A 1,500% lift in AI Overviews. Six out of seven AI platforms unable to read the data that supposedly caused it. The numbers look contradictory until you see how the data was actually collected.
We ran a three-month controlled experiment to test this exact question. We monitored brand coverage across 319 prompts in the US market, tracking a SaaS brand’s visibility across multiple AI search engines. The premise was simple: if we add robust structured data, will AI models directly ingest it and cite us more frequently? The results were less direct than expected, revealing a nuance that many marketers miss.
Why most AI platforms can’t parse JSON-LD
The first test asked a straightforward question: can these platforms read the code on the page? The results were stark. Out of seven major AI search systems, only one actually retrieved the structured data as intended.
We ran direct fetch tests to see if models could access the specific JSON-LD blocks we added. Gemini was the sole platform that successfully parsed the schema, accurately identifying the BlogPosting and WebPage tags. The other platforms failed in distinct ways. ChatGPT, Claude, and Perplexity could not interpret the data at all. One specific platform even blocked our fetch attempt entirely, redirecting the request to a manual extraction guide instead.
The most puzzling result came from Google AI Mode. It did not read the code; instead, it hallucinated a “Service” schema type that we had never implemented. This is a critical distinction. The model generated text that sounded like it was reading structured data, but it was actually fabricating the metadata. It was not interpreting the source; it was predicting what the source should look like.
| Platform | Fetch Result | Outcome |
|---|---|---|
| Gemini | Success | Correctly parsed BlogPosting/WebPage |
| Google AI Mode | Failure | Hallucinated “Service” schema |
| ChatGPT | Failure | Could not interpret data |
| Claude | Failure | Could not interpret data |
| Perplexity | Failure | Blocked fetch attempt |
| Microsoft Copilot | Failure | Accessed text only |
This failure is not a bug; it is a feature of the underlying architecture. Most AI search engines ingest content through HTML-to-Markdown pipelines designed to convert web pages into clean text for language models. These pipelines systematically strip out <script> tags to reduce noise. Since JSON-LD lives inside <script> blocks, the structured data is removed before the model ever sees it. The model simply doesn’t have the data to parse.
This explains why schema markup remains a tool for traditional SEO rather than a direct lever for generative search. The model reads the text on the page, not the metadata hidden in the code. If the pipeline strips the script, the schema is invisible to the AI.
How schema markup feeds the AI Overview trigger logic
The apparent paradox of a +1,500% lift in AI Overviews alongside a near-total failure rate in direct schema reading resolves when we look at the underlying plumbing. Schema did not boost visibility because AI models directly ingested the JSON-LD. Instead, the gains followed a traditional SEO pathway.
Structured data improves how search engines interpret page content, which increases eligibility for Rich Results. Stronger SERP presence then improves organic rankings, which AI Overviews use as a primary proxy for source quality. In this model, the AI system does not read the schema tags; it reads the resulting authority signal from the search index.
This indirect mechanism explains why the +377% increase in SERP features and the +1,500% rise in AI Overview appearances tracked each other so closely. These metrics align with known traditional SEO mechanisms rather than a new form of Generative Engine Optimization (GEO) consumption. If AI platforms were directly parsing structured data, we would expect a different pattern—one where schema presence correlates with citations even when organic rankings remain flat. That was not the case. The data suggests that AI Overviews inherit their trust and relevance scores from the same ranking factors that drive standard organic search, with schema acting as a clarifier for the crawler, not a direct input for the model.
Understanding this distinction is crucial for positioning your technical work. Google AI Overviews exist at the intersection of SEO, AEO, and GEO, but technical SEO remains the primary entry point. The structured data layer functions as a bridge between human-readable content and machine-interpretable signals. By improving the clarity of that bridge, you enhance your chances of being cited, not because the AI sees the code, but because the search engine is better able to vouch for your content’s relevance. This reinforces the value of technical hygiene while clarifying that direct “GEO” consumption of schema by most AI engines is still largely theoretical, with the exception of a few platforms that have integrated specific parsers.
Controlling for algorithmic shifts in your own schema experiments
When you see a sharp jump in AI Overviews or brand mentions, the first instinct is to credit the technical change you just deployed. We made that exact assumption at the start of our observation window. Then we looked at the data for competitors who had not touched their code. Their brand coverage moved in parallel with ours. That single observation shifted our entire interpretation.
The December lift was not driven by our new structured data; it was an algorithmic shift affecting the entire platform. If you are measuring the impact of schema markup on AI search visibility, tracking competitors is not optional. It is the control that separates a real technical win from a market-wide trend. Without it, you risk attributing a platform update to a specific intervention like adding JSON-LD tags.
The competitor control
During the study, we tracked brand coverage for several competitors in the same vertical. We implemented five types of schema, including FAQ and SoftwareApplication, on our SaaS pages. Our brand coverage rose from roughly 27% to 49%. At first, that looked like a direct result of the schema deployment. But when we checked the competitor data, we saw similar or even larger movements in their brand visibility during the same period.
These competitors had not changed their structured data. They did not add new Organization or Article markup. Yet their numbers moved. This parallel movement is the key indicator. When the entire market shifts, the lift is environmental, not structural. It suggests that Google’s algorithms or the AI models themselves changed how they select sources. Attributing that shift to your specific technical change is a misread. The only way to know if your structured data worked is to see your brand move while your competitors stay flat. If everyone moves, the algorithm moved.
