Why AI Citation Overlap Is So Low Across Search Engines

AI citation overlap is the amount of shared source material that different AI answer engines cite for similar questions. Research finds that overlap is usually low: most domains cited by one engine are not cited by another, even when the engines agree on which brands to recommend. That means one-platform optimisation can leave real gaps in how buyers discover and verify your brand.

The practical response is not to create a different website for every model. Build clear, useful answers on your own site, earn credible third-party coverage, and measure citations separately by engine. SearchFIT’s AI search resources can help you explore the topic, while its AI visibility and content articles offer related practical guidance.

What AI citation overlap means

An AI citation is a source an answer engine links to or names as evidence for a response. It might be a company’s product page, a review site, a news article, a forum discussion or a video. Citation overlap measures how often two or more engines cite the same sources across a set of questions.

For example, imagine ChatGPT, Perplexity and Gemini answer the question, “What is the best project management software for a small design agency?” One might cite a vendor’s pricing page, another a Reddit discussion, and a third a software review site. If all three cite the same domains, overlap is high. If each draws on a different set, overlap is low.

Overlap can be measured at different levels. A domain is the main website, such as example.com. A root domain groups pages under that site, including subdomains in some studies. An exact URL is one specific page, such as example.com/pricing. Two engines might cite different pages on the same domain, so domain overlap can look higher than exact-URL overlap.

A common measure is the Jaccard overlap. It compares the number of sources shared by two engines with the total number of unique sources cited by either engine. If Engine A cites ten domains and Engine B cites ten, with one domain in common, the shared portion is one out of nineteen unique domains, or about 5%. The formula is shared sources divided by sources in either set.

That is useful for comparing citation sets, but it is not a direct measure of how often a brand appears in answers. A brand mention, a linked citation and a recommendation are different things. Keep them as separate reporting measures.

How much citation overlap do studies find?

The short answer is: little, although the exact result varies with the question set, engines, time period and definition of a source. Several studies report low overlap between engines, but their samples are not interchangeable. As examined in broader research on cross-engine citation overlap, domain sharing remains rare across engines. Treat each figure as evidence from that study’s design, not as a permanent universal rate.

Foglift’s citation overlap study reports a mean pairwise Jaccard overlap of 0.094, or 9.4%. The figures reported by Kingy are similarly low for OpenAI and Claude, OpenAI and Gemini, and Claude and Gemini. Those numbers describe the evaluated prompts and API answers, not every possible user query.

The AEO Spotlight study provides a helpful view of how the shared and unshared groups relate. Across its 30 buying queries, it found 81 domains cited by all three tested engines, representing 5.7% of the domains in its analysis. Those domains accounted for 46.8% of aggregate citations. At the other end, 1,132 domains, or 79.1%, were cited by exactly one engine. The study’s breakdown of shared and engine-specific sources shows why a small shared group can still receive a large share of total citations.

Wellows reports that 79.6% of websites in its much larger study appeared in only one of five engines. When it matched exact URLs instead of root domains, single-engine isolation rose to between 85.1% and 93.2%. That distinction matters for marketers: being cited somewhere on your domain does not mean the same useful page appears across engines. Wellows explains its citation overlap research.

These results should not be reduced to one headline number. A study of a few dozen commercial questions can reveal patterns in those categories, while a large collection of questions may include many other intents. A study comparing three engines cannot establish the exact overlap of a different group of five. Prompt wording, geography, timing, the answer engine’s available search features and how researchers deduplicate domains can all affect the outcome.

The robust takeaway is not that overlap is exactly 5%, 9% or 11%. It is that a large share of cited domains can be engine-specific, so a single-engine report cannot describe your full AI search presence.

Why engines can agree on brands but cite different sources

A brand recommendation and its supporting evidence are separate parts of an answer. An engine may have strong reasons to name a familiar vendor but retrieve a different page or source to support that recommendation. The answer can look consistent to a buyer even while the citation trail changes from one platform to another.

Foglift reports high agreement on which brands engines recommend in its analysis, while finding much lower agreement on the domains cited to support those recommendations. This is not contradictory. A brand can be associated with a category across many sources, but each engine may select different evidence for a particular answer.

Think of a buyer asking, “Which accounting platform should a growing online retailer consider?” One engine may cite a vendor’s pricing page. Another may cite a review platform. A third may use a comparison article. The brand can appear in all three responses, but the sources shaping a buyer’s confidence are different.

