Why added schema may not increase AI answer citations

More AI implementation may not fix the visibility problem

A Manhattan team has work on frequently asked question markup, or FAQ schema, queued for 400 product pages while someone proposes replacing its content-management system, or CMS, because it is “not AI-ready.” Traffic has been declining since Google’s AI Overviews appeared, and the technical and search engine optimization, or SEO, teams want work they can act on. Yet nobody agrees on what AI readiness means, so the proposed migration has no clear test of success before engineering begins.

Without a clear goal, teams can easily confuse different outcomes. AI citations are references or links included in generated answers; website visits and revenue measure what happens beyond those answers. For pages already receiving citations, the spending question is whether more technical work will improve their chances of selection once crawlers can access them. Citation research can inform that decision, while traffic recovery remains a separate outcome to measure.

The author of Does Structured Data Actually Improve Answer Engine Citations? hopes to save the Manhattan team about six months of meetings and decisions. That estimate raises a practical question for customer experience, or CX, and marketing technology leaders: which remaining problem would the proposed work solve? Labels such as answer engine optimization, or AEO, and generative engine optimization, or GEO, become useful only when a proposal connects its work to a specific obstacle.

Added schema showed no clear citation benefit for already-cited pages

To identify that obstacle, start with what schema can do. Structured data gives machines explicit labels describing content, and schema markup uses a shared vocabulary for those labels. JSON-LD is a format for delivering that information; FAQ schema applies it to questions and answers. The conventional recommendation has been to add those labels so machines can interpret a page more easily and, in turn, use it more often.

Easier interpretation and more frequent selection are separate steps. A system must first obtain a page, read its parts, and identify what they represent; structured data can help with that work. It then chooses material for a particular answer, which requires a reason to select that page. So a page can be accessible and correctly classified while offering little worth citing.

Ahrefs tested the value of adding schema to pages already receiving citations in a study published in May 2026. Ahrefs sells SEO tools, so demand for measuring and improving search visibility benefits its business. Its researchers tracked 1,885 pages that added JSON-LD between August 2025 and March 2026, matching them with pages that had similar prior citation levels but did not add schema. They measured citations in Google AI Overviews, Google AI Mode, and ChatGPT during the 30 days before and the 30 days after implementation.

The comparison accounts for citation changes that could happen without implementation. Ahrefs used a matched difference-in-differences design: it compared the change among pages that added schema with the change among similar pages that did not. A simple before-and-after report could credit schema for an increase that both groups experienced. This observational design lets researchers distinguish those changes, although it differs from an experiment that randomly assigns pages to receive markup.

Within that comparison, AI Mode citations moved about +2.4% and ChatGPT citations about +2.2%; Ahrefs characterized both changes as statistically insignificant. AI Overview citations declined slightly. The study found no clear incremental citation benefit in its sample. Discussing statistical noise across thousands of URLs, the author treats the small increases as insufficient grounds for the usual implementation claim: a positive percentage needs to be read alongside the comparison group’s change.

Google’s guidance also separates eligibility for generated answers from special markup. Google says AI Overviews and AI Mode require no special schema; as the operator of those search products, it has a commercial stake in how publishers adapt to them. Four days before the Ahrefs study appeared, Google also deprecated FAQ rich results, the conventional search feature that could display marked-up questions and answers. That change concerns a separate feature and provides context for evaluating an FAQ backlog.

But the study’s population limits where its finding applies. Every page already had more than 100 AI Overview citations before the markup was added, so these pages had already been found and used in generated answers. The study therefore addresses schema’s value after discovery. A page that has never appeared in an answer still needs an assessment of access and discovery, where structured data may do useful technical work.

For the Manhattan team, that limit turns a broad schema proposal into a page-by-page diagnosis. The author describes structured data as having held a prominent place in standard AI-visibility advice for two years, which helps explain its familiarity as backlog work. But an accessible, already-cited page needs a more specific case for another implementation pass. The next question is what would make the page or brand more likely to be selected, including whether relevant outside organizations recognize it.

