Forty-seven seconds.
Four-minute listens.
Five levers.

Every newsroom has a name for the reader who lands on an article, scans, and disappears before the second paragraph: a fly-by. The pattern is older than the web. What’s new is how brutal the numbers have become.

Twenty years of screen-attention logging at UC Irvine, summarised in Gloria Mark’s Attention Span, puts the average reader’s window on any screen down to forty-seven seconds^[1] — down from two and a half minutes in 2004. The reporting is good, the lede is sharp, the photography is polished, and most of the audience never reaches the bottom of paragraph two.

Audio articles open a different axis entirely. The same investigation, read aloud rather than typeset, holds 75% of listeners through to the end across the BotTalk network of more than thirty European publishers running text-to-speech in production. Same words. Different attention budget.

This piece is about the five operational levers that turn that completion-rate number from a per-article win into a session-extending layer of the newsroom — and why the publishers running them are pulling away from the ones who aren’t. Written from inside BotTalk, the orchestration layer running audio articles across thirty European newsrooms today.

Why text loses and audio compounds

The fly-by isn’t a per-publisher anomaly — it’s the texture of the 2026 attention market. Discovery is platform-first: readers arrive from Google Discover, Apple News, or a push notification with a single intent gate, the headline. Scroll-velocity is muscle memory: a social-traffic reader has already scrolled past forty pieces of content in the last half-hour. The bar to stop scrolling on yours is higher than the bar to read yours.

Chartbeat made the cliff quantitative more than a decade ago, with an analysis of two billion page visits that found 55% of visitors spend fewer than fifteen seconds actively on a page^[6]. Pew Research’s cross-publisher benchmark, drawn from a 117-million-interaction dataset, refined the picture from the engaged-reader side: average engaged time on long-form is about two minutes; on short-form, closer to one^[2]. The article the editor designed for ten thousand words rarely gets read past two thousand.

Audio doesn’t beat text on this axis. It opens a new one. The reader who bounced at 47 seconds stays for four minutes when the article is read to them. The reader who would never have visited at all subscribes to the podcast feed and listens at the gym.

Underneath that shift sits a deeper change in how audiences consume any media at all. Up to 85% of mobile-feed video on Facebook and Instagram is now watched with the sound off^[5]. The screen time is already lost. What’s rising in its place is background audio — over the speaker while the kettle boils, on AirPods during the school run, in the car between meetings. Publishers shipping only text compete for the visible-screen minute. Publishers shipping audio articles compete for every other minute of the day too.

Five ways to fuel long sessions with audio articles

1. Front-load the play button — audio is the first surface, not a feature

Most publisher CMSs ship the audio player as a small affordance under the headline, easy to scroll past. The fly-by never reaches it.

The first move is to promote the player to the same visual weight as the lede. A wide, branded, one-tap audio bar — above the first paragraph, sticky on scroll, with the playback duration shown next to the play button so the visitor knows the commitment up front. Show “4:32” the way YouTube shows runtime. Listeners decide on duration, not on prose.

Two follow-ons. First, autoplay-on-tap rather than autoplay-on-arrival — respect platform rules and respect the reader, but make starting audio cost one tap rather than three. Second, persist the player when the reader scrolls so the body of the article stays visible while audio keeps playing. A miniature dock at the bottom of the viewport is the pattern that converts; a player that disappears off-screen the moment the reader scrolls is the pattern that doesn’t.

2. Clone the voices your readers already trust

Synthetic neutral voices are fine for utility content. They are not fine for long-session retention. Listeners stay with voices they recognise — and on a news site, the voices the audience already trusts are your own editors and reporters.

The taz play is the cleanest case study in the European market: clone the voices of named editors, attach each clone to a documented, revocable consent, and ship the editor’s actual voice narrating their own column. Seventy percent of taz readers now listen rather than read. The voice is not the gimmick — the voice is the reason the reader stays through the byline they came for.

For a newsroom that doesn’t want to clone editors, the second-best move is to pick a named, persistent voice per section. The same voice reads the politics section every day. The same voice reads the sports section. Repetition compounds familiarity; familiarity compounds session length. Rotating randomly through generic provider voices breaks that compounding effect.

3. Match the voice to the moment — orchestrate per story type

Even with cloned editor voices, no single voice wins every cell of the matrix. A breaking-news flash, a long-form investigation, an evening briefing, and a sports recap each have a different cadence — and the voice that fits the cadence holds attention longer than the voice that fights it.

