# Does Schema Markup for AI Actually Work? > Five studies tested schema markup for AI citations. Every one that measured a change found none. Here is the evidence and what is still worth shipping. Canonical URL: https://cite.solutions/blog/does-schema-markup-for-ai-work Source: Cite Solutions (cite.solutions) Published: 2026-09-05 --- [Technical Guides](/category/technical-guides)12 min read # Does Schema Markup for AI Actually Work? [Subia PeerzadaFounder, Cite Solutions · September 5, 2026](https://www.linkedin.com/in/subia-peerzada-75025764/) Key takeaways ## what five schema studies agree on The headlines contradict each other. The research designs do not. Sort the studies by what each method can prove and one distinction survives. 1. 01Both studies that added schema and watched what happened found no citation uplift. Ahrefs measured 1,885 pages against roughly 4,000 controls and saw AI Overviews move -4.6%. 2. 02The peer-reviewed model returned null for schema presence at OR 0.678, p = .296, while Google rank position held at OR 0.762 per position, p < .001. 3. 03Generic markup underperformed no markup at all, 41.6% against 59.8%. Attribute-rich Product and Review markup led at 61.7%. If you are deciding whether to spend a sprint on schema markup for AI search, the honest answer is that the question has now been tested five times and the results only look contradictory until you sort them by method. Two studies added structured data to real pages and watched what happened to citations. Three counted structured data on pages that were already being cited. Those are different questions, and only the first one is the question you are asking. Five studies, sorted by what they can prove Every study that measured a change in citations after schema was added found none The headlines disagree. The designs do not. Sorted weakest to strongest by what the method is capable of establishing, the contradiction disappears and one distinction survives. Relixir · 2025 · 50 sites No control group FAQ schema pages cited 41% vs 15% A snapshot of who already had schema. Nothing was added, so nothing was tested. Otterly · 2026 · \~1M citations No control group 350% FAQ schema citation lift Biggest sample, weakest design. Pages with FAQ schema also carry question-and-answer prose. AirOps · 2026 · 50,553 responses Covariate controls JSON-LD 38.5% vs 32.0% A 6.5-point schema gap, reported alongside a 44-point rank gap on the same pages. Ahrefs · 2026 · 1,885 pages Difference-in-differences No uplift on any platform Schema was actually added and citations were watched before and after. They did not move. Fischman · 2026 · 1,006 pages GEE, clustered errors OR 0.678, p = .296 Null for schema presence. Google rank position was the predictor that held at p < .001. What happened to citations on the 1,885 pages that actually added schema Google AI Overviews \-4.6% Google AI Mode +2.4% ChatGPT +2.2% The distinction that survived a controlled test Fischman's model returned null for schema presence and significant for schema specificity. Generic markup did not merely fail to help. It sat 18 points below pages carrying no markup at all. Citation rate by what the markup contains, 1,006 pages across ChatGPT and Gemini Attribute-rich schema 61.7% Product or Review with populated price, rating, specification fields No schema at all 59.8% Plain HTML, no JSON-LD block Generic schema 41.6% Article, Organization, BreadcrumbList with no factual payload Three numbers that settle the sprint decision 43.1 / 44.8Schema prevalence among AI-cited versus non-cited pages inside Google's own top ten. Indistinguishable, which is what collapsed the apparent effect. 0.762The odds ratio per Google rank position, p < .001\. Rank was the predictor that held when schema did not. 8 of 12Of the most-cited domains in our own 90,132-answer corpus, eight were not brand-owned. No schema deployment reaches those pages. Sources: Ahrefs schema study, 1,885 pages against roughly 4,000 matched controls, August 2025 to March 2026\. AirOps ChatGPT citation study, 16,851 queries run three times for 50,553 responses across 353,799 pages, reported by Search Engine Land. Fischman, Does Schema Markup Predict AI Citation, 730 citations across 1,006 pages and 75 commercial queries on ChatGPT and Gemini. Relixir 50-site sample and Otterly million-citation study as reported by their publishers. First-party figures from The CITE Index, 90,132 AI answers over 63 days, 19 May to 21 July 2026. ## Does schema markup for AI actually work? Schema markup does not increase AI citations on its own. Both controlled studies that added schema to live pages found no uplift, and the peer-reviewed model returned null for schema presence once Google rank was held constant. What did hold is specificity: markup carrying real facts outperformed generic markup by 20 points. That is the finding. The rest of this post is how it was established, why the older studies said the opposite, and which parts of a schema deployment still earn their place. > Schema does not earn citations. Specific schema earns eligibility. ## The five studies, sorted by what each one could prove Read them in this order and the disagreement dissolves. The weakest designs produced the loudest numbers. ### Study #1: Relixir counted 50 sites and found 41% against 15% The most-quoted figure in this debate comes from a [50-site sample](https://signals.sh/blog/faq-schema-for-ai-citation-41-vs-15-percent) comparing pages with FAQ schema to pages without. Sites with FAQ markup were cited at 41%. Sites without were cited at 15%. Nothing