28 Wishes sells ice cream in Los Angeles. Below 70°F, sales fall about 20%. The weather is out of their hands. Rent isn't.

So the owners started putting about $20 a day into Kalshi weather markets, taking the cold side. The days that keep customers away now pay something back.

This is hedging. Big companies have done it for decades, buying protection against bad weather, fuel spikes and rising rates. It used to take a broker, a trading desk, and an order size no corner shop could meet.

Kalshi opens it up. Contracts on weather, fuel prices, inflation, tariffs and regulation, starting at a few dollars. Take a position on the outcome that would hurt you. If it hits, the payout softens it. If it doesn't, the contract expires and the good month was the point.

See how other owners are doing it at Kalshi.com/hedging

There's a scene in Back to the Future Part II where Marty walks into Cafe 80s in 2015 and shows off on Wild Gunman, the Nintendo shoot-out game, only to be put in his place by two kids (one of them a very young Elijah Wood) who watch him play and say: "You mean you have to use your hands? That's like a baby's toy!"

I used that line in a presentation at WPP back in 2017, when I felt we were starting to see customers change how they got to answers, quicker and by voice. I was wrong for 2017, but it feels very right in 2026, because for the last few months I have barely typed. This section was dictated, as are most of my prompts, emails and Slack messages, and it's Wispr Flow doing the work.

About 10 years ago voice was meant to be the next big thing in search. The stat everyone quoted was that 50% of all searches would be voice by 2020, and for years it was credited to Comscore, wrongly as it turns out. It actually came from Andrew Ng, then chief scientist at Baidu, who told Fast Company in 2014 that "in five years time at least 50% of all searches are going to be either through images or speech". Somewhere along the way the images part fell off, Comscore's name got attached and it ended up in every agency deck going.

It didn't happen, at least not in the way anyone meant. Smart speakers turned out to be very good at timers, weather and music and not much else. You had to talk to them like a keyword ("weather London tomorrow"), and if they misheard you (“I’m sorry, I don’t understand”), you started again. The frustrating part is that the case for voice was already there, because a Stanford study in 2016 found speech was three times faster than typing on a phone. The speed was never the problem, it was what you could do with it.

Two things have changed. The first is that the model on the other end now understands you when you ramble, so you can think out loud, change direction halfway through a sentence and correct yourself, and tools like Wispr tidy it into something readable. The second is that there is now somewhere worth sending all those words. A voice assistant in 2018 gave you one answer, whereas an AI assistant takes a paragraph of context and does something useful with it.

It isn't just me. A Morning Consult survey for Menlo Ventures (who, full disclosure, led Wispr's latest round) found 46% of US AI users now speak to AI at least some of the time, and 19% say voice is their main way in. Google has said that conversations in Gemini Live run five times longer than text ones on average, which is the context point in a single number.

The platforms are building for this. OpenAI rebuilt ChatGPT's voice mode as GPT-Live in July, which can listen and speak at the same time and search the web mid-conversation, and this week it added email, calendar and Slack. The money is following too, with Wispr raising $280m at a $2bn valuation in August.

My own numbers make the point. I've dictated 246,380 words through Wispr, which it tells me is roughly two books, at 112 words per minute, which puts me in the top 1% of users. By my rough maths that's around 55 hours I'd otherwise have spent typing, or about seven working days. The bit I didn't expect was where the words go: 63% of my dictation (4,116 prompts) goes straight into AI tools, a quarter goes into work messages and only 4% into email.

It has also grown well beyond dictation. You can talk into your laptop or your phone with the same ease, it now takes meeting notes, which Wispr rolled out recently, and there is a scratchpad for jotting things down, so a lot of what I used to spread across three or four tools now sits in one place. At $15 a month it is ridiculously cheap for how much I use it.

The one thing I'd say to anyone trying it is not to judge it on day one. At first it feels slow and a bit awkward, because you are getting used to thinking out loud rather than typing, but you pick it up far quicker than you expect, and once you're in the flow of it, going back to the keyboard starts to feel like the baby's toy.

Here's the thing for anyone tracking visibility. When I type a prompt, I keep it short, as if I'm searching. When I talk, I give far more context: who I am, what I've already tried and what I'm worried about. If that holds for people generally, spoken prompts get longer and more specific, while most of the prompt sets brands are tracking still read like typed keywords. Last week I said to write your prompts the way people actually talk to an assistant, and voice makes that literal.

