What Citation Patterns Tell Us About Optimizing for AI Engines

What changed in AI citation patterns?

AI citation patterns have moved the operating question from how to rank one domain to where an engine finds support for a particular answer. Mencoro analysed 23,466 answers across ChatGPT, Google AI Overviews, Google AI Mode, and Perplexity between 1 July and 30 September 2026. Those answers contained 176,332 citations across 22,193 domains for 42 brands in eight Spanish commercial sectors.

The useful change is not that conventional search has stopped mattering. Google says its existing SEO best practices remain relevant for AI Overviews and AI Mode, and pages still need to be indexed and eligible for a Search snippet. What has changed is the evidence surface around the page. Google says its AI has may use query fan-out, issuing related searches across subtopics and data sources. A brand can therefore be absent from the supporting sources for a question even when its core site is technically sound.

That distinction matters because a citation is a visible unit of selection. Rankings helped people find pages. Citations show the sources an engine chose to support an answer. There is no page two in an AI answer, so a category team needs to know whether it is cited, where it is cited, and which competitors hold the citations that shape the recommendation.

1. Before: optimize the owned domain around keywords, links, and broad authority signals.
2. After: retain sound SEO, then measure cited sources by engine, query, market, and competitor.
3. What to do now: track Citation Share, Citation Count per day, Answer Presence, and Share of Voice against a stable set of commercial questions.

What does the before-and-after evidence show?

The Mencoro dataset gives a dated view of the shift from a website-centred assumption to an answer-centred evidence map. Across the study, company websites received 72.6% of citations, directories and marketplaces received 11.6%, and news and magazines received 4.3%. Yet the tracked brand's own website received only 4.8% of all citations. That does not mean owned content is unimportant. It means the cited web around a brand can matter as much as the brand's own pages for commercial questions.

Treat this as directional evidence from one country, time period, tracked question set, and research methodology. It does not prove that every category has the same proportions. It does establish that a visibility plan based only on a brand domain can miss much of the citation landscape. The right response is measurement, not copying a percentage into a forecast.

The before-and-after table is a practical operating model, not a claim that one tactic replaces another. Search fundamentals make a page eligible. Citation Engineering extends the work to the evidence that an engine sees and selects across the question universe.

1. Start with questions customers ask before they choose a vendor, provider, product, or location.
2. Capture answers separately by engine and preserve the cited URLs, domains, answer wording, date, and prompt.
3. Group cited domains by function: owned properties, official records, trade publications, directories, reviews, marketplaces, social platforms, and video.
4. Use gaps to commission accurate source-ready content or correct incomplete public information. Do not attempt to game model outputs.

Why is there no universal list of sites to target?

There is no universal list because 57.4% of citations in the Mencoro study pointed to domains cited for one tracked brand only. The study describes a fragmented source environment where niche relevance can outweigh a generic placement strategy. A local service, a regulated provider, and a software category may each face a different set of directories, official sources, comparison pages, specialist publications, and community platforms.

That finding affects teams that buy generic authority campaigns in the hope that one placement pattern will carry across every market. Broad authority may still help a business, but it is not proof that a source will appear in a particular answer. The relevant question is narrower: which domains are cited for the specific question, market, and buyer intent that matter to this business?

A useful audit begins with a finite prompt set. Include category questions, comparison questions, location questions where relevant, and questions about standards or eligibility. Run the same set across engines on a documented schedule. Then separate a domain that appears once from a domain that appears across several questions. This reveals whether a citation gap is structural, seasonal, or simply noise.

1. Do not treat a third-party mention as equivalent to a citation on a priority question.
2. Do not assume a source that works in one vertical will work in another.
3. Do record the page type and the factual claim that made a cited page useful.
4. Do prioritize sources that recur in the questions linked to revenue or qualified demand.

How do ChatGPT, Google, and Perplexity differ?

The engines differ enough that a single content format will not cover every citation opportunity. In the Mencoro study, Instagram was cited for 40 of 42 tracked brands, with 97.6% of those Instagram citations coming from Google AI Overviews and AI Mode. The study also found no ChatGPT citations of TikTok. These figures describe this Spanish commercial sample, not a permanent rule for every country or sector.

ChatGPT showed a different profile in the same study. Mencoro reports that it accounted for 84.1% of El País citations and 94.4% of citations to Spain's official gazette, BOE. Public bodies represented 11.7% of its citations. That pattern supports a practical distinction: for questions involving regulation, eligibility, market context, or established company facts, official records and well-supported written documentation deserve particular attention.

Perplexity also showed a distinct signal. It accounted for 191 of 246 YouTube citations in the dataset, or 78%. That is not an instruction to publish video for every topic. It is a reason to check whether useful demonstrations, explainers, or first-party visual evidence are already available for the questions where Perplexity matters. Content should match the evidence a buyer needs, not a channel quota.

