Speakable and FAQ rich results: two levers that stopped paying

On 7 May 2026, Google stopped showing FAQ rich results, and on 8 May it wrote the deprecation notice into its Search Central changelog. On 15 June the documentation page itself was removed.

I read that entry the week it landed. Then I opened the audit sheet for a hotel site I was midway through. One row said "add FAQPage to twelve pages, expected lift: rich result".

So that row was already wrong the day I wrote it.

And the row under it was worse. It said "add speakable to the blog", copied from a checklist I had been using since long before anyone said the letters GEO.

I did not know if either line had ever paid. This post is the answer I got when I went and checked, and it changed what NovaTechRay puts on an audit for an existing site.

Step 1: What speakable was for

The schema.org definition is honest about its scope: the property marks the sections of a page that are most appropriate for text-to-speech.

That is the whole design. It points at a headline and a summary with a CSS selector or an XPath, so a device can read them aloud without reading the navigation.

Google's own speakable documentation was still labelled BETA when I fetched it on 25 September 2026. Google's page says the consumer is the Google Assistant answering topical news queries on smart speakers, up to three articles per question.

News, smart speakers, one language, one market, publishers only.

But the checklists never carried those limits across. They carried the word "speakable", and speakable sounds like it should matter to an assistant that talks.

Step 2: Why to skip it now

Nothing in the speakable pipeline touches a hotel, a cafe or a tour operator, and nothing in it touches the assistants that recommend businesses in text.

The schema.org page reports usage of speakable at somewhere between 100K and 1M domains in its August 2026 aggregation from Google's index. Lots of people shipped it. That number says nothing about anyone getting paid for it.

I shipped it too, on my own site, and I have no evidence it did a single thing.

So the instruction for an existing site is short. If speakable is already in your JSON-LD, leave it, because it is inert and removing it is a deploy for nothing.

If it is on a proposal as a deliverable, strike the line and ask what else on that proposal was copied from a news publisher's playbook.

Delete less than you think.

That sounds backwards for a post about levers that stopped paying, so here is the defence. Valid, true structured data costs nothing to serve and cannot be penalised, which is the same position I took on schema for AI search.

Stale markup is cheap. Billing someone to add markup whose only consumer has gone is the expensive mistake.

Step 3: FAQ rich results are gone

The changelog entry is plain. Google added a deprecation notice on 8 May 2026, the feature stopped appearing in results from 7 May, and the documentation was removed on 15 June with the note that the feature is no longer shown.

The URL for the old FAQPage guide now redirects to the changelog anchor for that removal. I checked it this morning, and a redirect like that is the cleanest signal a platform can send.

But I want to be fair to the tactic, because it worked for years.

An FAQ rich result took two extra lines in a search listing, pushed a competitor down, and was cheap to earn. It was a good lever while it paid.

Then it stopped, and most of the audit templates I have seen this summer have not noticed.

Step 4: What that changes on your page

The rich result was the reason people wrote FAQs that nobody would read on the page itself.

Five questions in an accordion, collapsed by default, written for the snippet rather than for a guest.

That incentive is gone, and good riddance, because it produced pages where the useful answer lived in a hidden block and the visible page said nothing.

So the FAQ has to earn its place as content now.

Open by default. Above the footer. Answering the questions a front desk hears twice a day: late check-in, parking, whether the kitchen does a vegan option, whether a dog is fine.

And write each answer as a complete statement that survives being lifted out alone, because that is how a retrieval system will read it.

Ahrefs found in its brand-mention study that a brand's mentions across the web correlate with AI citation at r=0.66, far ahead of backlinks at 0.22. A self-contained answer is the unit that gets repeated.

Step 5: Keep FAQPage, generated from the visible FAQ

Here is where I nearly got it wrong. In June my first draft of the new audit said "remove FAQPage, feature retired", and I sat on it for a week because something felt off.

The thing that felt off was this: the rich result was the reward, but it was never the reason the markup was readable.

FAQPage is still the least ambiguous way to tell a parser that a question exists on this page and that the following text is its answer.

Google's structured data policies still stand. They still say markup must describe visible content.

So the rule I landed on has two halves. Keep FAQPage, generate it from the FAQ that is visible on the page, one entry per visible question, answers matching word for word, and expect nothing to appear in a search listing because of it.

I did not know if it would work as a machine-reading aid when I kept it, and I still cannot show you a controlled result.

What I can show you is a page that says the same thing in prose and in data, which is the cheapest form of corroboration there is.

Step 6: llms.txt, for completeness

Every audit I inherit this year has llms.txt on it, so it goes on the same table.

Ahrefs checked 137,210 domains in a study published 15 June 2026 by Louise Linehan and found 97 percent of the llms.txt files it discovered received zero requests.

Google's changelog note from 15 June says the file is not needed for Google Search and neither helps nor hurts visibility.

So it sits in the same bucket as speakable: inert if present, not worth a line item. I have written up the one case where it might help if you want the long version.

The pattern across all three is the same. A real feature existed, a checklist absorbed it, the feature went away, and the checklist kept billing.

Step 7: The audit list for an existing site

This is the table NovaTechRay now runs before anything else on a site that already has markup.

Speakable present: leave it, do not extend it, strike it from any proposal.

FAQPage present: keep it, diff it against the visible FAQ, fix any answer that does not match, and set the expectation to zero rich results.

FAQPage absent: add it only if there is a visible FAQ worth marking up, and write the FAQ first.

llms.txt present: check the access log for requests before you touch it, and delete it if it has gone stale.

Then spend the freed hours on the things that still read. The paragraph under the H1, server-rendered facts, an explicit price where you can stomach one, and consistent details across every profile that mentions you.

The hotel from the opening got exactly that list. The twelve-page FAQPage row became a four-page row tied to real questions, and the speakable row was struck with a note explaining why, which is the part clients remember.

The short version

• Skip speakable; its only consumer was a news feature for smart speakers.
• Stop promising FAQ rich results; Google retired them on 7 May 2026.
• Keep FAQPage, generated from a visible FAQ, answers matching word for word.
• Write each FAQ answer so it stands alone when lifted off the page.
• Treat llms.txt as inert until your access log says otherwise.
• Strike every retired lever from your proposal and say why in writing.
• Move the hours to the paragraph under the H1 and the facts a machine can quote.

Sorting retired levers from live ones is most of what NovaTechRay does at novatechray.com.