Does Schema Markup Help AI Citations? Ahrefs Tested 1,885 Pages

Adding schema is not a demonstrated citation-growth tactic for pages already being cited. Ahrefs' May 11, 2026 study makes that a concrete budget decision: maintain useful structured data, but do not buy a promised increase in AI citations just because someone will add JSON-LD.

FAQ blocks and generic GEO checklists are in the same position: the best tests we can find do not show that they get a brand cited. What the evidence supports is narrower. In ChatGPT's API, a page was cited only after the engine's own search had retrieved it (Gemini and Perplexity were not tested the same way). Across the engines, within a website, pages whose title and a paragraph matched the questions people ask were cited more. And for a brand the models barely know, credible outside pages that name it are the other route in. The sections below separate those effects and say how far each can be trusted.

The distinction matters. A study that finds no clear uplift in an established cohort is useful evidence against an easy sales claim. It is not proof that every schema type is useless, that all AI systems ignore structured data, or that removing valid markup will improve performance.

The Finding: Citation Growth Did Not Follow the Schema Change

Ahrefs tracked 1,885 pages that added JSON-LD between August 2025 and March 2026, matched them against 4,000 control pages, and compared citation changes across Google AI Overviews, Google AI Mode and ChatGPT.

Its preferred analysis was a matched difference-in-differences comparison: examine how the treated pages changed relative to comparable pages over the same period, rather than credit schema for a platform-wide rise or fall.

Platform	Reported relative citation change versus matched controls	The useful reading
Google AI Overviews	−4.6%	A small, statistically significant decline; Ahrefs says the cause is unexplained
Google AI Mode	+2.4%	Not statistically distinguishable from zero
ChatGPT	+2.2%	Not statistically distinguishable from zero

These are relative changes in citation counts in this study, not percentage-point changes in a brand's share of AI answers. They are not conversion rates or forecasts for a new implementation.

The AI Overviews result is not a reason to rip schema out. Both groups were declining; the treated group declined somewhat faster. Ahrefs does not identify why. On AI Mode and ChatGPT, the small positive estimates do not establish a dependable benefit either.

My read: the schema change is not the growth story here. A proposal that sells implementation as an established citation multiplier needs better evidence than the presence of markup on frequently cited sites.

The Cohort Is the Important Qualification

Every page in the dataset already had at least 100 AI Overview citations in February 2025, before the schema additions Ahrefs tracked. These were not invisible new pages waiting to enter a search system.

That changes the question the study can answer. It examines whether adding schema was followed by additional citations for pages already receiving them. It does not directly settle whether structured data helps a particular new page become discoverable, whether a specific schema type helps a particular feature, or whether a different kind of site would respond differently.

Ahrefs also pooled schema types. An average across Article, Product, Organization and other implementations is not a separate result for each one. Pages can change their content, links or technical setup when they add markup, and the researchers acknowledge that those co-occurring changes cannot be fully separated.

This was a matched observational analysis, not random assignment of schema to otherwise identical pages. Keep its finding strong and its scope intact: adding JSON-LD did not deliver a clear citation uplift in this already-cited cohort.

A Second Test: Compare Pages of the Same Website

Chen and Koudas's October 10, 2026 arXiv preprint asks the same question from another side. The preprint analysed 158,099 answers and 468,037 citations collected over 90 days from ChatGPT, Gemini and Perplexity in two markets, US home internet and small business-software vendors. Across websites an FAQ block went with 18.1% more citations, but comparing pages of the same website it went with 3.4% fewer, and none of the FAQ block, structured data, summary box, author box or page speed effects was significant within a website.

The reason is that the practices cluster on websites that are cited more for other reasons. FAQ blocks were more common on the most-cited quarter of websites than on the least-cited quarter, so a comparison across websites credits the block with the website's reputation. Compare pages inside one website and that credit disappears.

What did survive the within-website comparison was fit to the question: pages whose title and paragraphs match the questions people ask were cited more. That is a statement about the words on the page, not about an FAQ block as a format. Retrieval explained far more than page features did. The same preprint replayed the questions through ChatGPT's API and found that every page ChatGPT cited was one its own search had returned (100.0% of 22,315 cited URLs). Which pages that search retrieved explained 32% of the within-website variation in 90-day citation counts, against 2.8% for all measured page features. In the small-vendor market the figures were 30% and 10.0%, and the difference was not statistically significant.

How far to trust it

• It is a preprint, not a peer-reviewed result, and it is observational. The within-website design removes everything constant inside a site, not every difference between a site's pages. The results are associations.
• It measures engine APIs, not the apps people use. The answers came from ChatGPT, Gemini and Perplexity API calls with web search on, and Perplexity could not be replayed. Both datasets are English-language US questions, written for organizations that use the authors' monitoring platform, and about a third of the cited URLs could not be analysed.
• There is a commercial interest. The preprint thanks the team at ktau.ai for help with the data, ktau publishes the study as its own ("we", "our monitoring platform"), and its write-up ends in an offer to replay your category. Read it as you would Ahrefs' own study: informative, not independent.
• We have not checked its data. The preprint says data and code are released, but the text we read gives no link, and ktau's write-up gives none either, so only its reported results were read.
• It is not independent confirmation of the earned-media finding. An earlier paper from overlapping authors reported that AI search leans toward earned, third-party sources over brand-owned content.

