplication. | 0–2 |
| 5. Licensable rights | You hold clean, enforceable rights to licence it — no murky UGC ownership, no third-party content you can’t grant. | 0–2 |
| 6. Demand alignment | Your domain maps to where models are weak and queries are hot (niche expertise, current events, code, local data). | 0–2 |
| 7. Defensibility / moat | Your value can’t be fully captured by ingesting you once — there’s a flywheel that keeps producing what the model needs. | 0–2 |
How to read your score
11–14 — Genuine supply asset. You have leverage; pursue direct licensing and/or RSL terms (Section 7) and protect Factor 7 above all.
7–10 — Promising but leaky. Fix the weakest factor first — usually structure (2) or freshness (3) — before approaching anyone. Collective licensing fits you best.
Below 7 — Not a standalone training source yet. Your near-term play is AI visibility and earned links, not licensing revenue. Build a citable data asset first (Section 8).
Worked examples. Reddit scores close to the ceiling: a uniquely conversational corpus (2), inconsistent structure but improving (1), relentless freshness (2), heavy community moderation (2), licensable via its user terms (2), demand-aligned because models love real opinions (2), and a powerful flywheel — people keep posting (2). That is why it commands the prices it does, and why it is now reportedly the most-cited single domain in AI answers. Stack Overflow scores high on uniqueness, structure and signal — but, as we’ll see, near zero on Factor 7. The flywheel broke.
3. What Reddit got right (and is still fighting for)
Reddit’s playbook is almost a checklist for the rest of this guide. First, it stopped giving the asset away. In mid-2024 it blocked most automated crawlers, forcing AI companies to licence rather than scrape — and it has been willing to enforce, suing Anthropic in June 2025 for allegedly scraping the site after claiming to stop, and even restricting the Internet Archive to stop archived pages becoming a side-door. Scarcity created the negotiating table.
Second, it is repricing as its leverage grows. On its mid-2025 earnings call, Reddit pushed for “dynamic” compensation — pay that scales with how essential its data proves, tied to citations, conversions and even benchmark lifts, rather than a flat annual cheque. The flat → usage → dynamic progression is the template every content company will follow. Third, it is backing the standards — joining the Really Simple Licensing collective (Section 7) so the whole industry, not just Reddit, can charge.
Case study — Reddit (public)
Read the full account in Columbia Journalism Review’s “Reddit Is Winning the AI Game”. The transferable lesson: uniqueness plus enforceable scarcity plus a refresh flywheel = pricing power. Reddit didn’t out-write the web; it owned a corpus no one could rebuild.
There is a fourth lesson hiding in Reddit’s numbers — the discovery dividend. Being the licensed, trusted source doesn’t just pay directly; it makes you surface more often everywhere else. Reddit became the most-cited domain across AI answers, and a Google ranking change that elevated forums roughly tripled its readership in under a year. Licensing, citation and organic visibility turned out to be the same flywheel viewed from three angles: the more essential your data is to the models, the more they surface you, the more humans arrive, the more fresh content they generate, the more valuable your next licence becomes. That compounding is the real prize — and it’s why an ai training source strategy is never only about the cheque. It’s about becoming structurally hard to route around.
4. The Stack Overflow paradox: when being the best training source kills you
Stack Overflow had, arguably, the highest-quality training corpus on the open web for its domain: fifteen years of peer-reviewed, voted, structured programming answers. Models trained on it heavily. Then the bill came due. Once ChatGPT and its successors could answer coding questions directly — using knowledge distilled in large part from Stack Overflow itself — developers stopped visiting. Monthly question volume collapsed from a ~200,000 peak to under 50,000 by late 2025, with one month recording barely 6,800 — levels not seen since 2008. Questions were down roughly 76% since ChatGPT launched.
This is the paradox at the centre of any ai training source strategy. The very feature that made Stack Overflow valuable — a public archive of answers — also made it substitutable once ingested. Attribution inside a chatbot does not rebuild a community. And here is the part that should worry every would-be supplier: the asset is a flywheel that needs humans. If people stop asking and answering, the corpus stops refreshing, and the source that was once essential becomes a frozen, decaying archive — a Factor 3 and Factor 7 failure at once.
The company is not dead — Stack Overflow was sold to Prosus for $1.8bn in 2021 and has leaned into licensing and enterprise products, with reporting suggesting its licensing revenue has held up even as the public forum emptied. But the forum that created the value is a shadow of itself. The closest parallel is Chegg, whose business and share price were gutted once students realised an AI tutor was free. The lesson is blunt:
The Stack Overflow lesson
Do not licence away the thing that makes you necessary unless you protect the flywheel that keeps producing it. If your only value is a static archive, a one-time licence is the last big cheque you will write yourself — because once the model has the archive, it no longer needs you, and your audience no longer does either.