This distinction matters when diagnosing visibility. If your brand is named but another site is cited, you have brand visibility without source ownership. If your site is cited but your brand is not recommended, your content may be useful as evidence without connecting strongly to the product choice. Track both outcomes rather than treating “mentioned” and “cited” as synonyms.

Why AI engines cite different websites

AI answer engines do not all search the web in the same way. Each product combines its own model, search or retrieval systems, source selection and answer-generation process. A retrieval pipeline is the chain of steps used to find possible sources and select material for an answer. Its design influences which pages are available to cite.

The details can change, and outside observers do not have complete access to each platform’s systems. It is safer to treat claims about a platform’s preferred sources as observed patterns in a particular study, rather than a fixed rule or official guarantee.

Different search indexes and discovery paths

An index is a database of pages and other content that a search system has found and organised. Search partnerships, crawlers, refresh schedules and ranking systems affect what a platform can retrieve. If two engines start with different candidate pages, they may end up citing different sources even when they answer the same question.

The AEO Spotlight analysis, for example, found substantial differences among the sources used by ChatGPT, Perplexity and Gemini for its buying prompts. Other studies have reported platform-specific source patterns, such as more community discussion in some Perplexity results and more video sources in some Google AI Overview datasets. These are observations, not instructions to assume every answer will follow the same pattern.

The practical point is that a page’s presence in Google does not guarantee that every assistant will retrieve or cite it. Nor does an answer engine’s use of web search mean that it uses the same ranking order as a conventional search results page.

The question changes which evidence is useful

A prompt is the question or instruction given to an AI system. Small changes in the prompt can change the answer’s intent and the evidence needed. “What does this software cost?” calls for current pricing. “Is it suitable for a finance team?” may call for security details, integrations, customer experiences or independent assessments.

This is why one generic visibility check is not enough. Build a realistic prompt set around how your buyers research, compare, shortlist and validate products. Include category questions, use cases, alternatives, pricing, integration requirements and concerns. A B2B SaaS company might test “best inventory software for a multi-site retailer” and “does [product] integrate with [platform]?” An online store might test product comparisons, materials, sizing and delivery questions.

Engines may favour different kinds of evidence

Some answers need first-party facts, such as product specifications or a vendor’s current pricing. Others need independent validation, such as user reviews, industry coverage or practical discussion. An engine may draw on both, but choose different supporting pages depending on its retrieval process and how the question is framed.

A company’s own site is usually the best place to publish accurate, detailed product information. It is not automatically the only source a buyer or AI system will trust. A third-party review, an editorial comparison or a useful community discussion can supply evidence that the company’s own marketing page cannot.

This is particularly relevant for commercial questions. If an assistant is asked to compare vendors, it may use external sources to help establish differences and credibility. That does not make third-party citations a substitute for accurate product information on your site. It means both first-party clarity and independent validation matter.

The same engine can change its sources

Citation overlap is not only a difference between platforms. An engine can return different sources for the same prompt at different times. This is sometimes called run-to-run drift: variation between repeated answers, even when the question stays the same.

The Kingy AI Citation Index reported that, in its repeated API tests, OpenAI repeated 66.4% of its sources, Claude 51.5% and Gemini 31.9%. CueScout, in separate B2B buying-query observations, reported different day-to-day stability rates for Perplexity and ChatGPT. These studies use different methods, so their percentages should not be compared as though they were one test. Together, they illustrate why a single check is a weak basis for a strategic decision.

A source can disappear because the prompt was interpreted differently, the search results changed, the model or retrieval system was updated, or the answer drew on a different set of evidence. Track results on a regular schedule and look for patterns across repeated observations. Do not treat one missed citation as proof that a page has failed.

Why ranking first on Google may not earn an AI citation

A top organic ranking and an AI citation are related visibility outcomes, but they are not the same outcome. Organic ranking is a page’s position in a conventional search results list. An AI citation is a source selected for an AI-generated response. A page can do well in one system and not be selected by another.

Studies of Google AI Overviews have found that the amount of overlap with Google’s top organic results varies. Ahrefs has reported a 38% top-10 overlap in its analysis, while BrightEdge findings place overlap between 17% and 32%. These figures reflect different study designs and should be read as evidence that the relationship is not a guarantee, rather than as one settled rate.

Ranking still matters. Search visibility can help people find your content and can support discoverability. But an AI answer may select a page based on how well it supplies a specific answer, what information is available from other sources and how the retrieval system evaluates candidate pages. Traditional SEO can help a page be found; it does not ensure that the page is chosen as a citation.