Outside recognition is associated with visibility

Ahrefs’ brand-level research gives teams a reason to investigate recognition. Across 75,000 brands, it examined visibility in AI Overviews, ChatGPT, and AI Mode. Branded web mentions had a correlation of 0.664 with visibility, compared with 0.218 for backlinks, which are links from other websites. YouTube mentions had the highest reported correlation, at 0.737.

Those correlations describe relationships across brands, so a spending decision still needs an explanation of how the proposed work would help. Buying mentions, replacing link-building spending with mention-building spending, and expecting a particular return are separate interventions whose effects the correlations do not establish. For an accessible brand that is rarely selected, the findings make relevant recognition worth investigating. They give a campaign a question to test.

The author makes that distinction explicit by rejecting the interpretation “links are dead, chase mentions instead.” Consider a publication that buyers in a category actually read: an editor decides a company deserves a reference and publishes it. The reference may contain a hyperlink or name the company without one, and answer engines can count either form. Both can come from becoming useful and credible enough for an independent publication to include.

Because the editorial decision can produce either form of reference, changing the measurement need not create a new marketing discipline. A campaign focused on large volumes of unlinked mentions could overlook why a relevant publication chose to include the company. The author describes external mentions and links as having mattered for 20 years, including when the familiar search interface consisted of ten blue links. Recognition work therefore has a history to draw on, while selection in generated answers needs its own assessment.

Recognition may also matter before an answer engine retrieves a page. Seer Interactive proposes a sequence in which a model draws on knowledge acquired during training to choose brands to recommend, then searches for supporting material. Under that hypothesis, citations support a recommendation already formed, and an unfamiliar brand may struggle to enter consideration through a page-level change alone. As a marketing agency, Seer has a commercial interest in companies seeking advice about AI visibility.

Seer’s proposed sequence describes a possible constraint rather than a universal account of answer engines. If a system has little prior basis for recommending a company, additional markup would leave that problem unresolved even when the site works technically. Teams can investigate relevant outside recognition in that situation and assess the resulting changes. A CMS upgrade addresses a different part of the process.

Different answer engines select different material

The recognition hypothesis still leaves a platform question: which system is selecting the material? Moz’s 2026 analysis found that 88% of Google AI Mode citations came from outside the organic top 10, the leading conventional unpaid search results. ChatGPT showed only about 6.5% URL overlap with Google’s top results. Moz sells SEO software and benefits from demand for understanding search visibility; its findings challenge the assumption that conventional Google rankings provide a common citation target across answer engines.

Differences in selected pages also appear in brand-level results. The author cites a March 2026 analysis reporting that citation volume for the same brand differed by as much as 615x between platforms. Visibility on one service is therefore a poor substitute for measuring another. A team can be well represented in one answer environment while remaining largely absent elsewhere, so it needs to identify the platform behind any reported gain or decline.

The systems’ different routes to information help explain why their results can diverge. ChatGPT is described as relying on Bing’s index and substantially on knowledge acquired during training, while Perplexity uses live retrieval and favors recency. Google uses query fan-out, expanding a user’s query into subqueries and drawing material from those searches. These descriptions explain different routes to selection while allowing for the systems to change.

As systems change, teams pursuing platform-specific techniques face ongoing maintenance. The author characterizes that work as maintaining four technical playbooks for four opaque systems that change monthly. The author’s preferred alternative is to earn references from trusted organizations that several systems may encounter. The author describes this as one strategy offering four potential benefits rather than four separate strategies that may produce none; it is a judgment about where to concentrate effort.

But that preference needs to remain separate from the research findings. Ahrefs’ correlations concern the relationship between recognition and visibility, while the platform comparisons concern differences in selected material. Each informs a different part of the decision. A team considering an external placement should identify why it matters to buyers and measure its effects across the relevant platforms, while continuing to improve the evidence on its own pages.

On-page work can strengthen the evidence an engine uses

Improving that evidence gives page-level work a different purpose from adding labels. The original GEO research associated with Princeton and Georgia Tech reported citation-rate gains from adding primary-source references, specific statistics, and quotations from named experts. Each intervention changes the material available to support an answer. The work addresses what a page can substantiate once an engine has obtained it.

