JSON-LD Structured Data: The Quality Method for SEO and GEO
🇫🇷 Lire en français : Données structurées JSON-LD : la méthode qualité SEO et GEO
On 7 May 2026, Google removed FAQ rich results from its results — those accordions of questions and answers that expanded a listing within the page. Many concluded that FAQ markup was dead. It isn’t: the FAQPage type is still valid, Google keeps reading it, and it now serves AI engines above all. This turnaround sums up the whole stake of structured data in 2026: the same JSON-LD tag works twice — once for Google’s understanding, once for citation by AI. Here is our quality process for producing it cleanly, on our own sites and our clients’, plus a tool to apply it yourself.
Note
During a recent project meeting in the Jarry area, with a team of web developers working on an editorial quality process, JSON-LD came up very quickly in the discussion. The shared observation: everyone knows how to add it, but almost no one validates or maintains it — without really knowing whether it’s an asset or a liability for their site.
What JSON-LD actually does (and why it counts twice in 2026)
Tip
In one sentence: JSON-LD doesn’t create content — it makes machine-readable what is already on your page, so that Google understands it and AI engines cite it.
JSON-LD is structured data: a small block of code, written in the vocabulary of schema.org, that describes your page to the engines. It states, in black and white: “this is a business, here’s its address and phone number”; “this is an article, here’s its author and date”; “here’s the breadcrumb”; “here’s a real list of questions and answers”. The content itself doesn’t change. We simply hand the machines a map of it.
This markup is not a direct ranking factor. It won’t move you up a notch on a query. What it does is resolve ambiguity: an engine that understands your page precisely uses it better, and stays eligible for the enhanced displays that remain. Until now, this was a purely Google matter.
Important
Google is already testing AI Mode in the United States and Canada — a version of its engine that answers like an AI: conversational, synthetic, citing its sources rather than listing ten blue links. When it lands here, well-structured pages are the ones that will stay visible: get your site ready now.
What has changed is the arrival of a second reader. Search is becoming a web of answers as much as a web of links: AI Overview at the top of Google, ChatGPT, Perplexity, Gemini all synthesise and cite sources. Hence the double score for truly compliant sites: your page is now graded twice, by Google’s index and by AI engines. The right move isn’t to pick a camp, but to serve both with the same rigour. And the good news: what’s readable for the machine — a clear answer up front, explicit headings, well-identified entities — is also more readable for your clients.
The FAQ, a textbook case of structured data
The frequently asked questions page is an old web convention, inherited from forums and mailing lists: you grouped the questions that kept coming back, so as not to answer them ten times. When schema.org standardised structured data, this convention became a type in its own right — FAQPage — a way of telling the machine “this is a question, and here is its answer”.
Google seized on it in May 2019: it turned this markup into a rich result, the expandable accordion under the blue link. The effect was spectacular — two to three times more room in the page. Within a few months, FAQPage became one of the most widely deployed markups on the web… and one of the most misused, dropped just about everywhere, including on pages that had no real question to ask.
It’s that excess that did the feature in. In August 2023, Google restricted it to “authoritative” government and health sites only; for everyone else, the accordion vanished within a few weeks. Then, on 7 May 2026, it removed it for everyone — those sites included — via a simple banner added at the top of its documentation, with no fanfare. A three-stage sunset: the display in May, the Search Console reporting in June, the API in August.
The lesson of this trajectory applies to all of JSON-LD: markup only has value as long as it honestly describes the page. The day it serves mainly to grab space in Google, it ends up removed.
What the end of FAQ rich results changes (and doesn’t)
Important
Since 7 May 2026, Google no longer displays FAQ rich results. The FAQPage markup, however, is still valid and read: don’t remove it if it describes real questions and answers visible on the page.
The point many miss: the FAQPage type has not been removed. It’s the visual display that stops, not the markup. Google keeps reading the schema, and states that unused structured data poses no problem — no need to go cleaning everything out, as long as it describes real questions and answers that are shown.
Where the tag keeps its value is on the AI side. Generative engines pick up mainly the visible text of your questions and answers; the schema comes as support, as a structural signal — weighted more on some engines than others, at the time of writing. The conclusion fits in one sentence: you no longer mark up a FAQ for Google, you mark it up for the machines that remain — provided the FAQ is real and shown.