Back to the Future Part II got plenty wrong about 2015. It missed the internet entirely, filled the McFly house with fax machines and we're still waiting on the hoverboards. But the kids in Cafe 80s were onto something, as was Marty Jr shouting channel numbers at the TV: the future they imagined was one where you talk to your technology rather than use your hands, and that's the bit that is finally coming true.

Comscore's AI Intelligence Report for Q2 is out, and it is one of the better pictures of how people in the US are actually using AI, because it comes from panel data rather than a survey. AI assistant tools reached 35% of desktop users and 29% of mobile users in Q2, up from 25% and 12% in January 2025.

The bit that stood out for me is where the growth is coming from. In absolute terms the biggest groups of users are still aged 25 to 54, and the over-55s started from a much smaller base, which flatters their growth rates. But no other age group grew as fast. Visitors aged 65 and over more than doubled year on year, up 112% on desktop and 108% on mobile, and 55 to 64-year-olds were up 98% on mobile. If your audience skews older, this is definitely an area to explore now rather than later. It also ties back to the voice point above, because Menlo's survey found over-65s are just as likely to talk to AI as Gen Z.

Two other numbers are worth knowing. ChatGPT's share of US AI prompt volume fell from 70% to 50% between January and June, while Gemini went from 17% to 30% and Claude from 2% to 11%, so auditing one platform now means auditing half the market at best. Conversations are also getting longer, averaging 5.4 prompts on ChatGPT, 6.2 on Gemini, 6.9 on Copilot and 9.1 on Claude, which is a reminder that a single prompt is rarely the whole journey.

One caveat, though: this is a look back. Comscore's data runs to the end of June, before OpenAI released GPT-6 Astra in early September and GPT-6 Sol and Luna last week, so I wouldn't read it as ChatGPT being in decline. It is still comfortably the biggest platform, and the more useful takeaway is that the market is spreading out, not that one player is fading.

The most useful chart in the report, for anyone working on visibility, is in travel. Comscore split the sites ChatGPT read to build its hotel answers from the ones it linked as an inline citation in the answer itself, and the gap is big. Tripadvisor appeared in the sources behind 61% of hotel-related answers but as an inline citation in only 21%, while Booking.com dropped from 33% to 9%.

Even a citation doesn't guarantee a visit, because when Pew tracked real Google searches, people clicked a link inside the AI Overview on just 1% of visits. And a link still isn't a recommendation, as Comscore doesn't measure whether the brand was actually named or suggested.

This matters most for publishers and affiliates. We talk to a lot of them, and being cited by AI is fast becoming the currency they use when talking to potential customers, but the affiliate CPA model is dying when chatbots recommend brands using their content and nothing gets attributed back.

That said, being the source rather than the answer isn't necessarily a bad thing for Tripadvisor, because if its reviews are shaping which hotels ChatGPT suggests, it is still influencing the decision. Being read, being linked and being recommended are three different outcomes, and most reporting only measures the first.

A few more takeaways from the report:

Put together, AI use is spreading well beyond the early adopters, older audiences are catching up fastest, it is split across more platforms and the thing worth measuring is whether you show up in the answer, not just whether you were read.

On Wednesday Obsero co-hosted a dinner for finance marketers with our friends at Adclear, with marketers from brands including PensionBee, CMC Markets, IG, Zego and CME Group around the table. It ran for over five hours, which tells you how much there was to talk about.

We covered everything from growth over efficiency and the CMO's seat at the executive table to how people are working with creators, and plenty of GEO. Self-reported attribution came up again too, so it's clearly not just me who thinks "where did you hear about us?" is the most underrated question in marketing.

What struck me most was how many of these brands are already leaning into AI search properly, rather than waiting to see how it shakes out. We'll be running more of these over the coming months, so if you're a marketer and want a seat at the next one, reply to this email and we can chat.

I'm also out and about. I'm in Dublin this week for the AWS panel on Tuesday and at brightonSEO the week after, so if you're around either and want to talk AI search over a coffee, let me know.

That is it for this week, until next time.

Have a good one.

Andy