Look closely at the mechanism, because it generalises. Stack Overflow’s value depended on a loop: people hit a problem, asked, got peer-reviewed answers, and that exchange became the next person’s search result — and the model’s next training example. Break any link and the loop unwinds. AI broke the asking link: why post and wait when a chatbot answers instantly? Stack Overflow’s own 2025 developer survey, of nearly 50,000 respondents, found 84% now use AI tools in their workflow, with most leaning on GPT-class models. Fewer questions means fewer fresh answers, which means a staler corpus, which means the next model generation has less new signal to learn from — a slow-motion failure that hurts the source and, eventually, the models that hollowed it out. If you are a community-driven UK platform, this is the single risk to design against before you sign anything.
5. The four ingredients of a preferred training source
Strip Reddit and Stack Overflow down to mechanics and the same four ingredients appear. A preferred source has all four; a vulnerable one is missing the last.
- Proprietary or unique data. First-party experience, original research, community knowledge, or records nobody else holds. Re-stated public facts have no licensing value — the model already has them ten times over.
- Machine-parseable structure. Discrete, labelled chunks a model can extract cheaply: Q&A pairs, comparison tables, schema-marked entities, consistent headings. This is exactly why listicle placements are the structured format LLMs parse most easily — the same property that makes content citable makes it trainable.
- Continuous freshness. A dated, refreshing feed is an annuity; a frozen corpus is a depreciating asset. Models increasingly pay for the flow, not the archive.
- Clean, licensable rights. You can only sell what you can grant. Murky UGC ownership, embedded third-party content, or contributor terms that don’t allow sub-licensing will collapse a deal in due diligence.
The realistic version for UK SMEs and publishers
Most British businesses will never hold a Reddit-scale corpus — and don’t need to. The same model scales down. A regional recruiter holds proprietary salary data. A conveyancing firm holds anonymised timelines for thousands of UK property transactions. A trade supplier holds real failure rates by product. None of that exists elsewhere, all of it is structurable, and the freshest cuts of it are exactly what a model grounding a “typical cost / timeline / rate in the UK” answer needs. The move is to package it as a citable public number — the same discipline behind interactive calculators and data tools that earn 100+ links. A private answer (“your quote is £1,240”) earns nothing; a public finding (“the median UK figure is X”) is what both journalists and models reference.
UK SME shortcut
You do not need a billion comments. You need one dataset only you have, refreshed quarterly, structured cleanly, with rights you fully own. That is a licensable — and citable — asset. Original research is also the single most reliable way to earn the kind of editorial links that compound.
6. The UK legal reality in 2026: the opt-out is dead, licensing is the default
This is where the UK edition matters most, because Britain has spent 2025–26 having exactly this argument in public — and the outcome strengthens every content owner’s hand. Between December 2024 and February 2025 the government ran a consultation on copyright and AI. Its “preferred” option was a text-and-data-mining (TDM) exception that would let AI firms train on copyright works unless rights holders opted out. The creative industries revolted: the “Make It Fair” campaign, and more than 1,000 musicians releasing a silent album in protest.
The response was emphatic. Of 11,520 consultation submissions, 81% backed mandatory licensing and just 3% supported the government’s preferred opt-out. By January 2026 ministers conceded that expressing a preference had been a “mistake.” The Data (Use and Access) Act 2025 (Royal Assent June 2025) then forced the government to publish an economic impact assessment and a formal report by 18 March 2026. When that report landed, it abandoned the opt-out exception and adopted a “wait-and-see,” market-led licensing posture. In plain terms: in the UK, licensing is now the default expectation, not a broad free-training carve-out.
Three legal facts that change your leverage
- Licensing-first, by default. The House of Lords Communications and Digital Committee pushed a licensing-first framework, and the government declined to hand AI firms a free pass. Your content is not legally “free to train on” in the UK by statute.
- But enforcement has a territorial gap. In Getty Images v Stability AI, Getty dropped its primary-infringement claim and the High Court’s November 2025 ruling meant models trained overseas but deployed in the UK can be hard to pin down. Practical control — access restrictions, machine-readable terms — still matters more than waiting for the law.
- The EU is more directive — use it. Under the EU AI Act’s Article 53, general-purpose model providers must publish a summary of their training content (in force since August 2025), with fines up to €15m or 3% of global turnover. If you publish into the EEA too, that transparency is leverage in any licensing conversation — and it mirrors the GDPR-aware care UK agencies already apply across European markets.
UK rights-reservation decision tree
1. Do you hold a Scorecard ≥ 11 corpus? → Reserve rights, restrict TDM in your terms, and approach labs directly or via the RSL Collective.
