What Actually Earns Citations in AI Search
Learn what makes pages earn AI citations, from fan-out vocabulary and evidence to crawl access, entity consistency, and measurement.
Most advice on AI visibility describes what answer engines appear to do from the outside. The useful version comes from reading fan-out logs and watching which pages get pulled into an answer and which ones get skipped. This is the working list I keep for client sites. It covers what you say about yourself, what you say about competitors, where you put the information, how the page is built, and what changes when the site is an enterprise. The pressure is real. 6sense found that 94% of B2B buyers now use large language models during the buying process, and Gartner expects brands to lose 50% or more of their organic search traffic by 2028.
Fan-out is where discovery happens now
An answer engine does not run your prompt as one search. It breaks the prompt into sub-queries and runs them in the background. Semrush estimates that a complex prompt can trigger 8 to 20 or more of these. Google has not published the real count.
Two things follow from that, and both change how you write.
First, fan-out sub-queries are generated by the model. No person types them. Many of them are site-scoped lookups where search volume is irrelevant. A lookup restricted to your own domain returns whatever is on your site, no matter how many people search that string. What decides whether you show up is whether your page vocabulary matches the model's probe vocabulary. A keyword tool cannot tell you that. You have to read actual fan-out logs.
Second, there is no impression count to warn you when you are missing. In classic search the failure is visible. A page ranks position 2 for "search engine optimization services for enterprises" and gets 40 impressions a month, while it sits at position 11 for "best SEO agency" and misses 12,000. High position with low impressions is a positioning problem, not a ranking problem, and Search Console tells you it is happening. The AI version is worse. A page can be invisible to every fan-out on its topic and still look perfectly healthy in Search Console.
Match headings to probe vocabulary, not keyword volume
Here is a worked example. A runner asks a model for the best Nike running shoe for flat feet with a wide toe box. The model builds a candidate list from what it already knows, then checks each candidate against the brand's own site. Those checks are written in the buyer's words, not the brand's. Wide toe box. Heel to toe drop in millimeters. Stability or neutral. Weight in ounces.
Now look at what a shoe page usually says. React foam. Flyknit upper. Engineered for a smooth, responsive ride.
Those are material names and marketing language. The foam is the implementation. Fit and drop are what the model is checking for. If the page never states the drop as a number, never says whether the fit is wide, and never says whether the shoe is neutral or stability, none of the checks land. The shoe drops out of the answer, and a competitor with a plain spec table gets named instead.
The fix is to carry the checking words at the level of a page subject, not as a line buried in a features list. State the drop in millimeters. State the weight. State the fit as wide, standard, or narrow. State the arch support type. A model will not infer a specification you never published, and it will not guess on your behalf.
The payoff shows up fast when you fix it. Semrush ran a small experiment updating four blog articles to target fan-out phrasing and saw citations for those articles roughly double within a month. It was a volatile result on a small sample, so treat it as early evidence rather than a benchmark. The direction matches what I see. Vocabulary changes move citation rates before anything else does.
What you say about yourself
Never write your own exclusion criteria. Every "less suited to," "not ideal if," and "may not fit" gets extracted as a filter. A model building a shortlist is looking for cheap reasons to cut candidates, and a self-disqualifier is the cheapest one available. If a limitation is real, phrase it as who the product is for, not who it is not for.
Put a number on anything you want repeated. Answer engines run verification queries against published figures, including quoted exact-phrase site-scoped lookups. A number generates a verification query, and that query is another retrieval of your domain. An adjective generates nothing. "Trusted by leading brands" cannot be checked. "1,400 retailers across 70 countries" can, and the checking itself earns you a second visit.
Have one sentence that says what you are. If your homepage headline is a question or a slogan, there is no declarative entity statement to lift. A model answering "what is X" needs a plain sentence, not a value proposition.
Deliver what the title promises. If the headline says "best X solutions" and the body is a definitional guide with no vendors named, the page attracts comparison intent and answers a different question. That is retrieval waste, and you can forget about the citation.
What you say about competitors
Your own roundup can recommend a competitor. If you publish "best X software" and every rival entry carries specifics while yours carries positioning language, the model cites your page and names them. You have funded their placement with your own content. Only build a listicle if you can back your own ranking with third-party data, such as G2 or Gartner.
