The web is about to start talking

For thirty years a web page has been a thing you look at, in one language, silently.

All three of those assumptions are expiring at the same time, and the studios that treat them as optional extras are about to find themselves shipping the 2019 version of a website at 2026 prices.

Part one: one language stops being a default

Publishing in a second language used to be a budget line. Professional translation ran ten to twenty-five cents a word, so a modest site was a few thousand dollars before anyone touched the CMS — and it recurred, quietly, on every content update forever.

That cost has collapsed to approximately nothing.

When a cost goes to zero, the thing it was gatekeeping becomes a default. Nobody itemises HTTPS any more, or mobile layout, or image compression. Those were all line items once. They became table stakes the moment they stopped being expensive, and a site without them now reads as broken rather than as basic.

Multilingual is next, and it is further along than most people notice. 62 million people search Google from inside the United States in Spanish. In the US, 27% of small businesses still have no website at all — and a large share of the ones that do have exactly one language serving a customer base that does not.

The twist: translation is what stops working

Here is the part that gets missed. The same collapse that makes a second language free makes a translated second language worthless as a differentiator.

If translation costs nothing, everyone has it. Your competitor has it. The AI answering the query has it. Machine-translated Spanish is now the floor, not the ceiling, and readers identify it in about one sentence — the register is off, the idiom is Spain's or nobody's, the examples are all American, and the whole page has the tone of a document rather than a page written for a person.

So the bar moves. It moves from do you have Spanish to is your Spanish written.

And it moves further than copy. A genuinely localised page has different content, not different words for the same content.

What that looks like in practice

We shipped Rounds this month — a daily news briefing with an English edition and a Spanish one. The Spanish edition is not the English page in Spanish. Almost nothing crosses over.

That last row is the whole argument in one word. A translation engine turns "Wire" into Cable. A Spanish newsroom calls it el teletipo. One of those is a translation and the other is the language.

And the history module is the tell: on the same August morning, the English edition surfaced a Colorado mine spill and the Spanish edition surfaced a Spanish Civil War naval convoy and the death of Friedrich Engels. Same feature, same date, different past. What a culture remembers about a date is content, and no translation layer can generate it.

A translated site hands a Spanish reader an American morning in Spanish. That is a different product, and they can tell.

Part two: pages will read themselves

The second default is louder, in the literal sense.

Every major browser has shipped a speech synthesis API for years. Almost nobody used it, because the voices were terrible — flat, mispronounced, robotic in a way that made listening more work than reading. Read-aloud stayed where bad voices are tolerated because the alternative is nothing: assistive technology.

Neural voices fixed the quality problem. The reason nobody shipped read-aloud has quietly stopped being true, and the demand it was failing to serve was never small. People already listen to text in enormous volume — podcasts, audiobooks, saved-article apps, screen readers, the read-aloud built into every phone. The behaviour is established. The page just wasn't participating in it.

You are already being read aloud

This is the argument that should actually move you.

Ask an assistant about your business today and it reads your page out loud — in a synthetic voice you did not choose, with your product names pronounced however the model guesses, at a pace you did not set, with your structure flattened into whatever the model thought was important.

That is happening now, at growing volume. AI Overviews appear on 40–46% of queries, and ChatGPT is now roughly the third most common way people find local businesses. A large and rising share of the audience for your words never sees your page at all.

You do not get to opt out of being read aloud. You only get to choose whether it is your voice doing it.

What has to be true first

Read-aloud is not a widget you bolt on. Turn TTS loose on a typical page and it will earnestly recite the skip link, the cookie banner, the nav, "Home Services About Contact", every icon with a missing label, and then the article — which is a worse experience than silence, and is exactly why the first wave of read-aloud buttons got ignored.

The page has to be built so a voice can move through it:

• Real semantic structure. <main>, <nav>, <article>, headings in order and not chosen for size. A voice has no peripheral vision — heading order is the layout.
• A correct lang attribute, and lang on foreign phrases inside the text. This is the same hook that makes a page multilingual and the one that makes it pronounceable. A Spanish name inside an English sentence is either marked up or mangled.
• Alt text that reads as a sentence, because it will be read as one.
• Abbreviations and numbers that survive being spoken. "$1.2M" and "e.g." are visual shorthand; out loud they are a stumble.
• No content that only exists in a layout. Anything carried by position, colour, or a hover state does not exist to a listener.

That list is not new. It is the same discipline that has always served screen readers, and it is close to identical to what makes a page extractable by an AI engine. One investment, three payoffs: accessibility, voice, and citation. It is rare that those line up this cleanly, and it is a good reason to stop treating any of the three as a compliance chore.

Where the voice should not go

We are not neutral about this and we have already turned it down once.

Cuyar is our movement app for adults over 65, and it talks constantly — nine coach voices guiding people through exercises. We evaluated browser text-to-speech for it and rejected it: the cadence is wrong for instruction, it degrades unpredictably across devices, and it cannot be relied on offline. Every voice line in that app is recorded audio shipped with the product.

That decision holds, and it draws the line precisely.

Narration tolerates error. Instruction does not. If a voice misreads a comma while reading you a news brief, you lose a beat. If it misreads a pause while telling a 78-year-old how to lower themselves into a chair, that is a fall. Synthetic narration will become a default for articles, briefings, documentation, and product copy long before it should be trusted anywhere a wrong syllable has a physical consequence.

The web is going to start talking. That does not mean every page should say everything out loud.

The short version

1. A second language is no longer a budget decision, so it stops being optional — and because it stops being optional, machine translation stops being an advantage. Write the second language.
2. Localisation is a content decision. Different sources, different examples, different references. Same words in a different language is the cheap version, and it reads that way.
3. Your page is already being read aloud by machines you do not control. Shipping your own voice layer is how you take that back.
4. The work that makes a page speakable is the work that makes it accessible and citable. Do it once.

None of this is a prediction about some distant web. Two of the four are true today, and the other two are a build cycle away.

We build bilingual sites where the Spanish is written, not translated — see how we do it, or look at Rounds, where the two editions do not share a single source.