2. Score 7–10 with clean rights? → Join collective licensing (RSL) so you negotiate as a bloc, not alone.
3. Score < 7, or rely on others’ content? → Don’t chase licensing yet. Set sane access controls, keep AI-search crawlers allowed for visibility, and build a citable asset.
4. In every case: update website terms to state your TDM position, log crawler behaviour, and keep records — the practical self-help every UK adviser now recommends.
Where UK policy is heading — and why it favours you
The direction of travel is consistent. The government’s December 2025 progress statement reported that 88% of respondents favoured stronger copyright and licensing frameworks, and the four official working groups are focused on transparency, technical standards, licensing and creator remuneration — not on a free-training exception. Expect three things to firm up over the next legislative cycle: a transparency obligation (forcing labs to disclose, at least in summary, what they trained on — the UK echoing the EU’s Article 53), open technical standards for rights reservation and provenance (work like C2PA content credentials is already being adopted), and facilitated collective licensing so smaller creators can negotiate together. Each of those increases an individual content owner’s leverage. The smart move is to be standards-ready now rather than scrambling when disclosure becomes mandatory.
Worth noting too: some AI firms are experimenting with revenue-sharing rather than flat fees — Perplexity launched a $42.5m publisher revenue-share programme, tying payment to publisher content used in answers. For a mid-sized UK publisher, usage-based and revenue-share models can ultimately beat a one-off licence — provided you have the structured, fresh corpus to keep getting used.
7. The mechanics: RSL, robots.txt, pay-per-crawl and llms.txt
The old robots.txt gave you exactly two settings: allow or disallow. That binary can’t express “yes, for a fee” or “yes, with attribution and a royalty.” Three overlapping mechanisms now fill the gap. Getting these right is a technical-SEO job — crawlability, indexation and access control — so loop in whoever owns your infrastructure.
Really Simple Licensing (RSL)
RSL is an open, machine-readable licensing standard that reached version 1.0 in late 2025 with backing from 1,500-plus brands including Reddit, Yahoo, Medium and Quora. Instead of “don’t crawl this,” you declare “crawl it under these terms” — attribution, subscription, pay-per-crawl or pay-per-inference (you’re paid each time a model uses your content in an answer). The RSL Collective pools publishers into a single bargaining bloc, explicitly modelled on music-royalty bodies like ASCAP and BMI — the small-publisher route to terms you’d never win alone.
Illustrative RSL declaration (paid AI-training licence) — conceptual, not production config:
<rsl xmlns=”https://rslstandard.org/rsl”>
<content url=”/”>
<license>
<permits type=”usage”>ai-train</permits>
<payment type=”subscription”>
<custom>https://yoursite.co.uk/ai-licence</custom>
</payment>
</license>
</content>
</rsl>
Reproduce the spec exactly from rslstandard.org before deploying; the above shows shape, not syntax to paste.
Pay-per-crawl (HTTP 402)
Cloudflare — which sits in front of roughly a fifth of the web — revived the dormant HTTP 402 “Payment Required” status code so site owners can charge AI crawlers per request, or block training crawlers while letting AI-search crawlers through for visibility. With AI bot traffic up roughly 300% year on year, the CDN layer is where most non-enterprise publishers will actually enforce terms. The major declared crawlers you’ll be writing rules for: OpenAI’s GPTBot, Anthropic’s ClaudeBot, Google-Extended, PerplexityBot and Common Crawl’s CCBot.
Illustrative robots.txt stance — monetise training, keep AI-search visibility:
# Block pure training crawlers (negotiate licence instead)
User-agent: GPTBot
Disallow: /
User-agent: CCBot
Disallow: /
# Allow AI-search crawlers that drive citations + referral
User-agent: OAI-SearchBot
Allow: /
# Point machine clients at your licence terms
# (RSL / llms.txt referenced separately)
Verify current user-agent strings before deploying — crawler names change, and a wrong rule either leaks content or kills your AI-search visibility.
llms.txt and the enforcement gap
llms.txt is an emerging convention — a plain-text map pointing AI clients at your most important, cleanest content. Treat it as a discovery and hygiene signal, not a licensing gate or a ranking factor; no major model currently uses it as an access control. And the honest caveat across all three mechanisms: enforcement still depends on cooperation. As of early 2026, no major AI company had formally agreed to honour RSL, and badly-behaved scrapers ignore robots.txt routinely. The teeth come from infrastructure (CDN-level blocking, Web Bot Auth crypto-authentication) and, ultimately, the courts — which is why the legal posture in Section 6 and the technical posture here have to work together.
Where the money flows next
A layer of intermediaries now sits between you and the labs: TollBit and ScalePost operate bot “paywalls” and AI-content marketplaces; ProRata.ai tracks attribution and shares revenue on actual usage; Cloudflare acts as clearing house for pay-per-crawl. You don’t have to pick a winner today — but you should know the market exists, because it means even a Scorecard-9 publisher has a route to revenue that didn’t exist 18 months ago.