Counterevidence sits on the other side. GEO-16, an observational audit of cited B2B SaaS pages, found structured data among the page-quality pillars most associated with citation. Its abstract clusters errors by domain but does not describe a within-website comparison, and the preprint's general argument is that comparing across websites credits a practice with the website's own effect. Volpini and colleagues found in a controlled retrieval-augmented generation experiment that JSON-LD alone gave modest gains and a richer entity page format, which they built and release templates for, gave much larger ones. That tests retrieval accuracy in a lab system, not citations in public engines. Neither result reverses the budget call below. Both are reasons to keep markup accurate rather than to remove it.

Retrieved, Cited and Named Are Three Different Outcomes

Most advice treats them as one. The preprint's data separates them.

Outcome	What it means	What the preprint found	What you can act on
Retrieved	The engine's own search returns your page	Every page ChatGPT cited was one its searches returned; the engines rewrite each question into new sub-queries on every answer, and the pages they return are rarely the pages Google or Bing return	Make sure the page can be crawled and read, and word its title and a paragraph in the buyer's terms
Cited	The answer links your page	ChatGPT cited mostly from the top of its own ranking; at equal rank, FAQ blocks and schema went with slightly lower odds of being cited, not higher	Do not buy markup as the citation lever
Named	The answer mentions your brand	A model can name a brand it already knows with no cited page behind the mention	Be described accurately on pages the engines can retrieve

The preprint found that 46% of ChatGPT's mentions of a brand the question did not name were supported by no cited page that names that brand. In the preprint's small-vendor market, a cited third-party page that names a vendor went with 39.7 times the odds that ChatGPT's answer names that vendor, and 17.7 times for Gemini, against 4.2 and 6.3 times in the large-brand market. That is an association in one niche market, and the third-party pages in it are any outside websites, including reviews, directories and forums, not only earned editorial coverage. It supports making sure credible outside sources describe a little-known brand accurately. It does not show that a placement causes a citation.

The preprint also found that a brand name can work against a product. In its controlled tests on the models behind the engines, with constructed inputs, revealing brand names reordered rankings of products with identical facts. What the model already believes about a brand enters the answer, which is another reason to check what engines say about you and not only whether they link you.

Two Smaller Tests Check the Same Question From Other Angles

Two other tests are worth reading, as differently built checks with narrower scope, not as a second confirmation of Ahrefs.

OtterlyAI ran a three-month experiment on its own site, from December 7, 2025 to March 7, 2026, and says plainly that it is one SaaS brand, not an industry study. Otterly reported that six of the seven platforms it asked to return the markup from two of its pages failed to do so, and that Gemini was the one that returned it correctly, though Gemini showed no visibility gain on the tested pages. Otterly's brand coverage rose after the markup went live, but competitors that, per Otterly, had made no schema changes moved as much or more, and Otterly concluded the lift came from an algorithm change, not from schema.

Kurt Fischman's preprint collected 730 citations from ChatGPT and Gemini across 75 commercial queries. After correcting an artifact in its first pooled comparison, it found no significant link between schema presence and citation (odds ratio 0.678, p = 0.296); Google organic rank position was the dominant predictor. Its one exception: Product or Review schema with populated fields such as pricing and ratings was cited more often than generic Article, Organization or BreadcrumbList markup (p = 0.012). That is a single-author preprint, so read the exception as a hypothesis worth testing on a product page, not as a rule.

A Same-Website Count From the Machine Relations Index

Machine Relations Research, the Bramwell-owned publisher of the Machine Relations Index, ran the same-website comparison on its own answer ledger. It paired 687 pages that ChatGPT, Claude, Gemini, Google AI Mode or Perplexity cited with a page from the same website that the ledger never recorded as cited. More than eight in ten of both carried JSON-LD, and the gap between the two groups was about two points. FAQPage markup was the one type with a clearer gap, and the study could not separate it from the visible questions it describes. Roughly one cited page in nine carried no markup of either kind.

That fits the preprint's reading: the markup gap between cited and uncited pages is a gap between websites, not between pages of one website. It is a count, not a test of what adding markup does, and it read the HTML as served, so markup added by JavaScript after load is not in it.

Direct Retrieval Is a Different Question From Search Eligibility

The related searchVIU experiment, published December 2, 2025, tested five AI systems in October 2025 using a purpose-built product page. Its eight scenarios put information in visible HTML, JavaScript-rendered content and different structured-data formats.

The important result concerns what the systems returned: none of the five systems tested returned the information available only in JSON-LD or hidden markup. That is a reason to put material claims, product details and answers on the page people can read.