Include the vendors the model already has in mind. Models assemble a candidate list from priors before they search anything. A roundup missing the obvious names reads as incomplete and gets discounted. Look at who actually appears in the fan-outs, not at who your sales team calls a competitor.
Concede something. A comparison where you win every row reads as promotional. Losing one axis openly is what makes the other rows credible enough to quote.
Do not disparage. "Struggles at scale" and "limited for enterprise" are unverifiable and flag the page as marketing. Give competitor figures and attribute them to the competitor's own source.
Where the information lives on your site
Your specifications belong where the site-scoped queries land. Engines verify candidates against product pages and homepages. If your certifications, uptime, customer count, and integration count live only in a blog post, the verification step finds prose and moves on.
A term in a bullet is not a page about that term. Engines run quoted exact-phrase site lookups. If "share of voice" appears only as a line item inside a 35,000 word guide, that query surfaces a listicle instead of an answer.
Datasheets beat marketing pages. Named capabilities with mechanisms attached routinely outperform benefit prose on the same product, sometimes by a wide margin. "Validates on upload and flags blocked rows" is extractable. "Launch faster with confidence" is not.
Some claims can only be won off your domain. Support quality, satisfaction, and ease of use get verified at review sites no matter what you publish. Ahrefs analyzed 75,000 brands in December 2025 and found brand mentions across the web correlated with AI visibility at 0.656 to 0.709, while the number of pages on a site correlated at about 0.194. Publish the attestable specifics anyway, but treat that citation as a review generation problem rather than a content problem.
How the page is built
Server-side render anything you want cited. Vercel's crawler research found that GPTBot and ClaudeBot fetch JavaScript files but do not execute them. Googlebot renders JavaScript, so the assumption that anything Google indexes is also readable by AI crawlers is false. Content that only exists after hydration may as well not exist.
Read your own extracted text. Modal forms, country dropdowns, navigation, and consent banners frequently end up in the extracted body. Hundreds of junk tokens dilute the semantic signal on every affected URL, and you will not see it in the browser. The same research found roughly a third of ChatGPT and Claude crawler fetches landing on 404 pages, against 8.22% for Googlebot. Crawl efficiency is not a solved problem for these bots, and messy markup makes it worse.
Sourced statistics have to be about the page's subject. Heavy citation of off-topic research does not help. The citable material has to answer the query the page attracts, or the model quotes something that misses the point and moves on.
Compress the advice everyone else publishes. Generic "what to look for" sections attract retrieval and earn nothing, because every competitor has the same section. Lead with what is differentiated and push the shared knowledge below it. Keep dates current. A guide dated last year in a library that is otherwise current gets deprioritized.
Before you consolidate anything
Measure both channels first. The page that wins in AI retrieval and the page that wins in classic search are frequently not the same page, and sometimes not even the same URL. Semrush found that ChatGPT cites pages ranking in position 21 or lower roughly 90% of the time. Redirecting a ranking page to fix a citation problem trades a real channel for a hypothetical one. Rewriting in place is almost always the safer instrument.
Check for competing duplicates before building anything new. If three URLs cover one topic, each channel may be landing on a different one, and adding a fourth will not help.
What changes at enterprise scale
Everything above applies to a single page. At enterprise scale the problem shifts. The failure is rarely one bad page. It is that the same fact exists in six versions across six teams, and the retrieval system picks one of them for you. Six operating requirements follow.
1. Build a claim-evidence system
Enterprise sites usually fail not because they lack claims, but because the claim, its owner, its proof, and its update date are scattered across sales decks, product pages, support docs, and PDFs.
For every claim you want an AI interface to repeat, create a maintained internal record containing:
The exact approved claim.
A measurable qualifier: number, date, geography, cohort, methodology, or scope.
The primary source URL where it is stated.
The underlying evidence: customer count, uptime definition, benchmark method, certification, contract language, analyst report, or case study.
A named business owner and review date.
The external corroborator, where one exists: G2, Gartner, Forrester, ISO register, customer announcement, partner directory, government database, or audited report.
