The Value of Voice: A 2026 Bet for AI Search
- Andrew Riker
- Aug 13
- 8 min read
I consulted my astrologist last week. She said audio is the next big play.
I've been making this bet in conversations for a few months now and it's time to write it down, including the parts where the data doesn't fully cooperate (you'll see below).
The short tldr; version:
most brands are pouring budget into the one format that got commoditized to zero
most treat the formats that are still expensive to fake as a brand awareness play measured in downloads.
Finding your context layer through audio and video will help you craft that moat you want
That's backwards, and the citation data I’ve been seeing is starting to show it.
The slop math is more interesting than the slop panic
You already know the headline number. Graphite sampled 65,000 English-language articles and found AI-generated share went from roughly 10% in late 2022 to 52% by October 2025, then plateaued around a 50/50 split. Every LinkedIn post about the death of the internet cites it.
The second Graphite finding is the one nobody quotes, and it's better. When they looked at what Google actually surfaces, 86% of the articles in search results were human-written. Only 14% were AI. This was data from October, so grain of salt now.
So the slop is real, it just isn't winning. What it did was reset the floor. Machine content made publishing cost roughly nothing, and your content now has to clear a bar that used to be cleared by simply existing. Fifty pages a month used to be a moat. Now it's the price of entry, and it's an entry fee anyone can pay with a credit card and a prompt!
This is also why I keep seeing the same shape in data I see. Visibility spikes for six or eight weeks, then flattens, then slides. The best folks like Lily Ray, Cyrus Shepard and others are finding this all the time.
The closed loop does not break through
I say this with love, because I've watched genuinely smart teams do it.
They ask an AI tool what topics to cover. They ask an AI tool to draft the coverage. They ask an AI tool to optimize the draft for AI retrieval. Then they wonder why the output has the same density of original information as everything else in the category.
No new information entered the system at any point. It's a closed loop. You put the model's understanding of the world in, and you got the model's understanding of the world back out, formatted nicely.
Google has been telling us the answer to this for years and we've been treating it as a compliance checkbox. Information gain. First-hand experience. E-E-A-T. Proprietary data. All of it is one instruction wearing four outfits: publish something that wasn't already in the index.
And in July, a survey of 45 GEO studies published between November 2023 and July 2026 landed on arXiv with a conclusion that should have been a bigger deal than it was: no evaluated technique consistently improved discoverability, traffic, or business outcomes across platforms. I covered it in the newsletter. Some rewrites actually REDUCED retrieval.
Structure isn't the differentiator anymore, cause, well, everyone can have structure. Schema is a solved problem you can buy with your dev team cycles. What's left is whether you know something that your competitors don’t.
Your brain is the last proprietary database
Every team I work with has a first-party data problem, and almost none of them have a first-party data shortage. They just are not using it, most likely.
They have sales calls where the same objection surfaces for the eleventh time. Support tickets nobody has read as a corpus. A VP who can explain exactly why the standard approach in their category fails at scale, and has never written it down because writing it down would take four hours and she doesn't have four hours.
Try to get that out of her as a blog post and you'll get a calendar invite that slips three times. Put her on a mic for forty minutes with someone asking good questions and you'll get more original thinking than her last six months of content calendar produced.
Talking is the cheapest extraction method we have for the stuff only you know. It's not a new observation, journalists have run on it forever. What's new is that the extraction now compounds into a machine-readable asset instead of dying in a Google Doc.
The transcript is the asset
This is the part people miss when they hear "invest in video."
AI engines aren't watching your webinar. With the partial exception of Google's own ecosystem, they're reading whatever text sits on the page: transcript, show notes, metadata, schema. Audio and video are the capture method. Text is what gets retrieved.
The Podcast Citation Index study out of Everything-PR made this brutally clear, too. They tracked 5,214 citations across 187 shows from December 2025 through May 2026, and the finding buried in their methodology matters more than the rankings did. Most shows that missed the top ten missed on transcript access, not content quality. If the engines can't crawl the words, the show doesn't exist to them.
Two decades of SEO people have been saying transcripts matter. They were right for the wrong reason, and now they're right for a much better one: a one-hour conversation transcribed and published on a page you own becomes a permanent retrieval anchor.
Which also means the podcast you host on Spotify and Apple only, with no episode page and no transcript, is invisible. You built the asset and then rented it to platforms that don't share it to help build that a sweet, sweet search equity.
What the citation data says, and where it argues with itself
Now the fun part, because the numbers here are genuinely messy.
The bullish case: Ahrefs' Brand Radar looked at 1.9 million citations across a million AI Overviews in July 2026 and found youtube.com as the single most-cited domain at 21.1% mention share. Other studies put it between 20.9% and 29.5% depending on the window and sample. BrightEdge has it at 16.6% in AI Mode. Adweek's January investigation across 6.1 million citations found YouTube had overtaken Reddit as the most-cited social platform, appearing in around 16% of LLM answers, generating 18x the citations of Instagram and 50x TikTok.