The Princeton and Georgia Tech research reported improvements of 40% for inline citations to primary sources, 37% for specific statistics, and 22% for named expert quotations. Each percentage belongs to a separate intervention and should inform a separate editorial decision, because adding them would create an unsupported expectation of combined gains. Teams applying these techniques should measure their own results rather than treating the reported gains as a forecast for their websites.

Applying the techniques starts with the claims a page asks readers to accept. Add an inline reference where a primary source supports a claim, include a concrete statistic where it makes the subject more precise, and obtain an attributable expert quotation where informed judgment helps explain the issue. Each addition gives readers and answer systems more material for assessing the page. The changes may happen inside a CMS, but the underlying work is reporting and sourcing.

Reporting becomes a useful priority when a page is already easy to retrieve but offers weak support for its conclusions. Another markup pass would leave that weakness in place, whereas stronger evidence changes what the engine can use. Editorial investment can therefore sit alongside efforts to earn relevant outside references. Both should follow a diagnosis of the page’s remaining problem.

Bound the technical work before redirecting spending

That diagnosis begins with whether crawlers can obtain the content at all. The CMS should render text on the server or deliver it quickly enough for crawlers to read it. Teams should also check whether robots.txt, the file that gives crawlers access instructions, or a web application firewall, or WAF, blocks AI crawlers. Enterprise security reviews can introduce those blocks unintentionally, making access a concrete engineering issue.

Once access is working, delivery checks should cover clean HTML, meaningful headings, and stable URLs. Placement matters too: cited passages are described as clustering toward the beginnings of documents, so important material should appear early enough to be readily encountered. The author’s example of content buried 900 words down a page illustrates the problem. Engineering and editorial teams can inspect that placement together as part of a review with a defined scope.

The author estimates that a competent developer can establish the technical baseline in a few weeks and separately recommends allocating two weeks to fixing it before redirecting the budget. Both are planning judgments; the defects found determine the actual work. Necessary maintenance continues, and a migration can still make sense when delivery or business requirements justify it. The change in spending is to require a specific purpose for ongoing CMS work whenever visibility remains flat.

With access established, a CX or marketing technology leader can assess investments in reporting and relevant recognition. Options include references in category trade publications, work with credentialed practitioners publishing under their own names, and participation in communities where buyers discuss the category. Video and original data offer further ways to contribute material that others can reference. These activities can complement the evidence added to the company’s own pages.

Relevance should determine how those opportunities are valued. The author judges that one substantive editorial placement in a trusted outlet outweighs a hundred directory listings across platforms. As a strategic judgment, that comparison challenges purchasing based on volume: a proposal should explain why the publication matters to buyers and what makes the reference substantive. A placement earns its rationale through the audience it reaches and the contribution it recognizes.

The Manhattan team can apply the same requirement to the proposed migration and the 400-page FAQ backlog. Before approval, each proposal should identify the affected pages, the access or selection problem being addressed, and the outcome the team will measure. For pages already receiving citations, further schema work needs an incremental case tied to that remaining problem. Replacing the AI-ready label with those requirements gives engineering a defined assignment and the budget owner a decision they can evaluate.

Key executive takeaways

• Define the investment goal: Budget owners need each AI visibility proposal to identify an access or selection problem and a measurable outcome. Track citations, website visits and revenue separately so implementation success has a clear definition.
• Require an incremental case for schema: Ahrefs found no clear citation benefit from adding schema to already-cited pages. SEO owners reviewing markup backlogs need to distinguish those pages from pages still facing discovery or access problems.
• Test relevant outside recognition: Brand mentions correlated with AI visibility in Ahrefs’ research, but the findings establish no guaranteed return from buying mentions. Marketing owners evaluating placements need evidence of audience relevance and a plan to measure results.
• Measure each answer engine separately: Citation patterns differ substantially across Google AI Mode, ChatGPT and other services. Search analytics owners need platform-specific reporting to identify where visibility investments produce gains.
• Strengthen the evidence on each page: Editorial owners can improve answer-ready material through primary-source references, specific statistics and named expert quotations. Research reported citation gains from these interventions, making them candidates for measured tests on accessible pages.
• Bound the technical work: Engineering owners need to check crawler access, content rendering, clean HTML, headings and stable URLs before proposing broader changes. Once delivery works, further CMS spending needs a specific technical or business justification.