It is not a complete account of how an engine discovers, indexes or ranks a source. SearchVIU explicitly distinguishes its tests, which mainly show direct fetch (Perplexity and Google AI Mode answered from their indexes), from training, indexing and search phases. Its result should not become the universal claim that an AI system can never use information from structured data.

Google's own guidance makes the operating distinction clearer. To be eligible as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to appear in Search with a snippet. Google says no special schema.org markup is required for these AI features.

That is an eligibility statement, not a promise that an eligible page will be selected. Google also says structured data should match the visible page. Its general structured-data guidelines retain the role of supported markup in rich-result eligibility while making clear that correct implementation does not guarantee a rich result.

The Move: Keep the Plumbing, Fund What Gets You Retrieved and Named

I would maintain useful, accurate markup. I would not make a schema rollout, an FAQ block or a generic checklist the centerpiece of an AI-citation growth plan on the strength of this evidence. The implementation reference covers which types to implement and how, and the how-to guide covers the rollout.

Fund these instead, in this order:

1. Make the page readable where engines look. For Google's AI features a page must be indexed and eligible to appear in Search with a snippet, so fix that first for them. The preprint found that ChatGPT's retrieval is its own and rarely returns the pages Google or Bing return, so indexing in Google is not a proxy for being retrieved there. If a page's decisive information exists only in hidden markup, put it in the visible page.
2. Answer the buyer's question in the buyer's words. Put the question's terms in the title and in a paragraph that answers it directly; the paragraph can sit anywhere on the page. The title match and the best-matching paragraph were the two page properties that passed all five of the preprint's robustness tests within a website; the opening paragraph alone did not, and the title match did not show for ChatGPT. Questions and answers on the page are fine; the evidence is about how well the text matches the question, not about the FAQ block format.
3. Measure retrieval, citation and naming separately, per question, over time. Run the questions your buyers ask, with search on, in each engine, and record whether your page is cited, whether your brand is named, and what the answer says. The cited set changed week to week in the preprint's data, so one snapshot is not a standing.
4. Get credible outside sources to describe you accurately. For a brand the models barely know, this is the closest evidence of how an answer comes to name it. Earned coverage in relevant publications, analyst and review mentions, and accurate profiles all count as sources an engine may retrieve. Judge them by whether your answers change, not by placement counts.

Machine Relations is the discipline of earning AI citations and recommendations for a brand by making that brand legible, retrievable and credible inside AI-driven discovery. Steps two to four are that work. Jaxon Parrott coined Machine Relations in 2024; AuthorityTech, which he founded in 2018, is The World's First Machine Relations Agency; Machine Relations is open to anyone and machinerelations.ai is its home. The strategic job is to earn relevant, credible coverage around a position the brand can own, and to check that engines retrieve it. It is not to claim that one placement automatically produces a citation or that coverage is a prerequisite for every possible source.

For the broader execution decision, the GEO strategy guide puts source-level citation work inside that larger authority strategy.

The budget call is straightforward: keep schema doing its legitimate job. Fund citation growth against a documented access, answer or authority gap, and measure retrieval, citation and naming separately instead of selling the markup as the result.

Frequently Asked Questions

Does FAQ schema help AI citations?

No test we found shows it does. Ahrefs found no clear citation uplift when already-cited pages added JSON-LD (it did not test FAQ blocks separately), and the 2026 preprint found no significant FAQ-block effect once pages of the same website are compared. Valid FAQ markup that mirrors visible questions is fine to keep if it serves readers. No test we found shows that it earns or costs citations, and Google no longer shows the FAQ rich result in Search, so it has no rich-result benefit there either.

Do I need schema markup to be cited by ChatGPT or Perplexity?

Nothing we found says you do. Google says no special schema.org markup is required to appear in its AI features, and the preprint found that every page ChatGPT cited was one its own search had retrieved, so being retrieved comes first. Keep markup accurate and matched to the visible page.

Do GEO and AEO checklists get a brand cited?

Not as checklists. The items that went with citation in the evidence are being retrieved by the engine's own search and having a title and a paragraph that match the buyer's question. Items such as FAQ blocks, summary boxes, author boxes and page speed did not show a significant effect within a website. Use a checklist to find access and answer gaps, not as a promise of citations.

What should a brand invest in instead of schema and FAQ blocks?

Make important pages crawlable and readable, answer the buyer's question in the buyer's words, measure whether your page is retrieved, cited and your brand named for each question, and get credible outside sources to describe you accurately. For little-known brands, outside pages that name the brand are the closest evidence of how an answer comes to name it.

Does this apply to the ChatGPT or Perplexity apps people use?

Not directly. The preprint replayed questions through engine APIs with web search enabled, which need not behave like the consumer apps. Treat it as evidence about how retrieval-based answers are built, and check your own questions in the products your buyers use.

Updated November 3, 2026: added the same-website preprint test, the retrieved, cited and named distinction, what to fund instead, and this FAQ. The Ahrefs finding above is unchanged.

Updated November 9, 2026: added the Machine Relations Index same-website count.