Nashville Briefing: What a Good AI Deal Actually Buys

Next week the sell side gathers in Nashville. Before the handshakes and hallway promises, we should get clear on what has actually changed with AI content licensing.

In the last 60 days, AI licensing jumped from polite exploratory calls to signed checks. Some of the deals are sharp. Some are not. Either way, the pace tells you this is not a distant conversation anymore.

This is not the endgame. It is the opening position.

If you run a newsroom or a portfolio of sites, the question is not “should we sign,” it is “what are we selling, and what are we keeping.” Training rights fill an AI model’s memory. That is a one-time transfer of value for you and evergreen value creation for them. Real-time use is where your work shows up in front of a user, and it must be tracked, attributed, and paid each time it happens.

This brief is meant to be useful. What follows is a plain read on recent deals and what a good deal looks like: the rights they actually buy, the credit your brand keeps, and the controls that return value to you. Walk into Nashville with that frame, compare notes with peers, and you’ll be ready to evaluate AI licensing deals on solid footing.

The New Map: Purpose, Attribution, Transformation

If you want to make sense of these agreements, and see the risk buried in the fine print, you need a better map than “training rights” and “distribution rights.” Those were built for the search engine era.

We use a different framework - a content usage taxonomy - built for the AI economy. It cuts through the marketing gloss and tells you exactly what’s being bought, sold, and risked. Three dimensions matter most: Purpose, Attribution, and Transformation.

Purpose: Learning vs. Real-Time Use

Why is the buyer touching your content at all?

• Learning is about building something that persists: an index, a model, a knowledge graph, insights and analytics. It’s “train once, use forever” territory. Once the model has your content in its weights, you can’t pull it back.
• Real-Time Use is about applying your content in the moment: answering a user’s question, powering a surface, generating an experience on demand. Every use is a new transaction, which means every use is a new opportunity to track, attribute, and bill.

These are not interchangeable rights. Training is a one-off premium. Real-time use is a metered utility. Smart deals price them separately.

Attribution: From Source of Truth to Footnote

Once your content is in the output, what does the user actually see?

• Citation is clear, clickable, and on the surface where the user’s eyes land.
• Mention is buried, a hover state, a footnote, a secondary tab.
• No attribution: the user sees the AI as the source.

AI platforms have every incentive to downplay source credit. Unless your deal locks in on-surface citation, you’re betting on good faith in an environment where the economics reward the opposite.

Transformation: Verbatim, Summarized, Synthesized

How does your content survive the trip through the system?

• Verbatim is quoted exactly as you wrote it.
• Summarized is paraphrased or shortened, often losing nuance.
• Synthesized is blended with other sources into a “new” answer, the most common, and the most brand-erasing.

The more transformation applied, the harder it is to prove your content’s value, which is exactly why AI companies like synthesis. Your deal needs to recognize that and price accordingly.

How the Current Deals Map to the Taxonomy

The New York Times’ multi-year deal with Amazon looks impressive at first glance: training rights, real-time use in Alexa answers, and a reported $20–25 million a year. But here’s the danger: it’s a blended scope. Amazon gets to learn from the Times’ archive, a one-way transfer that lives forever in the model weights, and also to use that content in real time without coming back to the well. That’s permanent value for them, recurring cost for the Times. Attribution? Likely there, but Alexa’s UI doesn’t exactly make citations the hero; it’s a polite mention at best. Transformation is almost always summary or synthesis, which means Alexa owns the final voice. Without ironclad language on visible credit and detailed usage telemetry, this is the kind of deal that pays well up front and quietly underperforms over time. At minimum, require per-use telemetry that captures surface, timestamp, query class, and content URI, then tie monthly revenue recognition to those logged events.

Condé Nast and Hearst’s decision to feed Amazon’s Rufus is a different play: pure real-time use in a high-intent commerce environment. When someone is asking for “the best winter boots for icy streets” and your review powers the answer, that’s influence at the moment of purchase. Done right, that’s gold. The risk? In retail flows, attribution is often reduced to a product mention, and your content is rarely left intact. Rufus will summarize or blend your work with competitors’, stripping away your brand’s authority. If the rate card scales with actual conversions, this is sharp. If it’s a flat fee for a crawl and some summaries, Amazon walks away with the upside. Add explicit controls: rate limits, category or SKU scoping, and a hard kill switch if placement degrades or usage exceeds agreed terms.

Gannett’s arrangement with Perplexity should have been a local-news power move. Two hundred markets’ worth of coverage flowing into real-time Q&A could anchor Perplexity’s authority on community-level queries. But instead of focusing on per-answer monetization and placement guarantees, the deal headlines “free Perplexity Pro access for Gannett employees” as part of the value. That’s a perk that will get cheaper every year as AI assistants commoditize, while the cost of producing local journalism doesn’t. Attribution in Perplexity answers is better than nothing, but with heavy synthesis across sources, click-through is not guaranteed. This is the danger of trading durable rights for short-term perks. Define the unit price: per displayed answer that includes a Gannett source, or per unique session where the answer cites Gannett, with a higher rate for top placement or featured callouts.

Google’s active licensing talks might be the most important test case yet. Their Gemini-powered AI overviews combine Learning and real-time use in a single scope, just like the NYT–Amazon deal. But the interface leans heavily on synthesis, and current citation patterns are mention-level at best. Without hard terms on placement, attribution, and per-use payments, this risks becoming the template for “pay-to-be-summarized,” locking in a model where AI is the front door to your work, and the front door is unmarked. Anchor economics to a minimum guarantee plus per-use fees for each overview impression that includes your content, with escalators as usage grows and as new Gemini surfaces roll out.

If I Were Negotiating Today

If I were in those rooms, I wouldn’t start with “what will you pay me?” I’d start with how will we measure value, and how do I stop you from taking more than you’re buying? Because in this market, the price tag is the last thing you negotiate, not the first.

• Separate the scope. Training rights and real-time use are different products with different economics. Price them separately.
• Lock in attribution with teeth. On-surface, in-context, unavoidable. Not a hover state.
• Price for transformation. The more they strip your voice, the more they pay.
• Demand measurement and reporting. Per-use telemetry, placement logs, audit rights. No data = no deal.
• Enforce with controls. Rate limits, kill switch, per-class toggles, machine-readable policy compliance.

Pushback is inevitable. You’ll hear, “If we can’t get terms we like, we’ll just scrape.” Maybe they will, but the market is shifting. The biggest buyers are paying because they need to. The sellers who set enforceable terms now will define the benchmarks. The ones who don’t will be negotiating against those benchmarks later, from a position of weakness.