If your page ranks well but is rarely cited, check whether it answers the actual prompt directly. A broad landing page may rank for a category term but contain no clear answer to a specific question about pricing, compatibility or limitations. Add accurate, easy-to-find information where it genuinely helps the reader, and make sure the page can be crawled and understood.

A practical cross-engine strategy

The goal is not to force every engine to cite the same page. The goal is to make your brand, product facts and evidence easy to discover across different routes, then learn which sources appear for the questions your buyers care about.

1. Define the questions that matter

Start with a list of buyer questions, not a list of platforms. Group questions by intent:

• Discovery: What tools or brands are relevant to this need?
• Evaluation: How do the options compare, and which is suited to a particular team?
• Validation: Is the product reliable, well reviewed or credible?
• Decision: What does it cost, what does it include and how does someone get started?
• Implementation: Does it work with our current systems, process or location?

Write prompts the way a buyer might ask them, including natural variations. Record the exact wording, engine, date, geography where relevant, answer, brand mentions and citations. This makes later comparisons more meaningful.

Do not build a huge prompt list before deciding what you need to learn. A focused set of commercially important questions is easier to review and act on. Expand it when you see a new use case, a competitor or a recurring source appearing in answers.

2. Measure each engine separately

Report brand mentions and citations by platform. A blended “AI visibility” score can hide the fact that a brand is well represented in one engine and missing in another. Break out at least:

• Whether the brand was mentioned.
• Whether the brand was recommended or compared.
• Whether a page on your domain was cited.
• Which exact URL was cited.
• Which external domains appeared alongside it.
• Whether the answer was accurate and current.

Use a consistent prompt set and observation schedule. A weekly or fortnightly check can help reveal direction without overreacting to a single answer. Keep a record of major product, content and platform changes so that you can interpret shifts. Avoid treating a small sample as a precise market-wide estimate.

If you work with clients or several brands, keep the same clear definitions across reports. SearchFIT’s pricing page can help you assess a platform for a regular tracking workflow. For an agency comparing approaches, see SearchFIT versus agencies and consider what reporting, content operations and client needs require.

3. Make your own pages easy to cite

Write pages that answer real questions clearly and accurately. Place the direct answer near the relevant heading, then explain the context, qualifications and next steps. This makes information easier for a person to scan and gives an answer system a self-contained passage to interpret.

A self-contained passage is a short section that makes sense without requiring the reader to piece together facts from several distant parts of a page. For example, a pricing page should state what a plan costs, what the price includes and any important qualification together, rather than scattering those details across unrelated sections.

Useful formats include:

• Pricing pages with clear plan names, inclusions and conditions.
• Product pages that explain who a product is for and what it does.
• Comparison pages that name the criteria used and distinguish fact from opinion.
• Technical documentation with current compatibility and setup details.
• FAQs that answer genuine questions, rather than repeating keywords.
• Case studies that explain the situation, approach and evidence without overstating results.

Use headings that describe the question a section answers. Keep paragraphs focused. Define specialist terms on first use. Include dates where information changes over time, and identify the source of factual claims. This is good content practice for people as well as machines.

Structured data, sometimes called schema markup, is code that labels information on a page, such as a product, organisation or article. It can help search systems interpret page details, but it does not guarantee a citation or special placement. Use valid structured data when it accurately reflects visible page content, and do not treat it as a replacement for useful information.

4. Build a credible presence beyond your own site

Your website controls what you say about your product. It does not control all the sources that buyers and answer systems may consult. Look for places where your customers genuinely seek advice or leave feedback, and maintain accurate, useful profiles there.

For a B2B software brand, that might mean keeping software directory profiles current, responding professionally to review feedback and making product documentation easy to verify. For an e-commerce brand, it might mean clear product information, reliable retailer details and useful video demonstrations. For an agency, it may include credible case studies and a consistent description of services across appropriate profiles.

Community participation is not a shortcut to citations. Do not flood forums with promotional posts or disguise advertising as independent advice. Contribute where your team has a real answer, disclose relevant affiliations and respect each community’s rules. Useful participation can support trust with people, whether or not it produces a citation in an AI response.

Earned media, reviews and community discussion can provide independent evidence, but each has limits. A review site may be useful for customer feedback but wrong about a product detail. A company’s documentation may be authoritative about an integration but not independent evidence of customer satisfaction. Help buyers assess the source, and correct material errors through the source’s proper process.