was added and nothing was controlled. It is a photograph of who already had schema, and the sites that invest in schema are the sites that invest in everything else. ### Study #2: Otterly reported a 350% FAQ schema lift across a million citations Otterly analysed over a million AI citations and reported that FAQ schema produced a 350% citation increase against unstructured content. We covered that study in detail when it landed, in [what Otterly's million-citation FAQ schema study found](/blog/faq-schema-ai-citations). Biggest sample in the set, same design problem. A page carrying FAQPage markup almost always carries question-and-answer prose underneath it, and the prose is the thing a model lifts. The study cannot separate the two. We published that 350% figure as the headline in April 2026 and we were reading it too generously. The number is real and the design cannot support the causal claim we hung on it. Two controlled studies have landed since, and this post is the revision. ### Study #3: AirOps found a 6.5-point schema gap next to a 44-point rank gap AirOps ran 16,851 queries three times each for 50,553 ChatGPT responses across 353,799 pages. Pages with JSON-LD were cited 38.5% of the time. Pages without were cited 32.0% of the time. On the same dataset, [reported by Search Engine Land](https://searchengineland.com/chatgpt-citations-ranking-precision-length-study-474538), the top search position was cited 58.4% of the time and position ten 14.2% of the time. Both numbers came out of one run. One of them is 6.5 points wide and the other is 44.2 points wide. The coverage quoted the smaller one. ### Study #4: Ahrefs added schema to 1,885 pages and citations did not move This is the first study in the set that changed something. Ahrefs tracked [1,885 URLs that added JSON-LD](https://ahrefs.com/blog/schema-ai-citations/) between August 2025 and March 2026, against roughly 4,000 matched control pages, using difference-in-differences to strip out platform-wide drift. Google AI Overviews moved -4.6%. Google AI Mode moved +2.4%. ChatGPT moved +2.2%. Only the AI Overviews figure was statistically significant, and it points the wrong way. Ahrefs is also candid about its own limits: all schema types were pooled rather than analysed separately, only HTML-delivered JSON-LD was tested, the window was 30 days on either side, and the pages studied were already earning citations. ### Study #5: Fischman's peer-reviewed model returned null once rank was controlled Kurt Fischman collected 730 AI citations from ChatGPT and Gemini across 75 commercial queries in five verticals, plus Google's top ten for the same queries as a control set, for 1,006 unique pages. The paper is [on SSRN](https://papers.ssrn.com/sol3/papers.cfm?abstract%5Fid=6284518). The first pooled model said schema actively hurt citation odds: OR 0.546, p < .001\. That result is the most interesting thing in the paper, because it is wrong and the paper explains why. Google's ranking algorithm enriches its own top ten with schema-bearing pages. That inflates schema prevalence in the non-cited control population and manufactures a negative association out of nothing. Inside Google's results, schema prevalence among cited and non-cited pages was 43.1% against 44.8%, which is no difference at all. Corrected with generalised estimating equations and query-clustered standard errors, schema presence came back null: OR 0.678, p = .296\. Google rank position held at OR 0.762 per position, p < .001. > Every study that measured a change found nothing. Every study that measured presence found rank. ## What the evidence stack actually asks The gap between how teams audit schema and what the citation data responds to is the whole problem. **What a schema audit usually asks:** * •Does the page have JSON-LD? * •Does it validate in the Rich Results Test? * •Are all the recommended types present? * •Is coverage above 80% of the site? **What the citation data asks:** * •Does the markup contain a fact that is not already in the prose? * •Would a model answering this query need that specific field? * •Does the page rank well enough to be in the retrieval set at all? * •Is the claim in the markup visible on the page, or hidden from the reader? Coverage and validity are the two things every schema tool measures. Neither appears on the right-hand list. ## Why generic schema is the part that fails Fischman's paper split markup by what it contains rather than whether it exists, and that split is the only one in the entire evidence base that survived a controlled test. ### Attribute-rich markup was cited at 61.7% and generic markup at 41.6% Attribute-rich means Product or Review markup with populated fields: price, aggregateRating, specifications. Generic means Article, Organization and BreadcrumbList, which describe the page without asserting anything about the world. The difference was significant at p = .012. Pages carrying no markup at all landed at 59.8%, between the two. ### Generic markup is not neutral, it sits below having no markup That ordering is the uncomfortable part and it needs stating plainly rather than explaining away. In this sample, generic schema sat 18 points below pages carrying nothing at all. The most likely reading is selection rather than penalty. Sites that ship a plugin default and stop are sites that stopped at other things too. But if you were expecting generic markup to be a free floor, the data does not support that either. > Generic markup is not neutral. In the one controlled test that split it out, it sat below no markup at