11 May 2025
News
Is AI finally hitting search traffic?
Apple’s Eddie Cue, testifying in the remedies phase of the Google search antitrust case, said in passing that Google search queries from Safari dropped for the first time last month, which he thought was probably due to people using AI instead. He also said Apple is looking at revamping search on iPhones around LLMs (i.e. OpenAI, Perplexity, Anthropic and others as well as Google). Google stock fell 10%, and Google rushed out a statement saying that its traffic from Apple devices hasn’t fallen. (Can more people call Eddie as a witness, please? He always says interesting stuff.)
You can read this different ways. Apple gets $20bn a year from Google and so Cue has a motive to defend the status quo - ‘the market is already competitive - no action needed! Let us keep the $20bn!’ It’s not clear why Google and Apple disagree, given they should both have the same data. It’s not clear how much LLM use is substitutional versus additive to traditional search, nor how much such use covers the kind of search that drives ad revenue (is ‘find me a mesothelioma lawyer in Chicago’ better on Google or ChatGPT?) It’s not even clear whether LLM search will be captured by OpenAI or Google (note that for all the buzz it gets in tech, Perplexity struggles to break the top 100 in app store ranks where OpenAI has been at the top for a year).
However, maybe that lack of clarity is the point. No-one really knows how any of this is going to go, so, as we saw with DeepSeek at the beginning of the year, people jump at anything that looks like signal. CUE, GOOGLE
A busy week at OpenAI
Bloomberg reports that OpenAI has reached agreement to buy Windsurf, an AI coding assistant, for $3bn. It hired Fidji Simo, the CEO of Instacart (and former Meta exec) to be COO. Sam Altman testified in the US Congress, flipping his previous lobbying line and asking for less regulation instead of more. And Altman backed down from his attempt to change OpenAI to something more like a ‘normal’ company structure, which doesn’t resolve the tensions in its relationship with Microsoft and other investors.
Sam Altman seems to be a polarising future amongst those who’ve worked with him (many of whom have quit), and he also has a lot to do: from the outside it sometimes looks like his three jobs are lobbying, fund-raising and stakeholder politics. But somehow, he has to keep Open AI setting the agenda as the whole tech industry jumps onto AI, and as all the model companies try to work out what an actual product and strategy beyond ‘build better models!’ should be. Moving up the stack is part of that, and so is hiring a good COO. Meta already has a ‘head of product’ in Kevin Weil, though, and Sam is still CEO, so… more politics? WINDSURF, COO, CONGRESS, STRUCTURE
Figma moves up the stack
OpenAI is trying to move up the stack by buying Windsurf - ‘code an app that looks like this!’ Today, apps and services are designed in Figma, though, and so naturally Figma is also moving up the stack. This week it had a big AI day with a lot of new coding tools to help teams turn their mockups into software. LINK
Meta wins a spyware case
Meta was awarded $167m in damaged after suing NSO, the best-known of the (mostly Israeli) ‘mercenary spyware’ companies that find ways to hack WhatsApp, iOS, Android and other systems and spy on specific, targeted people, and sell those tools to (allegedly) anyone who can pay. Apple also sued NSO in 2021, but withdrew the suit last year saying that it would be require to disclose too much of its own security systems if the case went to trial. LINK
Self-driving trucks are here?
Aurora launched commercial autonomous truck service in Texas. I’m old enough to remember when this was supposed to be easier than doing passenger service in suburbs, and would be working by 2020. LINK
The week in tech law
I write about this stuff here more for the record than because it’s interesting or even important - these are arguments from a decade ago, where even the most dramatic proposed outcomes won’t change much. But there's a lot going on right now, and these are still billion-dollar questions, so...
Google and the DoJ spent the week arguing whether Chrome should be (or even could be) spun off. LINK
Somewhat more plausibly (but mattering even less to anyone except digital ad people), the DoJ wants Google’s ad network (~10% of revenue) to be spun off. LINK
Apple filed to stay a court’s order last week that it allow developers (in the USA) to link out freely to any payment method they want without paying Apple a commission. Apple argues that this amounts to an order for it to provide its services (i.e. access to the platform and APIs) for free, and that US courts aren’t allowed to do that. I have no view on the legal position, but it strikes me that Tim Cook always talks about ‘user sat’ and ‘delighting our customers’, yet forgot to treat developers as users to be satisfied and as customers to be delighted - instead Apple treats them as suppliers to be squeezed. LINK
Meanwhile, the Kindle app and a few others immediately started offering an in-app Buy button after the Apple ruling, after 15 years in which this was impossible (for ebooks and streaming music in particular, they never had the margin to give Apple 30%). There is no way that Apple’s rules here were good for anyone. LINK
Meta was in court defending its use of (sometimes pirated) ebooks as LLM training data. Not a good look, but I find it very hard to see how using a book by Sarah Silverman in the training data affects the market for that book - this is not Napster. LINK
And there was probably something else that was so boring I’ve already forgotten it.