5. Match the content to the business model

For B2B SaaS, make product capabilities, integrations, security information, pricing and implementation details easy to verify. Explain what the product does and where it may not fit. When comparison content is appropriate, use transparent criteria rather than unsupported claims about competitors.

For e-commerce, keep product attributes consistent across product pages and relevant feeds. Explain materials, dimensions, compatibility, care, delivery and returns in plain language where applicable. Accurate information reduces ambiguity for both shoppers and search systems. SearchFIT’s guide to AI search for e-commerce brands is a useful next resource.

For agencies, separate client reporting by brand, market, prompt group and engine. An agency should be able to explain whether a citation gap comes from missing content, weak third-party evidence, a technical access issue or simple variation in a small sample. Use the findings to choose a specific next action rather than promising a particular citation outcome.

6. Diagnose gaps before changing everything

When one engine cites your site and another does not, compare the answer and sources before rewriting content. Ask:

1. Did both engines receive the same question and intent?
2. Is the page accessible and clear about the fact the answer needs?
3. Does another source provide fresher or more independent evidence?
4. Is the page a useful destination for the query, or is a more specific page needed?
5. Is the pattern repeated across observations, or did it happen once?

If a review site is repeatedly cited for a product comparison, do not automatically try to displace it. Check whether the profile is accurate and whether your own comparison page covers the criteria buyers need. If a platform consistently cites documentation for integration questions, make sure the documentation is current, public and easy to navigate.

A low citation count can also reflect the prompt set. If the tracked questions do not match your buyers’ language or purchase stage, the measurement may be unhelpful. Update the prompts before making large content changes.

Common mistakes to avoid

Optimising for one engine only. A strong result in ChatGPT does not tell you how Perplexity, Gemini, Claude or Google AI Overviews will behave. Use a representative set of engines that matter to your audience, and report each separately.

Treating every citation as equally valuable. A citation on a relevant, trusted source may help a buyer more than a link on an unrelated page. Review the source, context and accuracy, not just the count.

Confusing brand mentions with citations. A model can mention a company without linking to it, or cite a page without recommending the brand. Keep these measures distinct in dashboards and client updates.

Assuming one universal content template will work. Clear answers and reliable facts are broadly useful, but buyers ask different questions and engines retrieve different evidence. Start with user needs, then test whether your content is visible for those needs.

Making major decisions from one observation. Engines change, prompts vary and answers can drift. Repeat tests before concluding that a content change caused a citation gain or loss.

Chasing a magic formatting rule. Direct answers, descriptive headings and clear tables improve usability. They are not a guarantee that an AI system will cite a page. There is no reliable formatting trick that makes all engines select the same source.

How to interpret citation overlap in reporting

A useful report answers three separate questions. First, are buyers seeing the brand? Second, is the brand’s own website being used as evidence? Third, which independent sources are shaping the answer? The answers tell you whether the work should focus on brand awareness, better on-site information or a stronger external presence.

For example, if a SaaS brand appears in recommendations across several engines but its own pages are rarely cited, the next step might be improving product and documentation pages while checking whether independent reviews present accurate details. If the brand appears in one engine but not others, examine the prompts and source patterns before assuming a technical problem. If neither mentions nor citations appear, check that the tracked questions are relevant and that the brand has credible, accessible information available.

Use overlap as a diagnostic, not a target in isolation. You do not need every engine to cite the same page. You need the right information to be discoverable and trustworthy across the ways your customers search. The 5.7% all-engine figure from AEO Spotlight, for example, does not mean the other domains have no value. Its study found that the shared domains represented a large portion of aggregate citations, while the long tail still contained many engine-specific sources.

The takeaway for marketers

AI citation overlap is low because engines use different retrieval routes, respond to different prompt details and can change their own sources between runs. Research from Foglift, AEO Spotlight and Wellows all finds substantial engine-specific citation patterns, though each studies a different sample. That makes single-engine reporting an incomplete view of AI search.

Build the fundamentals that travel well: accurate first-party pages, clear answers, useful documentation and credible external evidence. Then track important prompts across engines separately, repeat observations and investigate gaps before changing your strategy. The objective is not uniform citations. It is useful visibility wherever your buyers look.

To put this into practice, start with a focused prompt set, review citations by engine and choose one content or reputation gap to address. Explore SearchFIT’s AI search resources, then review SearchFIT pricing if you need a scheduled way to track visibility and content work.