This makes your web content internally consistent and reduces a major enterprise AEO liability: one page says "global," a case study says "12 markets," a sales page says "70 countries," and a chatbot selects the lowest-confidence version.
Treat every important marketing assertion as an evidence object, not as copy. Each claim should have an approved wording, a number or scope qualifier, a first-party source URL, corroborating evidence where available, an owner, and a review date. If the business cannot point to the proof, do not expect an AI system to repeat the claim confidently.
2. Govern the entity, not just pages
Large organizations commonly publish conflicting descriptions of themselves across regional sites, investor pages, careers pages, acquisition announcements, product subdomains, partner listings, and old PDFs. This creates entity ambiguity. The system can retrieve a true statement from an outdated or narrower page and use it as the current company description.
Create an entity source of truth for the organization, each product, each business unit, each acquired brand, and each key executive or subject-matter expert.
At minimum, standardize:
Canonical company description and category.
Product names, abbreviations, legacy names, and deprecated names.
Customer segment and geographic availability.
Parent, subsidiary, and brand relationships.
Headline proof points: customers, locations, integrations, certifications, funding, compliance, uptime, and data coverage.
Official social profiles, app listings, partner directories, knowledge-base hubs, and newsroom pages.
The date each description was last validated.
This is especially important after rebrands, acquisitions, platform renames, market exits, or product packaging changes. Those are the moments when retrieval systems can connect the wrong entity, old pricing model, or obsolete capability to your brand.
Enterprise AEO is partly an entity-governance problem. Maintain one canonical description for the company and each product, then reconcile the same facts across the homepage, product pages, help center, newsroom, partner ecosystem, executive bios, regional sites, and structured data. A system cannot reliably cite a business that describes itself six different ways.
3. Make answer ownership explicit
Enterprise content often has broad topic coverage but no owner for the specific questions that buyers, analysts, procurement teams, or AI systems actually ask. Marketing owns the product page. Support owns the technical answer. Legal owns the compliance qualifier. Sales owns the competitive battlecard. The result is a fragmented answer.
Create an answer inventory around decision-stage prompts and assign an accountable owner to each answer.
Answer engine optimization
Every buyer question has one authoritative home. And one name on it.
| Buyer question | Authoritative destination | Accountable owner |
|---|---|---|
| “Does X support SSO, SAML, or SCIM?” | Security or identity-integration documentation | Product and Security |
| “How long does X retain data?” | Product documentation With plan and scope details | Product |
| “Is X HIPAA, SOC 2, or ISO 27001 compliant?” | Trust center and certification documentation | Security and Legal |
| “What does X cost?” | Pricing page With packaging, exclusions, and effective date | Pricing and Finance |
| “What are X’s API limits?” | Developer documentation With versioning | Engineering |
| “What is X best for?” | Product or category page Substantiated by use cases | Product Marketing |
| “How does X compare with Y?” | Transparent comparison page With sourced criteria | Product Marketing |
The goal is not to write a FAQ for every imagined question. It is to ensure the organization has one authoritative, crawlable, current, extractable answer for high-value questions.
Give high-value buyer questions a named owner and a canonical answer URL. In enterprise organizations, the best answer may live in product documentation, a trust center, developer docs, or a pricing policy, not necessarily in a marketing blog. AEO improves when the answer is both authoritative inside the company and easy to retrieve outside it.
4. Treat crawl access as policy
Do not let a default robots.txt, CDN rule, WAF policy, consent platform, or staging configuration decide whether an AI-search crawler can access your public commercial content.
OpenAI distinguishes its search crawler from its model-training crawler. OAI-SearchBot surfaces sites in ChatGPT's search features, GPTBot crawls content that may be used to train foundation models, and the two can be allowed or blocked separately in robots.txt. ChatGPT-User is a third case: it fetches pages when a user asks, and OpenAI states that robots.txt rules may not apply to it because a person initiated the action. Google similarly emphasizes that pages must be crawlable and indexable to be eligible for its AI features.
At enterprise scale, establish a crawler-policy register that records:
Which user agents are allowed, blocked, or rate-limited.
Which domains, subdomains, language folders, and CDN properties the rules apply to.
Whether important HTML, PDFs, help-center content, and feeds are accessible without authentication.
