AI Schema Generator Fixes: Review Generated Markup Before It Goes Live
The AI Schema Generator reads the content of a page with an AI model and writes JSON-LD structured data to match it, then validates the result against schema.org and Google's rich result criteria before handing it over.
That saves a lot of typing, but a language model can still pick the wrong type, miss a field that is on the page, or fill a gap with a value that is not true. This page explains how generated markup goes wrong, how to review it quickly, and which guide fixes each finding.
Why generated markup needs a human check
An AI model writing schema is doing two jobs at once: understanding the page and translating it into a strict vocabulary. Mistakes in generated markup almost always come from one of four habits of language models, and knowing them makes review much faster.
- It fills gaps. Asked for a complete Recipe or Product, a model may supply a plausible cooking time, rating or author name when the page does not state one. The output looks tidy and validates perfectly, which is exactly what makes it dangerous.
- It guesses the type from surface features. A recipe written in a chatty blog style can be labelled BlogPosting; a product comparison can be labelled Product. The model reads the style of the text as much as its purpose.
- It has no sense of today's date. A model's knowledge stops at a training cutoff. Unless the date is taken from the page or the system, it can produce a year that is out of date, or a date without a time zone.
- It does not know your site. The generator sees one page. It cannot know that your SEO plugin already outputs an Organization, that your authors have their own pages, or which URL is canonical, unless you tell it.
Validation catches format problems, such as a missing required field or a malformed date. It cannot catch a value that is well formed but false. That part needs a person who knows the page.
A five-minute review of generated markup
Before pasting generated JSON-LD into a page, put it side by side with the page and check these properties in this order. They are where generated markup is most often wrong, and a mistake in each has a different cost.
| Check | Compare it with | Why it matters |
|---|---|---|
@type | What the page is mainly about | The wrong type makes every other field irrelevant to the result you want |
| Ratings, review counts, prices | Numbers actually printed on the page | Invented figures break Google's guidelines |
author and publisher | The visible byline and your organisation | A made-up or text-only author weakens trust signals and loses links to author pages |
| Dates | The published and updated dates shown, or your CMS | A wrong year or missing time zone misleads search engines about freshness |
url, image, @id | The canonical URL and real image files | Relative, broken or non-canonical URLs point search engines at the wrong resources |
| Required fields | The rich result documentation for the type | A missing required field removes eligibility even when all else is right |
If you review in bulk, for example after generating markup for a hundred old blog posts, sample several pages from each template rather than checking one. Errors in generated markup tend to cluster: a template with no visible byline will give every page the same author problem.
The fix guides
Each guide covers one kind of finding in generated markup, how to spot it and how to correct either the output or the generation settings.
Accuracy of what was generated
Fixing the wrong schema type covers the page patterns that mislead type detection and how to override the type before or after generation. Fixing hallucinated properties covers finding and removing invented authors, dates, ratings and other values, and the checking routine that catches them. Fixing missing required fields covers the opposite problem: information that is on the page but was not picked up, such as ingredients in an image caption or prices loaded by script.
Format of the values
Fixing incorrect URL references covers relative paths, links to the wrong version of a page and image URLs that do not resolve. Fixing date format errors covers ISO 8601 format, missing time zones and wrong years.
Fitting into the rest of your site
Fixing duplicate schema generation covers generated markup that clashes with what a plugin or theme already outputs. Fixing Organization and Person reference loss covers authors and publishers that come out as plain text instead of linked entities. Fixing rich result eligibility gaps covers markup that is valid schema.org but is missing what Google asks for, and the difference between required and recommended fields.
Workflows by platform and content type
WordPress covers generating markup per post, working alongside an SEO plugin and handling older content in bulk. Shopify covers product pages, review apps and exporting and importing markup. Editorial content covers Article and NewsArticle markup, author pages and keeping published and updated dates right.
Fitting generated markup around what you already have
Most sites already output some structured data before the generator is involved. A WordPress SEO plugin usually adds WebSite, WebPage, Organization and often Article; a Shopify theme usually adds Product; a review app may add its own ratings. Pasting generated markup on top creates two descriptions of the same thing, and the two rarely agree.
Decide ownership type by type before you start:
- Site-wide entities stay with one source. Your Organization, WebSite and author Person
entities should come from a single place, usually the SEO plugin or theme, with a fixed
@idfor each. - Use the generator for what is missing. Typical good uses are types your plugin does not handle: Recipe details, Event data, Course information, or richer Product properties.