Google's documentation reinforces the need for engine-level measurement. AI Overviews and AI Mode may use different models and techniques, so their answers and links can vary. Google also says there are no special technical requirements or special structured data needed for inclusion beyond normal eligibility. Teams should improve content quality and crawlability, then observe citations rather than chasing a fictional AI markup shortcut.

1. For Google AI features, confirm indexing, snippet eligibility, internal links, page experience, images, video, and structured data where it accurately describes the page.
2. For ChatGPT-relevant questions, maintain precise product, policy, and official-source information that can withstand scrutiny.
3. For Perplexity-relevant questions, test whether a concise written explanation, a video, or both answer the searcher's real question.
4. For every engine, log citations at the URL level instead of treating an answer mention as the same thing as a source link.

Who does this change affect most?

This change affects B2B SaaS and tech growth teams that need to know whether AI recommends them or a named competitor for category and comparison prompts. It also affects local, multi-location, and service businesses that depend on being present when a person asks for the best service in a city. Both groups can be invisible if they measure only conventional rank positions and organic traffic.

For a SaaS team, the highest-priority questions often ask for the best tool, an alternative, a fit for a workflow, or a comparison between named vendors. The cited pages may include vendor documentation, independent reviews, trade coverage, marketplaces, and official material. A local business can face a different source map that includes map-oriented results, directories, review surfaces, local publishers, and social profiles.

The operating risk is false confidence. A team may see healthy search traffic and assume it is represented in AI answers, or see one isolated citation and assume it has durable presence. Neither conclusion follows. Answer Presence shows breadth across the question universe. Citation Count per day shows volume. Citation Share shows the proportion of relevant answers that cite a brand. Share of Voice places that result against competitors.

1. Growth leaders need a question set tied to category demand and commercial intent.
2. Content teams need a source map that distinguishes owned pages from the external evidence surface.
3. Local operators need location-specific monitoring rather than a national visibility assumption.
4. Leadership needs a recurring report that separates citation movement from traffic and pipeline outcomes.

How should you respond without chasing shortcuts?

Respond by building a citation measurement loop before expanding production. Define the categories, competitors, markets, and questions that matter. Capture answers across ChatGPT, Perplexity, Gemini, Copilot, and Google AI Overviews where they are relevant to the business. Then review cited URLs for accuracy, gaps, and recurring source types. This turns AI visibility from anecdote into a repeatable editorial and distribution problem.

Next, make the owned site easy to cite. Publish direct answers to real questions, distinguish product claims from evidence, keep material current, and link related pages so a reader and a crawler can understand the topic. Google explicitly says helpful, reliable, people-first content and ordinary Search requirements apply to its AI features. There is no special AI file or special schema type to add for AI Overviews or AI Mode.

Then build coverage beyond the domain through truthful, useful information in the places that your category actually uses. That may mean completing a business profile, correcting directory records, publishing a technical explainer, supplying accurate marketplace information, or developing a video that demonstrates a complex task. The method is quality, coverage, and freshness. It is not an attempt to manipulate models.

Finally, make the review cadence strict. A citation snapshot is only a snapshot. Track the same questions over time, flag changes in cited domains and competitors, and connect movement to content releases or verified source updates. Do not promise a citation count. Use the data to decide what deserves the next editorial investment.

1. Measure the existing answer set before selecting topics.
2. Prioritize pages and external sources that solve a specific evidence gap.
3. Publish accurate material that can be checked by a reader.
4. Re-run the same question set and report the movement with source URLs.

What should an AI citation report include?

An AI citation report should include the questions tested, engines tested, date, country or market setting, answer presence, citation counts, Citation Share, cited URLs, and the competitor cited in each answer. Without that evidence, a citation claim is difficult to audit and a month-over-month change is easy to misread.

The report should also separate action from observation. A cited directory with incorrect company details needs an information correction. A priority comparison question with no credible owned explanation may need a new or updated page. A competitor's recurring citation is a research lead, not proof that copying its format will work. The underlying job is to improve the evidence available to the engine and the buyer.

This is why Citation Engineering is a publishing discipline rather than a trick. It treats AI answers as a visible selection layer, then earns inclusion through authoritative content, accurate coverage, and ongoing measurement.

1. Use a stable prompt library and document any changes to it.
2. Store answer text and cited URLs together so a reviewer can recertify the finding.
3. Label findings by engine instead of combining incompatible source patterns.
4. Assign an owner to each verified evidence gap and record the outcome after the next measurement cycle.

Sources

Source: Mencoro

Mencoro analysed 23,466 AI answers, 176,332 citations, and 22,193 cited domains across 42 brands in eight sectors in Spain during Q3 2026. Mencoro, 2026-09-30

Source: Google Search Central

Google says existing SEO best practices remain relevant for AI Overviews and AI Mode, with no additional technical requirements or special structured data required. Google Search Central, 2025-12-10

Source: The AI Journal

The AI Journal reported the Mencoro findings, including differences in cited source profiles across ChatGPT, Google AI features, and Perplexity. The AI Journal, 2026-10-05