Ideas
College students are using ChatGPT, a lot. So, what are the tests measuring? Should they measure something else? We let students use calculators and spell check now, after all: it’s easy to say this is different, but what does that mean? LINK
Equally, big companies are all starting to use AI for hiring, and that creates just as many questions. LINK
US ‘dollar stores’ are using delivery apps. I would have guessed that this would be a perfect case of adverse selection, but apparently not. LINK
Google is now running small on-device LLMs inside Chrome to look for scams. LINK
Johnson & Johnson moves from many AI pilots and experiments to a few focused projects. LINK
WPP CTO Stephan Pretorius on all the ways it’s operationalising generative AI in the ad business. LINK
Pair that with Martin Sorrell (WPP Founder) shifting his S4 Capital from adland’s traditional hourly billing to charging by output. This is a general issue for professional services - if AI means you can do the same work in less time, do you still bill for time? Was the time the value, or just a proxy for value? LINK
JP Morgan estimates that an India-assembled iPhone is only $20 or so more expensive at US retail than a Chinese one pre-tariffs. (Apple plans to supply the US from India to mitigate Trump’s tariffs.) LINK
Tiktok is banned in the USA, but Donald Trump keeps postponing enforcement of that (without any actual legal authority to do so, if that matters anymore), and so the team is out there pitching to advertisers. LINK
Elizabeth Holmes’ partner has a blood-testing start-up. Yes, really. Satire is dead. LINK
Outside interests
At auction, a copy of Paul Rand’s presentation book for the NeXT logo. LINK
Dubai’s supercar graveyards. LINK
When satellites de-orbit and burn up, the vapour isn’t good for the atmosphere. Now that we have fleets of tens of thousands of LEO satellites, that might become an issue. LINK
The London mansion whose owner is asking for permission for an underground helipad. LINK
Data
This IBM CEO survey on AI adoption is more interesting than one would expect - lots of discussion of FOMO and pilots that didn’t make it to production (which is why you do pilots, but still). LINK
Owl & Co estimates that the global podcast market is worth over $7bn, double the IAB’s number. LINK
Column
Distribution
Imagine you have a new service with millions or tens of millions of daily users, that you want to get to billions. The product sells itself to some users, but many others have looked a few times and not come back. You don’t have an obvious viral loop and don’t seem to have network effects. And at the moment, your product is a pretty thin ‘wrapper’ on an underlying technology that seems to be a commodity - half a dozen other very well-funded companies are trying to do something very similar. What do you do? How do you get breakout adoption for an undifferentiated product? What do OpenAI, Anthropic, Perplexity or indeed Google, Meta and Microsoft do?
What do you do? Well, at a first past, there’s a list of answers that’s so predictable that you could ask ChatGPT. You do the obvious ‘build a platform stuff (remember when OpenAI was trying to build an app store?’) You can hire some big-name heads of product with experience of big consumer apps in the previous platform shift (Kevin Weil at OpenAI, Mike Krieger at Anthropic). If any viral moments do seem to be happening (remember the OpenAI studio Ghibli thing?) you can try to jump on them, but it’s not clear what retention that gives you. And then, you can go for distribution.
What kind of distribution? Do you buy users? (Meta ads worked very well for TikTok, at vast expense). Run a TikTok campaign (Perplexity)?
Or, can you leverage your existing base? Anyone with a legacy user base of any scale has spent the last year or two on one hand adding LLM-powered features and enhancing existing features with LLMs: Google with ‘AI Overviews’, for example. But can you use those surfaces to drive people to the New Thing itself?
So, Microsoft added a ‘Copilot’ button to Bing, but Google so far has decided not to put ‘Gemini’ on the home page, nor add it to Chrome (anti-trust concerns?). Meta added ‘Meta AI’ to the search box in Instagram and WhatsApp, and then added a floating blue circle, but it’s not clear that anyone knew what that meant: then it decided it had to create its own stand-alone Meta app as well. Google and Microsoft also have stand-alone ‘Gemini’ and ‘Copilot’ apps, but all of these apps struggle to break the top 100 in the app store. ChatGPT has been number 1 in the rankings for a year but the others are nowhere. There’s an innovation dilemma here, especially for Google - should it tell people to use Gemini instead of ‘Google’ or as well as ‘Google’, and if so, what for?
Some of this also feels like arguments we had a decade or more ago about apps (and Kevin and Mike know all about these). It is easier to drive adoption by making the new thing a tab in your existing product, or by unbundling it? Meta got Reels working by adding it as a tab in Instagram, but is Meta AI ready to be a tab? Conversely, how do we get people to install the app? On Android, Google can (for now) get OEMs to preload Gemini, but do people know what it is?
All of this is very early, and meanwhile, you can still try to make the app social and viral, even if we don’t know how yet. Meta’s new ‘Meta AI’ app has a social feed that tries to get more virality going (and also just tries to help people work out use cases), though a cynic would suggest that Meta is hobbled by not having anyone to copy. Will someone else work out a new growth pattern or a new UX?
For now, though, OpenAI is at the top of the charts through sheer user pull rather than product push (though it’s done some of its own distribution growth hacking, such as the integration deal with Apple). But the distribution teams are cranking up. Indeed, Google’s own growth is an interesting comparison - in the early years it leaned on distribution deals with Yahoo, a genius IE toolbar growth hack, and all sorts of other creative ways to get people to try the product. Of course, it benefited from the fact that Google search was so much better, which isn’t really true of any of these products at the moment. But the other part of the Google story is that it wasn’t there at the beginning. It came after the first wave of portals and search engines and showed them the right way to do this. And we’re still at the beginning of LLMs.