Our JSON-LD quality process, step by step
Badly managed JSON-LD doesn’t show: the page renders, the code is there, and no one notices it’s telling nonsense. Quality, here, is a method. Here’s ours.
1. Model the entities before writing a single line of code. Who are you (Organization, or LocalBusiness for a locally rooted activity)? What does the page contain? We link the identity to its external proofs via sameAs — Google Business Profile listing, LinkedIn page, official profiles. That’s what anchors your entity, for Google as for AI.
2. Choose the right schemas, by page type. A homepage doesn’t have the same needs as an article. In practice: Organization / LocalBusiness on identity pages; Article or BlogPosting on posts; BreadcrumbList for the breadcrumb; FAQPage only if real questions and answers are shown. One well-chosen type beats ten types piled up.
3. Apply the golden rule: the markup describes what’s visible. Never information absent from the page. Never a rating or a review that doesn’t exist. Never a FAQ marked up but invisible on screen. It’s the rule Google keeps repeating, and the one that separates a serious site from one that gets ignored.
4. Validate, then check consistency. We run the markup through the schema.org validator, and check that the identity is rigorously the same everywhere — the famous NAP (name, address, phone), which we detail in our article on local visibility in Guadeloupe and our Google Business Profile guide. One address mismatch between your site and your listing, and trust erodes.
5. Maintain. A page lives; its markup must live with it. A schema frozen on a page that has changed becomes a silent lie — exactly the kind of signal engines are learning to penalise.
The pitfalls that sabotage your structured data
Warning
Blackhat SEO recipes that misuse JSON-LD and other standards produce the opposite of the intended effect: engines and AI eventually catch you out and withdraw their trust.
The most common trap: a schema that’s incomplete, not respected, or that describes something other than the page. A FAQ marked up but invisible, invented reviews, information absent from the text. Google ignores it — and along the way, it dents the engines’ trust in all of your markup.
The errors we see most often come down to five points. FAQPage laid down for a rich result that no longer exists: since May 2026, adding it to “grab space in Google” makes no sense — it only makes sense if the FAQ is genuinely shown and useful. The FAQPage / QAPage confusion: FAQPage is you giving the official answer (service page, article); QAPage is a page where several users answer one and the same question (a forum, a discussion thread). Choosing the wrong type makes the tag purely and simply ignored. The generator copy-paste with no review: generic entities, false sameAs, values left at default. Over-markup, piling up types to “look rich”, which muddies more than it helps. And data injected in JSON-LD only, invisible to the visitor: that’s the direct violation of the golden rule.
None of these pitfalls is dramatic on its own. Stacked end to end, they turn a technical asset into a liability: code that gives the illusion of seriousness and works against you.
The skill and the prompt we use (and that you can reuse)
To hold this process at scale — our sites and our clients’ — we rely on a skill: a tool that produces clean JSON-LD by page type, from the real content and the company’s identity sheet. It generates the Organization or the LocalBusiness with its sameAs, the Article for posts, the BreadcrumbList, and only adds a FAQPage if real questions and answers are present. Above all, it builds in the guardrails: no information absent from the page, the right type in the right place, the golden rule applied by default.
The other half of the work is the audit. Here’s the prompt you can paste into ChatGPT or Claude to put one of your pages under scrutiny:
You are an expert in schema.org structured data. Here is the content and the JSON-LD of my page: [PASTE THE URL OR THE CODE].
Check point by point:
1. Is each marked-up entity actually visible on the page?
2. Are the schema.org types the right ones (Organization, LocalBusiness, Article, BreadcrumbList, FAQPage)?
3. Does the FAQPage match real questions and answers that are shown (and not QAPage)?
4. Is the identity consistent (name, address, phone, sameAs links)?
5. Are there any JSON-LD syntax errors?
Give me a quality score out of 10, the list of problems in order of severity, and the corrected JSON-LD.We’ve gathered the full skill, the audit prompt and a “JSON-LD quality process” checklist into a kit we keep on hand. If you’d like to receive it, signing up to our newsletter is the simplest way — we send it to you directly.
Clean JSON-LD isn’t an expert’s gadget: it’s good housekeeping for a site, the same rigour that makes you proofread a text before publishing it. In the age of generated answers, that housekeeping shows twice. If you want to know what your pages really tell the machines, that’s exactly what we look at together.