Whether WAF, bot management, geo-blocking, cookie walls, soft 403s, or rate limits interfere with legitimate search and retrieval agents.
Whether published robots.txt rules match edge behavior in production.
Who must approve any change that blocks discovery, search, training, data collection, or customer-requested browsing.
The point is not that every organization should permit every AI crawler. The point is that access should be a deliberate commercial, legal, privacy, and visibility decision, not an accidental side effect of infrastructure.
Separate crawler policy by use case. Search discovery, user-initiated browsing, model training, ads validation, monitoring, and internal automation are not the same activity. Decide which access you permit, document it by agent and domain, and test the live production response, not just the robots.txt file.
5. Publish reusable evidence assets
Generic explanatory content is easy to retrieve but rarely differentiates. The enterprise extension is to publish citation-grade evidence assets that competitors cannot easily reproduce.
High-leverage examples include:
Original benchmark datasets with methodology, sample size, field dates, limitations, and downloadable tables.
Product performance documentation with definitions and test conditions.
Integration directories that identify the integration mechanism, plan requirements, data sync direction, authentication method, and maintenance status.
Trust-center pages with certification scope, report date, covered products, and renewal status.
Transparent pricing and calculator pages with a version date and clear inclusions and exclusions.
Customer proof pages with named companies, quantified outcomes, timeframe, and approval status.
API and platform-limit documentation with rate limits, export limits, retention windows, storage rules, and version history.
Market maps or category databases with inclusion criteria and periodic updates.
Avoid "research" pages that cite only third-party material without adding original evidence. A system can retrieve those pages, but it has little reason to rely on them over the original source.
Create assets that other sources need to cite: benchmarks, definitions, datasets, method notes, product limits, integration specifications, compliance scope, and quantified customer outcomes. The strongest enterprise AEO asset is not merely optimized copy. It is a source of record with evidence competitors do not own.
6. Measure answer share and business return
Do not optimize only toward citations. A citation can be accurate yet commercially irrelevant. A brand mention can shape consideration but produce no tracked session. A crawler can consume substantial volume without returning meaningful referrals.
Build a reporting model that separates presence, selection, traffic, and commercial impact.
AI search measurement
Visibility fails in five different places. Measure each one on its own.
| Layer | Core question | Example metrics |
|---|---|---|
| 01 Retrieval | Did the relevant URL enter the candidate set? |
|
| 02 Answer selection | Was the brand or page included in the final response? |
|
| 03 Quality | Was the portrayal accurate and useful? |
|
| 04 Referral | Did answer exposure generate visits? |
|
| 05 Revenue | Did it influence pipeline or customers? |
|
Use a stable prompt set by persona, category, geography, funnel stage, and purchase question. Run it on a fixed cadence, preserve the model, interface, and version context, and annotate changes in site content, crawler access, product claims, and competitor moves.
Platform-native reporting is now worth pulling in. Bing Webmaster Tools added an AI Performance report in February 2026 and expanded it in June 2026 with intents, topics, citation share, and period comparison, covering citations across Copilot, Bing, and select partner AI experiences. Citation share is the useful one, because it puts your citations against every other site cited for the same grounding query. Use it alongside analytics and controlled testing. Do not treat any single vendor dashboard as a complete view of AI visibility.
Where to start
Pull your fan-out logs and list the words the model uses. Compare that list to your product page headings. Fix the vocabulary gap first, remove your self-disqualifiers second, and put a number on every claim you want repeated. If you are working at enterprise scale, run the claim inventory and the entity reconciliation in parallel, because those two determine whether the page-level fixes hold. That work is cheap, and it is the part that shows up in citation data before anything else.
Sources
6sense, 2025 Buyer Experience Report (survey of nearly 4,000 B2B buyers)
Gartner, prediction on generative AI search and brand organic traffic through 2028
Semrush, AI search traffic study (July 2025) and query fan-out analysis (August 2026)
Ahrefs, AI brand visibility correlation study of 75,000 brands (December 2025)
Vercel, AI crawler behavior research
OpenAI, crawler documentation on OAI-SearchBot, GPTBot, and ChatGPT-User
Bing Webmaster Blog, AI Performance (February 2026) and AI Visibility Insights (June 2026)