Isolated lifts as the true signal
How do you know if your schema markup actually influenced your AI Overviews appearances? You look for an isolated lift. An isolated lift happens when your brand’s visibility increases while your competitors’ visibility remains stable or decreases. That divergence is the threshold for crediting a specific change.
In our case, the initial broad increase was not a valid signal for the schema’s effectiveness. It was a platform-wide event. We only gain confidence in a technical intervention when we see a divergence. If your brand jumps 20% while the top five competitors stay within a 2% range, that is a strong candidate for a technical cause. If all six brands jump 15%, that is an algorithmic event. This distinction is critical for SEO and AEO strategy. It prevents you from scaling an initiative that did not actually work. It also prevents you from stopping a valid initiative because a market correction masked its impact. You need to isolate the variable. The variable is your structured data. The control is the rest of the market. If the control moves, the result is not caused by the variable.
Beyond the schema
It is also worth noting that our brand coverage growth from 27% to 49% was not solely due to the schema test. Our broader GEO efforts, including content updates and entity consistency work, contributed to that increase. The schema was one component of a larger strategy. This multi-factor reality makes isolation even harder.
If you are testing a single technical change, ensure that no other major changes are happening simultaneously. If you are running a full GEO strategy, accept that you may not be able to isolate the specific contribution of structured data. In those cases, use the competitor control to validate the baseline. If the baseline is moving, you cannot claim the entire lift as your own. You can claim your relative position change, but not the absolute number. This nuance matters when reporting results to stakeholders. It keeps the narrative honest. It ensures that your technical team does not take credit for algorithmic shifts. It also ensures that your content team understands that their work is part of a larger ecosystem. The goal is not just to increase AI visibility, but to understand what drives it. Schema markup is a tool. It works within a system. To use it well, you need to see the whole system.
What to do with structured data in a GEO strategy
The core takeaway is clear: structured data is an SEO lever, not a GEO growth lever. Keep schema markup on your technical checklist for rich results and indexing signals, but remove it from the top of your GEO priority list. The data shows that most AI search engines do not consume your JSON-LD directly, so optimizing for it as a primary citation driver is a misallocation of effort.
Instead, redirect your resources toward the signals AI platforms actually use for citation selection. Focus on content clarity, external citation building, and entity consistency. When an AI model looks for a source, it evaluates how clearly your content answers the query, how many independent third-party sites reference your entity, and how consistent your brand name and details are across the web. These factors influence whether your content gets cited in AI Overviews or other generative answers. Structured data supports the technical foundation, but it is not the content signal that earns the citation.
The Gemini exception and AI Merchant Programs
One notable exception exists: Gemini is the only platform that successfully reads and retrieves schema markup during direct fetch tests. However, its share of total AI citations is currently small, making it a low-ROI optimization target for most brands.
If you are a product brand, the higher-leverage alternative is participating in AI Merchant Programs. Programs like OpenAI’s Merchant Program and Perplexity’s “Buy with Pro” allow you to integrate live product feeds directly into AI search results. This gives you a direct path to commerce visibility that schema markup cannot provide, bypassing the parsing issues entirely. For these platforms, the connection is direct and transactional, rather than reliant on generic structured data interpretation.
Does schema markup work for AI search in your industry
The answer depends less on what you add and more on how your platform ingests data. Here’s how to interpret the findings for your specific context.
Can ChatGPT or Perplexity read my schema markup?
No. Most AI search platforms strip script tags during their ingestion pipeline, meaning the models never see your JSON-LD. In our direct fetch tests, only Gemini successfully retrieved the structured data. ChatGPT, Perplexity, and Microsoft Copilot all failed to parse the schema, with some platforms redirecting to manual extraction or accessing only readable text. If you are optimizing specifically for these tools, structured data is not currently a viable channel for direct consumption.
Will adding schema increase my AI Overview appearances?
Possibly, but indirectly. The correlation we observed between schema implementation and AI Overviews visibility aligns with traditional SEO mechanics rather than direct AI citation. Structured data improves SERP feature eligibility and indexing signals, which in turn influences organic rankings. Since Google’s AI Overviews often use organic ranking as a proxy for source quality, better technical SEO can lead to more frequent citations. However, this is a byproduct of improved search visibility, not a result of the AI reading your tags. Don’t expect a direct cause-and-effect relationship between adding schema and immediate AI brand mentions.
Should I remove schema if I’m focusing on GEO?
No. Keep your structured data for technical SEO health and rich results. Removing it would degrade your traditional search presence without guaranteeing any gain in AI visibility. The value of schema lies in its role as a foundation for standard search features. If your goal is Generative Engine Optimization, prioritize content clarity, entity consistency, and external citations, which are the signals AI platforms actually use to select sources. Schema remains a necessary baseline, just not a primary lever for AI-specific growth.
Does this apply to ecommerce sites?
Caution is advised. Our data comes from a SaaS brand, and product or pricing schemas may behave differently in commerce contexts. Ecommerce sites often rely on product feeds and merchant programs (like OpenAI’s or Perplexity’s) that interact with AI platforms differently than standard web content. If you run an online store, test these signals on your own site before assuming the same limitations apply. The mechanics of how AI tools handle transactional data versus informational content are not yet identical.
The data points to a clear mechanism: schema markup drives visibility in AI Overviews through traditional SEO pathways, not by being ingested directly by AI models. While the structured data improved SERP feature eligibility and organic rankings, the actual citation logic of these platforms relies on content clarity and external signals they can access without parsing JSON-LD. This distinction shifts the focus from technical implementation to strategic content placement. If AI platforms can read your content without schema, what is the real bottleneck for citation selection?