all. ### Schema is an eligibility layer, and eligibility is not the same as ranking The useful mental model is a gate rather than a lever. Product and Offer markup is what makes a listing eligible for AI shopping surfaces. Organization markup is what lets an entity resolve to the right company instead of a similarly named one. Neither of those is a citation boost. Both of them are prerequisites for being in the pool at all. That is why the studies keep returning null. They measure whether a lever moved something. Schema is not the lever. ### Find out whether schema is your actual constraint Most sites that are invisible in AI answers have a retrieval problem or a source problem, not a markup problem. Our audit tells you which one you have before you spend the sprint. [Book an AI Visibility Audit](/contact) ## What our own corpus says about where citations came from We ran a 63-day study from 19 May to 21 July 2026 covering 90,132 AI answers across ten consumer categories. The full numbers sit on our [AI search statistics page](/ai-search-statistics). Three of them bear directly on the schema question. Reddit was the single most-cited source in the corpus, with 14,698 citations, appearing in 13.6% of all answers. Eight of the twelve most-cited domains were not brand-owned. And the category leader changed on only 18.7% of day pairs, never changing at all in four of the ten categories. Read those together and the schema sprint looks different. Most of the citations went to pages nobody on your team can edit, and the pages that did win held their position for weeks without anyone touching their markup. > You cannot deploy schema onto a Reddit thread, and that is where most of the citations were. This is also why we treat schema as a hygiene task inside a larger engagement rather than a growth programme. A [managed GEO agency](/geo-agency) earns its fee on the source pool and the passage structure, not on the JSON-LD block. ## How to spend the schema budget without wasting the sprint Schema still has jobs. They are smaller and more specific than the guides suggest, and they are worth doing in this order. ### Step 1: Confirm the page is being fetched before touching the markup Structured data on a page an AI crawler cannot read changes nothing. Check server logs for GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot and Google-Extended before anything else. Our [crawlability audit for AI retrieval](/blog/geo-crawlability-audit-ai-retrieval) covers the specific failure modes. ### Step 2: Fix the entity layer with Organization and Person markup This is the disambiguation job. One Organization block per site with sameAs pointing at the LinkedIn, Crunchbase and Wikipedia entries that already exist. Author pages get Person markup with real credentials. Do it once and stop. ### Step 3: Ship attribute-rich markup only where you hold real facts Product, Offer and Review blocks with populated price, rating and specification fields. If the field would be empty or vague, leave the type out. An empty attribute is what the study called generic, and generic was the losing bucket. ### Step 4: Keep FAQPage markup and stop expecting it to do the work Google retired FAQ rich results on 7 May 2026, and the [structured data documentation](https://developers.google.com/search/docs/appearance/structured-data/faqpage) confirms the schema itself was not deprecated. We wrote the full case in [should you remove FAQPage schema](/blog/should-you-remove-faqpage-schema). Keep it, because the parsing cost is near zero. The question-and-answer prose beneath it is doing the extraction work. ### Step 5: Match markup to the job each page type performs Service, pricing, comparison and expert pages need different types for different reasons, and deploying the same block sitewide is what produces the generic bucket. The mapping is in our [GEO schema deployment matrix](/blog/geo-schema-deployment-matrix-page-types), and the audit that checks it is in [how to run an AEO schema audit](/blog/aeo-schema-audit-entities-answers-proof). ## Ship or skip, by schema type | Schema type | What it actually does for AI | Evidence | Verdict | | ---------------------- | ------------------------------------------------------------------------------ | -------------------------------------------------------------------- | ----------------------------- | | Organization, Person | Resolves your brand to the right entity in a knowledge graph | No citation lift measured; disambiguation is a separate job | Ship once | | Product, Offer, Review | Gates eligibility for shopping and comparison surfaces, carries liftable facts | 61.7% citation rate when fields are populated, against 41.6% generic | Ship where facts exist | | FAQPage | Mirrors the question-and-answer shape a model reaches for | 350% and 41-vs-15 claims are uncontrolled; 45.6% in the AirOps split | Keep, expect little | | Article, BlogPosting | Describes the page without asserting a fact about the world | Sits in the generic bucket that underperformed no markup | Low priority | | BreadcrumbList | Site structure signal for Google, no answer payload | Top of the AirOps split at 46.2%, which rank explains | Low priority | | HowTo | Step sequences a model can enumerate | Not isolated in any study; Ahrefs pooled it with everything else | Ship only for real procedures | The verdict column is about sequencing, not about deleting markup you already have. Removing valid structured data buys nothing. ## How to test this on your own site in two weeks None of the studies above ran on your pages. The design that produced