The bearish case: from the same body of research: ChatGPT cites YouTube at roughly 0.2%. Otterly's analysis of 100 million-plus citations found only 5.54% of all citations came from social and video platforms at all, with YouTube taking about a third of that slice. And Otterly reports Gemini's chat surface barely touches YouTube, even as Google's search-integrated surfaces lean on it heavily.
Both things are true because they're measuring different denominators on different products. "YouTube is the most-cited domain" and "video is 5% of citations" are not in conflict. One is share within AI Overviews, the other is share of everything.
So be precise about what you're buying. This is a Google-surface bet first, a Perplexity bet second, and close to a non-bet on ChatGPT today. If your buyers live in ChatGPT (it’s hard to even know that unless you’re doing self-reported attribution cause referral is so low, video is a slower play.
I'd still make it. Google is where the volume is, Google owns the video platform, and Google's models process the visual, the spoken audio, and the metadata as one input. Nobody else has that integration and nobody else can buy what they’ve curated over a couple decades.
You need an audience AND you need to be retrievable
The stat that changed how I talk about this: in Otterly's 2026 study, 40.8% of AI-cited videos had fewer than 1,000 views at the time they were cited.
Read that again if you run a brand channel with 400 subscribers and have been told it isn't worth the effort.
Citation is uncorrelated with popularity. The engine isn't looking for the video everyone watched, it's looking for the passage that answers the question cleanly, from a source it has reason to trust. Your 300-view customer-education video with a clean transcript and a real answer in it can beat a creator with a million subscribers who never published theirs.
So, it’s about the retrieval, but also distribution is the moat that some companies have over others, and having an engaged audience is obviously very important.
The attention argument, which is the one I actually believe most
Set the citation data aside for a second.
Audio is the only content format people consume while doing something else. Driving, running, weeding the garden, walking the dog (My Scout & Tilly are good girls).
Nothing else in your mix gets to borrow time from an activity that isn't content consumption. Every article you publish is competing for the scarcest resource on earth, time.
And the supply side is asymmetric in a way that I don't think is priced in yet. An affiliate operator can push 100 pages a week without breaking a sweat. Nobody has figured out how to fake being in a room with someone for an hour, taking hard questions, unscripted, with a recognizable voice and a name attached.
If provenance and authorship signals get any weight in future models, and I'd put decent odds on it given where the publisher fights are heading, the archive of you talking is worth more than the archive of you publishing.
Where this breaks
In 4 areas:
You make video slop. Same closed loop, new format. AI-scripted talking head reading an AI-generated outline into a webcam is not first-hand experience, it's the same commodity content with worse production values and a longer production cycle. If the person on camera isn't saying something they'd defend in an argument, don't record it.
You publish to platforms you don't own. Spotify, Apple, and YouTube are distribution. The episode page on your domain, with the full transcript, speaker attribution, headings that mirror how buyers actually phrase questions, and PodcastEpisode plus AudioObject schema, that's the asset. Platform-only is a bet on the platform's continued cooperation with each AI surface, and that cooperation is being renegotiated constantly. Ask Reddit and Google how their deal is going.
You measure it with the wrong scoreboard. Downloads and watch time are attention metrics. Fine for the social team, useless for this. The scoreboard is whether the transcript gets retrieved and cited, and whether you show up when your buyer asks the question.
Nobody owns it. This is the one that actually kinda kills it. It lands between brand, content, social, and demand gen, which means it lands nowhere. 61% of companies still haven't assigned AI search ownership to anyone. Adding an unowned podcast to that is just adding another orphan.
If I were standing this up on Monday
It would not be just a content project…more like a new system.
Pick one recurring internal conversation that already happens and already contains original thinking. The monthly customer call review. The post-mortem. Whatever. The thing your sharpest SME explains to every new hire. Record it. You're not adding a meeting, you want to record the whole damn thing.
Then build the pipeline once and stop thinking about it: record, transcribe, publish to an owned page with the transcript in HTML rather than a PDF, mark it up, cut the clips for distribution afterward. Clips are the byproduct that you share on the socials. The page is the point you need for search.
Two mediocre conversations a month beat one beautifully produced quarterly thing, because you're building a corpus.
And give it a name on someone's goals. Their personal goals. Give them a bonus if they do it!
The diagnostic question I'd ask your team this week: what did we publish last quarter that literally no other company in our category could have published?
If it takes more than ten seconds to answer, you already have your content strategy for the rest of the year, and it starts with turning a microphone on, and understanding the VALUE of VOICE.
Comments