This is where AI voice orchestration matters operationally. The control layer above every voice provider routes each article to the voice that wins that specific cell:

• Investigations → the deep, deliberate editor clone.
• Breaking news → the brisk, high-clarity neural voice.
• Evening briefings → the warmer conversational voice.
• Sports recaps → the punchy, faster cadence.

The result reads to the listener as editorial judgement, not vendor choice. The fly-by who would have bounced off a tonal mismatch finishes the article because the voice fits the genre. For the architecture behind this move, see also our piece on text to speech for publishers and the orchestration layer.

4. Ship a podcast feed alongside the on-site player

The on-site player wins the on-site session. The podcast feed wins the away-from-tab session — and the away-from-tab session is where background audio actually lives.

Concretely, every newsroom shipping audio articles should also ship an auto-generated, per-publisher podcast feed to Apple Podcasts, Spotify, and the open-podcast ecosystem. The same articles, the same voices, the same brand. The reader who listened to two pieces on your site subscribes to the feed; the next morning, your editor’s voice is the first thing in their car commute.

This is the lever that converts audio articles from a website feature into a distribution channel. It compounds because podcast listeners have an order of magnitude longer average session than scroll-bounce readers. They don’t fly by. They listen through to the end of the episode because that’s the medium’s grammar.

Edison Research’s Infinite Dial 2025 puts that surface in context: 73% of Americans aged twelve and over have now consumed a podcast, 55% listen monthly, and 40% listen weekly — an estimated 210 million people, with the weekly number up from 15% in 2017^[4]. The trajectory isn’t subtle.

The publishers in BotTalk’s network who turn on the podcast feed see a second curve open up alongside on-site listening — different listener profile, longer sessions, and a brand surface in the listener’s car that the publisher’s home page never reaches.

5. Catch pronunciation failures before synthesis ships

A mispronounced name kills a session faster than a tonal mismatch. The fly-by who hears “B-M-W i-X-one” stays; the fly-by who hears “biX-1” laughs and closes the tab. Long-session retention is a function of how often the listener has to forgive the audio.

The operational move is to run a pre-synthesis quality engine — the article is inspected and normalised before a single voice model sees it. At BotTalk this is five checks per article:

• Numbers — currencies, dates, percentages, decimals normalised.
• Tone Shift — abrupt register changes flagged before synthesis.
• Phonetics — uncommon spellings mapped to phonetic transcriptions.
• Dialect — Austrian, Swiss, Berlin, Bavarian register decisions per article.
• Dictionary — 50,000-entry pronunciation dictionary with publisher-specific entries for local names, neighbourhoods, sports teams.

This is the unglamorous lever. It doesn’t ship a new feature, it prevents an old failure. But the difference in average completion rate between a pipeline with the quality engine and one without is the difference between fly-by content and full-listen content. The reader hears the article they expected. The session compounds.

The session lengthens because the medium fits the moment.

What changes when audio articles run on these five levers

Numbers from the BotTalk network, June 2026:

• 24,000 hours of attention captured per day across the network — the cumulative effect of audio articles holding listeners past the text fly-by point.
• 6,000 articles narrated per day across 30 publishers (200 per publisher per day on average).
• 5 AI voice providers routed through one orchestration policy so the right voice fits each story.
• Zero customer-facing outages through three documented provider incidents in the last twelve months — listeners never noticed.

The number to anchor on is the 75% completion rate. That is what “audio articles fuel longer sessions” looks like when you multiply it across a day of publishing. Not a per-article win — a per-newsroom attention compounding curve.

Reuters Institute’s Digital News Report 2025 gives the editorial half of the same picture from the listener’s side: 73% of news-podcast listeners say the format helps them understand issues more deeply than the equivalent text^[3]. The longer engaged time isn’t background noise. It’s real attention.

Two publishers on what changed

Two newsrooms. Two different reasons audio compounded. Both kept the reader past the text fly-by cliff.

How to know if your audio is actually compounding sessions

A four-question audit for any newsroom that has already shipped audio articles but isn’t yet seeing the session lift the BotTalk network sees.

1. Where on the page is the play button? Below the lede is too late. Above the lede, with duration visible, sticky on scroll, is the bar.
2. Whose voice is reading the article? If it’s the same provider default voice every newsroom is using, you’re not differentiating. Named editors or named per-section voices compound; rotating generic voices doesn’t.
3. Is there a podcast feed? If audio only lives on the article page, you’re capturing on-site sessions and ignoring the much larger background-listening session. Ship the feed.
4. What runs before the voice model? If raw newsroom copy goes straight to ElevenLabs or Polly, listeners are forgiving an average of two pronunciation errors per article. A quality engine removes that tax.

Four questions. Ten minutes. Most “we have audio articles” pitches fail two or more.