Will the real Aggregator's AI Risk please stand up?

A few weeks ago, tech analyst Ben Thompson wrote an article entitled Aggregator’s AI Risk, where he talks about the risks Facebook, Google etc. face from offering AI tools.

I was interested because my AI vs. ads series is, in large part, about risks of AIs to Aggregators. (Aggregator is Thompson’s term for large tech platforms, based on his Aggregation Theory.)

Here’s my summary of his argument:

• Currently, Aggregators work by ‘providing a discovery mechanism for the abundance of supply’

  • In other words, they take in a ton of content and show it to users in a convenient way, like a feed or search results

• However, Aggregators are now moving towards giving AI tools to users

• These AI tools function like an ‘anti-printing press’, giving users ‘simply one answer’

• This exposes these platforms to a new risk:

  • If the user doesn’t like the one answer provided by the Aggregator, they will go elsewhere

  • Related: if politicans don’t like that one answer, they’ll impose costs via laws. Google exposed itself to this risk when Gemini showed ethnically diverse Nazis

• This risk can be mitigated by providing personalized AIs

  • For example, prompts to AIs automatically add context about your preferences to them

This is a fascinating argument, but I don’t think it’s the biggest risk to Aggregators from AI — nowhere near, in fact. For me, Thompson’s argument begs an interesting question:

Why are Aggregators moving towards giving AI tools to users?

Because ChatGPT, of course. But what has ChatGPT done? Many things, but the scariest thing as far as Aggregators are concerned is:

Created a new Last Mile of content delivery.

Aggregators used to bring content straight to the user. In Thompson’s now-famous diagram, it looks like this:

Credit: Ben Thompson (Stratechery)

On the left, you have an Aggregator bringing together a ton of content into one place, and on the right, the user consumes the content in that place.

But now AIs are able to intercept the content and deliver it to the user in a new Last Mile. That’s the anti-printing press part, of collapsing things to one answer. So things start looking like this:

Credit: Ben Thompson (Stratechery), with my edits in purple

There’s a whole other stage now, between where Aggregators dropped off content and where users consume it! This new Last Mile of content delivery looks like Aggregators still doing the same aggregating work, but AI saying ‘cool, thanks — I’ll take it from here’, and adding a layer of curation (and potentially mutation) to the feed or search results the Aggregator came up with.

That’s why the Last Mile is a funnel. AI boils the large amount of content delivered by Aggregators down to the ‘simply one answer’ Thompson wrote about. Hence the problem for Aggregators: one answer is one answer — not, lots of answers and a bunch of ads.

See that spot on the diagram where AI rudely blocks the flow of content from the Aggregator to the user? Let’s call that The Aggregator’s Lair. It’s wonderful being an Aggregator when people consume content in your Lair, which is usually a feed or search results packed full of ads. It’s much less wonderful when AI shows up in the Lair as the user’s ambassador, creating a huge gap between you and the user.

True Aggregator power comes from controlling demand to the extent that you can impose ads on those using your service, forcing them into your Lair. Without ads, there is no money. Without money, there is no power.

Given all this, we can now understand a major reason why Aggregators are moving towards giving us AI tools: they want to own the Last Mile, to create a new Lair at the end of it.

Last Mile Risk

But it’s not clear that the Last Mile will be a great place to introduce ads. What will people’s expectations be of Last Mile output — will they be OK with it becoming an ad-filled Lair? What will regulators think? Will Last Mile output be an information-dense environment where it’s easy to squeeze in plenty of ads, or will this output — typically in a ‘simply one answer’ format — have less surface area for ads? In the best case scenario where Aggregators always own the Last Mile, there’s at least a lot of risk for them.

Even worse for Aggregators though: they won’t always own the Last Mile. We can, in fact, expect there to be an arms race for it. Potential competitors like OpenAI, Anthropic, and the universe of startups building with their APIs — many of whom, crucially, do not rely on ad revenue but instead charge a monthly subscription to consumers willing to pay for information — are lining up to either:

• Compete: Effectively become Aggregators themselves, by scraping a lot of content (OpenAI-style). Then, add the Last Mile over the content they scraped — a much longer and arduous Last Mile, since they don’t get to pick up from the point of the relatively filtered Aggregator’s Lair. (This is how ChatGPT created the first new Last Mile, kicking off the whole arms race). This effectively increases the number of Aggregators in the world, which is bad for all existing Aggregators’ competitive positions.

• Interoperate: Provide an alternative Last Mile on top of an existing Aggregator’s Lair (which in the real world, is an app or a website). Whether competitors are able to do this will be a function of how much interoperability they can impose on Aggregators. This, in turn will be a function of politics and regulation, of legal pushback by Aggregators, and of permissions we are willing to grant AI in our AI OSes. If interoperating becomes easy, it directly endangers the ability of Aggregators to keep users in their Lair.

• Pressure: As a result of Competing and Interoperating, pressure the Aggregators to make their own Last Mile output less ad-rich. As we’ve seen, many Last Mile competitors won’t rely on the ad revenue model. They’ll charge a monthly subscription to consumers willing to pay for information. This means they’ll be able to afford to keep their Last Mile output ad-free, putting pressure on Aggregators to reduce the amount of ads in their Last Mile output.

One of the perverse dynamics here is that: the more Aggregators create the Last Mile, the more they start socializing it, inviting competition for it, and shifting the paradigm towards it! So they are damned if they do (create the Last Mile, since this speeds up the coming of the Last Mile arms race) and damned if they don’t (since others will get ahead in the arms race).

All of this is what I call Last Mile Risk.

One final point. I’m analysing this all through the lens of the AI vs. ads series, hence the focus on how AI may affect the large tech platforms: even so, it would be a huge oversight not to mention the Last Mile Risk to humanity — from a world where the Last Mile of content delivery is outsourced to black box AIs.

Plenty has been written on the dangers of AI intermediating information (and intention), but I’m going to consider whether the Last Mile framework means I can write something useful on the topic.

---

A few side notes:

1. This highlights the huge risk to (certain) Aggregators of people being socialized by ChatGPT, Copilot and others into a model of paying for information directly with money, instead of paying for information with attention.

2. Most of the biggest Aggregators have weakened their positions by replacing the social graph with the interest graph (in other words, giving you content based not on what your friends post/like, but based on your interests). This could be an article in itself, but its bearing on Last Mile risk is simply that: if you can consume TikTok without going into TikTok, you’re now more likely to do that because you don’t really use it (or Instagram etc.) as a place to chat with friends. There is no innately social reason to go into the app. Now, the Last Mile is coming along to make it so you don’t ever have to. And integrated AIs will eventually be able to create personalized content with no other Aggregators involved at all.

3. I agree with Thompson that the risk he points out to Aggregators can be mitigated by simply allowing personalized AIs. That’s a trend that I’d expect to see across all AIs, especially as we move towards AI OSes.

4. Thompson writes, in Economic Power in the Age of Abundance:

The Internet, though, is a world of abundance, and there is a new power that matters: the ability to make sense of that abundance, to index it, to find needles in the proverbial haystack.

Yes. But I’d modify the original analogy. Aggregators find needles in the proverbial haystack and move those needles into an ads haystack. But now AI has come along to simply give you the needle. That power is going to matter a whole lot.