8. How this connects to link building (and why it compounds)
Becoming a training source and earning links look like different games. They aren’t — they’re fed by the same asset. The content that models pay to ingest is, overwhelmingly, the content that journalists, bloggers and AI answers cite: unique data, structured clearly, kept fresh. Build that once and it works three ways at once — it earns the editorial links behind the 15 link building strategies that still move rankings, it gets you quoted in AI answers, and it becomes a licensable supply input.
The convergence is sharpest with original research. A proprietary UK dataset, turned into a public number, is simultaneously a link magnet and a training asset. The freshness that models reward is the same freshness that powers real-time newsjacking, and the structure that makes content trainable is the same structure that wins featured snippets and AI Overviews. If you want the hard evidence on how links and citations still drive visibility, our 2026 link building statistics keep the numbers current.
The compounding asset
One proprietary, structured, refreshed dataset = (1) editorial backlinks, (2) AI citations, (3) licensable training supply. Audit which of these three you’re currently leaving on the table — most UK businesses are capturing one of the three from an asset that could deliver all three. The right tools make the data-and-outreach workflow repeatable.
9. Five mistakes that turn a training asset into a liability
The downside cases are predictable, which means they’re avoidable. These are the five that recur most often when a content owner tries to play the supply game without thinking it through.
- Blocking everything. The blunt “ban all AI bots” toggle protects your archive and erases you from AI answers at the same time — because most systems use the same or similar user agents for live retrieval as for training. You want to charge or block training crawlers while keeping AI-search crawlers welcome. Blanket blocking is how sites quietly disappear from the new discovery surface.
- Licensing a static archive with no flywheel. The Stack Overflow trap. If the deal hands over everything that made you necessary and nothing keeps refreshing, you have sold your future for one cheque. Protect Factor 7 before you sign.
- Selling content you don’t cleanly own. User-generated content, guest posts, embedded third-party media and contributor agreements that don’t permit sub-licensing all blow up in due diligence. Confirm your rights before you pitch, not after.
- Optimising for volume over uniqueness. Ten thousand re-stated, undifferentiated pages are worth less to a model than one dataset only you hold. Thin, derivative content has no licensing value and, increasingly, no citation value either.
- Treating licensing as a substitute for visibility. Unless you’re at Reddit scale, licensing revenue alone won’t carry you. The durable position is both: get paid where you can, and stay cited and linked everywhere else. Falling out of AI answers is recoverable, but only if you diagnose why the citations stopped and rebuild the signals — don’t assume a licence deal makes visibility someone else’s problem.
Where this breaks in practice
Enforcement is the soft spot. RSL, robots.txt and llms.txt are declarations, not locks; non-compliant scrapers ignore them, and as of early 2026 no major lab had formally committed to honour RSL. Real teeth come from CDN-level blocking, cryptographic bot authentication, and — for serious breaches — litigation. Build the legal posture (Section 6) and the technical posture (Section 7) together, or neither protects you.
10. Your 30/60/90-day plan
Days 1–30 — Audit and decide
- Run the Training-Source Value Scorecard (Section 2) on your top three content types. Record the weakest factor for each.
- Confirm your rights position: can you actually licence this, or is it UGC / third-party content you can’t grant?
- Audit crawler access in your CDN and robots.txt. Are you accidentally blocking AI-search crawlers (losing visibility) or leaking training crawlers (losing leverage)?
Days 31–60 — Structure and signal
- Fix the weakest scorecard factor. Usually: restructure your best dataset into clean, chunked, schema-marked, dated form.
- Package one proprietary dataset as a citable public number. Publish it; pitch it; let it earn links while it sits as a licensable asset.
- Set your formal stance: update website terms on TDM, decide attribution vs subscription vs pay-per-use, and review the RSL spec.
Days 61–90 — Licence and protect
- If Scorecard ≥ 11: approach labs directly or via the RSL Collective. If 7–10: join a collective to negotiate as a bloc.
- Deploy enforcement at the CDN layer (pay-per-crawl / selective blocking) and reference your licence terms via RSL.
- Protect Factor 7. Build the flywheel — the recurring reason your corpus keeps refreshing — so you never become Stack Overflow.
The throughline of every winning ai training source strategy in 2026 is the same: own something the model can’t rebuild, keep it fresh, structure it so it’s cheap to use, hold the rights cleanly — and never licence away the flywheel that makes you necessary. Get that right and you stop being free training data and start being a supplier with a price, a citation, and a link profile that compounds. For the ground-floor fundamentals underneath all of it, our guide to what backlinks are and how authority moves remains the place to start.