- Link, do not repeat. Where generated markup needs an author or publisher, replace
the generated copy with a reference to the existing
@id, so there is one organisation and one author, not several near-copies. - Turn off what you replace. If generated markup does a type better than the plugin, switch that type off in the plugin rather than running both.
Where the generated block will live matters too. Markup stored in a page's custom field or content will not update when the page does, so a price change or a new updated date needs the markup regenerating or editing. For facts that change often, such as stock or price, markup produced by your platform from live data is usually safer than a generated snapshot.
Which findings to fix first
- Invented values. Ratings, reviews, prices or people that do not appear on the page. These carry a policy risk, so remove them before anything else, even if it costs a rich result.
- Wrong type. Correct the type next, because it decides which fields matter.
- Duplicates with existing markup. Resolve these before refining fields, or you will be fixing two copies.
- Missing required fields. Fill them from real page content so the page becomes eligible.
- Author and organisation references. Turn plain-text names into linked entities.
- URLs and dates. Correct format issues, usually quick wins once the structure is right.
- Recommended fields. Add those that are genuinely on the page to improve how eligible results display.
A sensible order of work
- Run a page through a structured data checker to record what markup it already has and where it comes from.
- Open the AI Schema Generator and generate markup for the page, setting the type yourself if the page is ambiguous.
- Review the output with the table above, deleting any value you cannot find on the page.
- Replace generated Organization and author details with references to your site-wide entities.
- Correct URLs to absolute canonical addresses and dates to ISO 8601 with a time zone.
- Remove or switch off any other source producing the same type.
- Add the markup through the template or field that suits your platform, then test the live page.
- For bulk work, generate a small batch first, review it, fix the settings, and only then generate the rest.
The AI Schema Guide covers the generator in depth. To check or repair markup once it is live, see the JSON-LD Schema Guide and the Schema Debugger Guide; to build markup type by type by hand, see the Schema Builder Guide.
1. Is AI-generated schema allowed by Google?
Google cares about whether structured data is accurate and follows its guidelines, not how it was produced. Markup written by an AI tool, a plugin or a person is judged the same way: it must describe content visible on the page, use the right type and avoid misleading values. The risk with AI generation is that invented details can slip in unnoticed, so the review step matters more than it would with hand-written markup.
2. Why did the generator add a rating my page does not have?
Language models are trained to produce complete, plausible output, and a Product or Recipe with a rating is a very common pattern in the examples they learned from. When the page has no rating, the model may supply one to fit that pattern. Delete it. Ratings and review counts in markup must come from genuine reviews shown on the page, and invented ones break Google's rules on misleading structured data, which can cost rich results across the whole site.
3. Can I generate schema for every page on my site at once?
You can, but do it in stages. Generate markup for a handful of pages from each template, review them carefully, and adjust the settings or type choices based on what you find. Then generate the rest and spot-check a sample from each batch. Problems in generated markup tend to repeat across similar pages, so catching them early saves correcting hundreds of pages later. Keep a note of which pages you have reviewed so nothing is published unchecked.
4. The markup validates, so why does the checker still flag it?
Validating against schema.org only confirms that the types and properties exist and are used in allowed ways. Google's rich result features have extra requirements, and some fields that schema.org treats as optional are required for a particular result. The markup can therefore be valid schema.org but still not eligible for the rich result you want. The eligibility gap guide explains which fields Google requires for each type and how to fill them from content that is genuinely on the page.
5. Should I let the AI choose the schema type?
For clear-cut pages, such as a standard product page or a recipe with ingredients and steps, the generator usually chooses well. For pages that mix purposes, such as a blog post that includes a recipe or a service page with a price list, set the type yourself. You know what the page is mainly for and which search result you want it to be eligible for, and the model can only infer that from the wording.
6. Why are the dates in my generated markup wrong?
A language model has no built-in knowledge of today's date and learned from text written before its training cutoff. If it infers a date rather than reading it from the page, it can produce the wrong year or leave out the time zone. The fix is to take dates from the page or your CMS, format them as ISO 8601 with an offset, and check the published and modified dates on every generated block.
7. Will AI-generated schema help my site appear in ChatGPT or other AI answers?
Accurate structured data can help any system that reads it understand what a page is about, and some AI tools do read JSON-LD. But AI providers say little about how they use it, and no markup guarantees a mention in an AI answer. Treat structured data as a way to state facts clearly and consistently. The same accuracy rules apply: markup that disagrees with the visible page is more likely to confuse than help.
The AI Schema Generator reads your page content with an AI model, writes JSON-LD tailored to the content type and validates it against schema.org and Google rich result criteria before delivery, ready for you to review and publish.
Run the AI Schema Generator →