For the 31 participants on INMA’s Media Tech & AI Week study tour, the last stop before the conference opened was AWS’s San Francisco office. The hosts shared that this wave of technology is unlike any publishers have weathered since the 1990s.

“This time feels different because AI is really at the core and heart of your business,” said Eyal Knoll Wel Roth, EMEA industry specialist for monetisation in media at AWS. 

AI creates content that competes with publishers’ own, it is increasingly the audience, and it decides how journalism gets discovered. AWS’s response comes in three parts: protect content, distribute it, and monetise it.

A different page for every visitor

It starts with knowing who is at the door. 

“The traffic coming from bots is now larger than the traffic that you see from humans,” said Megha Kande, senior manager of product management for Amazon CloudFront. 

Search crawlers still send readers back, roughly one referral for every 11 to 15 crawls. Training scrapers, which account for the largest volume, send none.

CloudFront now identifies about 650 bots classified according to IETF standards. It shows publishers which bots read which articles, and lets them rate-limit or block specific bots. Publishers can also answer an agent with a payment request instead of the article, currently through Coinbase, with Stripe coming soon.

The next step is routing. 

“Can you actually serve different content to an agent?” Kande asked. 

Unverified scrapers could get stripped-down pages or yesterday’s articles. Commercial bots could get a payment request. Personal agents acting for readers could get pages enabled for WebMCP so they can take actions on the reader’s behalf.

One publisher described why this matters now. Bots trigger ad requests and analytics events, which makes inventory look spammy and adds cost. The publisher would rather let a bot read than block a single genuine reader but would rather block than fire a bad ad request. 

Kande said bot classification will never be perfect. Confidence-score labels can already route uncertain traffic to pages without analytics JavaScript or ad hooks.

Licensing on three levels

Dev SenGupta, product leader for AI data and content at AWS, described three levels of content licensing taking shape:

• Bilateral deals between large AI labs and individual publishers.

• Pay-per-use arrangements, still emerging.

• A messy middle: collective deals such as a Danish publishers’ group representing nearly the whole market, and enterprise IT teams licensing specific content for internal tools.

Any deal, he said, must answer four questions: Who is asking? What can they do with the content? What does the publisher get back? And how does anyone know it happened?

A Dutch publisher described a collective agreement with Google and Microsoft covering every publisher in the Netherlands. It took about three years and brought in €3.2 million. That is modest, he said, but it establishes a legal baseline for further action and negotiation.

The weak point is holding together. The Dutch publisher said members of one industry coalition were already striking their own deals with OpenAI. Marcos Stupenengo, digital-media executive and strategist at AWS, predicted those bilateral deals won’t last more than two or three years.

Scarcity only works if publishers block together, Stupenengo said. As long as trusted outlets give their journalism away, AI companies have little reason to pay. If the major news organisations held back at the same time, those companies would have no alternative.

Pricing information by its value

Participants pushed past access to the harder question of price. Not all content is equal, one said. Publishers need to classify what they hold by value, archives included, and block rates need to exceed 30% before buyers take a marketplace seriously. 

Another participant compared it to programmatic advertising. Publishers could segment content by what buyers are looking for and price it accordingly, rather than treating their output as one lump.

So far, the money is thin. Asked whether publishers are earning from charging bots, Kande said: “Not yet.” 

The buyer side is weak because so much content is freely available to AI models. Stupenengo said AWS partners are seeing revenue mainly from corporate buyers. An individual user’s agent might pay a cent for content, while a licensed enterprise use could pay US$0.80.

Who can afford to wait

How urgently publishers act depends on where their revenue comes from. 

A study tour attendee from a local U.S. publisher, subscription-first with a 40-person newsroom, said the company has no interest in being an early adopter. It will keep bots out and watch while local TV and ad-dependent chains, which need the revenue, go first.

A Canadian media company, with two-thirds of revenue from readers and churn at a record low, is putting registration first to build direct relationships as search declines. INMA’s tour leader said the more diversified a publisher’s revenue, especially from readers, the longer its runway.

Another study tour participant was the most pessimistic. Publishers don’t know how they are discovered or presented in AI answers, they said, and have no position in the value chain to change either. The attendee expects significant collapse across the industry.

Routine work first

Ranjith K., principal solutions architect at AWS, estimated about 70% of newsroom effort goes to lower-value work such as research, translation, and approvals. That is where agents help most, provided they have governed access to archives and analytics.

One executive pushed back. AWS shouldn’t try to replace general-purpose AI tools, they said. It should help publishers build direct relationships with audiences, using assets it already owns such as Twitch.

Stupenengo offered two examples of the routine gains already in use:

• An audio-to-text system alerts journalists when a minister says something newsworthy on morning radio.

• A content system prompts reporters to request comment from companies or police named in a story, then sends the email.

AWS is about to enter a planning cycle for its 12-month roadmap. Kande asked the room what it should build.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .