The short answer
No, not the way it is being sold. Schema markup did not make my content findable, and I have data rather than an opinion on that now.
I ran a controlled test where I hid the same unique nonsense string in nine different places across three pages, then watched which places Google could find it in. Plain paragraphs, table cells, image alt text, content inside collapsed FAQ blocks, even content written onto the page by JavaScript after load, all found within 48 hours. The string inside JSON-LD structured data was never found on any page. I checked again at the next checkpoint and it still had not been.
If a human reader can eventually see the words, Google could find them. If only a machine could see them, Google could not. That one line explains every result in the test, including the odd ones.
What Google actually says
This is not a secret. Google says it in its own documentation for AI features in Search, in a sentence anyone can read for free:
"Structured data isn't required for generative AI search, and there's no special schema.org markup you need to add."
That is about as direct as Google gets. It appears alongside four other things Google explicitly says you do not need, which I went through in detail in what Google actually says about ranking in AI answers.
The trouble is that plenty of people are still selling schema markup as an AI visibility tactic. Some of them are selling it as the AI visibility tactic. And a documented statement from Google is evidence, but it is not proof. Google has been wrong about its own systems before, and companies have an obvious interest in downplaying anything that looks like a lever.
Why I tested it anyway
Rather than take the sentence at face value, I built something that could contradict it.
I published three reference pages on my own site and hid a unique nonsense string in nine different places on each. Strings like the one further down this page. They exist nowhere else on the internet, so if a search engine ever returns one, the only possible explanation is that it read my page. There is no other source it could have come from.
All nine placements sit on the same page, so page level factors like topic, links and authority apply equally to every one of them. Placement is the only thing that changes. Three pages rather than one, so a single fluke does not read as a finding. Three more strings were generated and published nowhere at all, as a check that the engines were not simply inventing agreement.
The full design, the predictions I wrote down before I had any results, and the running checkpoint log are all on the study page.
What happened
Almost everything I expected to fail worked, and the thing I expected to fail was the only thing that did.
| Where the string was hidden | Found by Google |
|---|---|
| Plain paragraph text | All 3 pages |
| Image alt attribute | All 3 pages |
| Inside a table cell | All 3 pages |
| Inside a collapsed FAQ block | All 3 pages |
| Written in by JavaScript after load | All 3 pages |
| In an H2 heading | 2 of 3 pages |
| JSON-LD structured data only | None |
| In the llms.txt file | None |
| In a PDF version of the page | None |
Image alt text surprised me. I had predicted it would never be picked up, and it was on all three pages inside two days. JavaScript surprised me more. I had predicted over a month and Google ran the code and had the content in two days.
Schema markup was the only prediction I got right in the direction I wanted to be wrong about.
Google had it the whole time
Here is the part that changes the shape of the answer.
Search for the JSON-LD string in Google and you get nothing. But open the same URL in Search Console's live inspection and look at Google's own rendered copy of the page, and the string is right there in the stored HTML. Google fetched it, rendered it, parsed it and kept it.
So the honest version is not that Google ignores your structured data. Google reads it. It just never turns it into something a search can reach.
Stored and findable are two different states. A lot of AI visibility advice quietly assumes that getting content into an index is the same as getting it into an answer. My schema string has been sitting in Google's stored copy of the page since the day it went live, proving those are separate problems.
My marker is a meaningless string. That is what makes it a clean test of retrieval, and it is also what makes it useless for testing use. No real question would ever call for it, so it could never surface in a generated answer whether Google wanted to use it or not.
So this shows that schema content does not become findable. It does not show that schema content is never used. Another researcher running parallel experiments tells me he found the opposite on that second question, that meaningful procedural information placed only in JSON-LD did turn up in Google's answers. I have not verified that myself and my design cannot check it. Worth knowing before you take the headline too far.
Google's five claims, checked
Google names five things in that guide and says none of them help. My study only had the design to check some of them. Here is where each one stands, including the ones I have no data on.
| Google's claim | Did my study check it? | Result |
|---|---|---|
| Special schema.org markup for AI is not needed | Yes | Google's claim holds. Never findable on any page. |
| llms.txt and other AI text files are not needed | Yes | Google's claim holds. Never findable on any page. |
| Chunking content into tiny pieces is not needed | No | No data. My design cannot speak to it. |
| Writing in a special AI friendly style is not needed | No | No data. My design cannot speak to it. |
| Chasing inauthentic mentions does not help | No | No data. My design cannot speak to it. |
I would rather show that table with the blanks in it than round it up into a cleaner story. Where I checked Google's claim, Google was right. Where I did not check, I do not know, and neither does anybody selling you the opposite.
What schema is genuinely for
None of this makes structured data a waste of time. It makes it a different tool than it is being sold as.
It still earns rich results in classic search, and those are still real. Star ratings, FAQ dropdowns, breadcrumbs, event dates, product prices, the visual furniture that makes a listing take up more room than the one above it. Google's own sentence even says so: it is worth using for standard rich results, just not for this reason.
It also does something quieter that matters more for a small brand. Person, Organization and sameAs let you state plainly who an entity is and which profiles belong to it, in a form that leaves no room for interpretation. That is useful when a system is trying to work out whether two mentions of a name are the same person. It is entity hygiene. It is not a ranking lever, and it will not put you in an AI answer by itself.
So keep your schema. Validate it. Just stop paying anyone a premium for it on the grounds that ChatGPT is reading it.
What to do instead
The test points at something plainer than most GEO advice wants it to.
Every placement that worked had one thing in common. The words were part of what a reader would eventually see. The paragraph, the table, the alt text behind an image, the answer inside a collapsed block, the sentence JavaScript wrote onto the page. All visible to a person, all found. The two that failed were machine only channels that no reader ever encounters.
Which means the useful version of the advice is boring. Put the claim you want repeated in the visible body text of a normal HTML page, in plain words, near the question it answers. Do not bury it in markup. Do not hold it back for a PDF. Do not put it in a file only crawlers read.
If you want the checkable version of the underneath, whether the AI crawlers can reach you at all, I built a free checker that runs through it.
This page is also a test
Everything above came from pages published on 22 August. Those pages answered where content gets picked up from. They did not answer how fast, because I submitted them for indexing and could not cleanly separate my submission from Google's own crawl schedule.
So this page carries one more string, in plain body text, the placement that already works. The string is bugavip1670. It has never appeared anywhere on the internet before this page went live, which makes it a clean clock. I know exactly when this page was published, and I will know exactly when the string first becomes findable.
That gap is the thing being measured. I am running the first check a few hours after publication, alongside a question about this page that cannot be answered without actually reading it. Results will go on the study page either way, including if nothing happens at all.
If you found this page by searching for that string, the test worked, and you are welcome to say so.
Does schema markup help you show up in AI search?
Does Google read the schema markup on my page at all?
So what is schema markup actually for in 2026?
If schema does not work for AI search, what does?
Primary sources: Google Search Central, AI features and your website · Google Search Console URL Inspection, live test and stored HTML, August 2026 · My own ingestion study, three pages, 25 markers, published 22 August 2026, full method and checkpoint log. All Google documentation accessed September 2026.