the null results is cheap enough to repeat in-house, and running it once is worth more than another guide. Pick 20 pages that already earn AI citations and 20 matched pages that do not. Record citation counts across ChatGPT, AI Overviews and Perplexity for 14 days with no changes at all, so you have a baseline that includes normal drift. Then add attribute-rich markup to half of each group and leave the other half alone. Keep the prose identical. Most schema experiments fail here, because the team rewrites the copy in the same commit and can no longer attribute anything. Measure for another 14 days and compare the change in the treated pages against the change in the untreated ones. That subtraction is the whole method Ahrefs used, and it is the difference between knowing and believing. If your treated pages move and the controls do not, you have found something the published research missed on your page type. If both move together, you have learned that the sprint belongs somewhere else. We built the measurement side of that loop in [how to measure GEO and AI visibility](/blog/how-to-measure-geo-ai-visibility). ## FAQ ### Does schema markup help AI? Not directly. The two studies that added schema to live pages and measured the change found no citation uplift, and the peer-reviewed model returned null for schema presence at p = .296\. Schema helps with entity disambiguation and surface eligibility, which are prerequisites rather than boosts. ### Is schema markup for AI search different from schema markup for SEO? The markup is identical. The payoff is not. Google rewards several types with rich results, which is a visual benefit inside its own SERP. AI systems read the underlying facts, so types carrying concrete values matter more and types describing page structure matter less. ### Which schema types get cited most by AI? In the AirOps split, BreadcrumbList led at 46.2%, then FAQPage at 45.6% and Organization at 44.3%. That ordering tracks the kind of site that deploys each type rather than the type itself, since the same dataset showed rank position moving citation rates by 44 points. ### Does schema markup help ChatGPT find my site? Discovery is a crawler question, not a markup question. ChatGPT reaches pages through OAI-SearchBot and third-party retrieval, and structured data on a page it cannot fetch does nothing. Check crawler access first, then look at whether you rank for the queries the model fans out to. ### Should I remove FAQPage schema now that Google dropped the rich result? No. Google retired the rich result on 7 May 2026 and left the schema itself in place. The parsing cost is negligible and the markup still describes the page accurately, so removal is work with no return. ## What to do with this on Monday If schema is currently the top item on your AI visibility plan, move it down. The evidence supports one afternoon of entity cleanup, attribute-rich markup on the pages where you hold real numbers, and nothing else. Then spend the sprint you just freed on the two things every study in this set kept finding underneath the schema signal: where you rank for the queries that trigger retrieval, and whether the third-party pages a model already trusts mention you at all. ### Stop guessing which lever moved your citations We measure citation share across ChatGPT, Google AI Overviews, AI Mode, Perplexity and Gemini every week, then work the sources those engines actually pull from. [Talk to Cite Solutions](/contact) Tags [GEO](/tag/geo)[AEO](/tag/aeo)[AI citations](/tag/ai-citations)[AI visibility](/tag/ai-visibility)[ai search optimization](/tag/ai-search-optimization)[structured data](/tag/structured-data)[technical SEO](/tag/technical-seo)[b2b ai visibility](/tag/b2b-ai-visibility) ## Continue the brief [01Technical GEOShould You Remove FAQPage Schema From Your Site?Google killed FAQ rich results on May 7\. The FAQPage schema itself is not deprecated, and it is now a primary AI citation signal across every major LLM.May 12, 2026Read→](/blog/should-you-remove-faqpage-schema)[02Technical GuidesDoes Gated Content Get Cited by AI?Gated content is invisible to AI crawlers, and in software the citation pool runs on company sites. Here is what to ungate and what to keep behind a form.Aug 3, 2026Read→](/blog/does-gated-content-get-cited-by-ai)[03Technical GuidesDeepSeek SEO: How to Get Cited by DeepSeekDeepSeek SEO is how you get cited by DeepSeek, the open-weight engine that answers from memory by default. Here are the five steps that win it.Jul 24, 2026Read→](/blog/deepseek-seo-how-to-get-cited) [FrameworkLearn the CITE framework behind our GEO and AEO workSee how Comprehend, Influence, Track, and Evolve turn AI visibility into an operating system.](/framework)[ServicesExplore our managed GEO services and AEO execution modelAudit, prompt discovery, content execution, and ongoing monitoring tied to AI search outcomes.](/services)[AuditStart with an AI visibility audit before executionUnderstand prompt coverage, recommendation gaps, source mix, and where competitors are winning.](/ai-visibility-audit) On this page On this page ## Work with us on this [AEO ServicesAnswer engine optimization: be the answer AI quotes.Explore→](/aeo-services)[GEO AgencyManaged generative engine optimization for B2B brands.Explore→](/geo-agency)[AI SEO ServicesAI SEO for B2B brands across every answer surface.Explore→](/ai-seo-services) ## Ready to become the answer AI gives? Book a 30-minute discovery call. We'll show you what AI says about your brand today. No pitch. Just data. [Book a Discovery Call](/contact)