How Do LLMs Choose Sources and Citations?
AI answer systems choose sources through a sequence of decisions. They interpret the prompt, decide whether current information is needed, create one or more retrieval queries, find candidate pages and passages, rerank those candidates, select evidence and generate a response. Citations attach to claims that connect to a specific source. Your page can shape an answer without receiving a visible link.
The 5-minute explanation
When someone asks ChatGPT, Perplexity or Google a question, the large language model is only one part of the product handling the request. Search indexes, retrieval systems, ranking models, safety rules, account context and citation interfaces all play a role.
Most search-backed answers follow eight stages:
- The product interprets the prompt. It identifies the subject, the type of answer required and any ambiguity in the wording.
- It decides whether live information is useful. Stable questions often draw on the model's existing knowledge. Questions about current prices, recent news or a specific source are more likely to trigger retrieval.
- The prompt becomes one or more search queries. The product shortens conversational requests, rewrites them or splits them into several related searches.
- Candidate content is retrieved. Relevant pages and passages are pulled using keyword matching, semantic similarity or both.
- The candidates are reranked. Initial results are scored again in the context of the full question. Relevance, clarity, freshness, authority, corroboration and source diversity all affect this stage.
- Evidence is selected. The product chooses the passages most useful for building the answer.
- The response is written. The model combines and paraphrases the selected evidence. Several sources often contribute to a single paragraph or sentence.
- Citations are attached. Where the product supports citations, visible claims link to sources that appear to support them.
This explains why traditional search position and AI citation visibility often diverge. One page ranks well because it satisfies Google's organic ranking system. Another supplies a cleaner passage for the specific question being answered.
So the second page gets the citation.
It also explains why a source can affect an answer without receiving a link. The product uses a passage to understand a topic, establish context or resolve a conflict, then attaches the visible citation to a different source whose wording supports the final claim more directly.
Mentions, citations and recommendations are three different outcomes. A mention is when a company, person or product appears in the answer, from model knowledge, retrieved content or both. A citation connects part of the answer to a source URL and shows which page the product chose to expose as supporting material. A recommendation goes further: the answer presents an entity as a suitable option for a particular user, problem or set of constraints.
A well-known company gets mentioned from model knowledge without a citation. A research paper gets cited without its publisher being named. A product gets cited as evidence but excluded from the recommendation because it doesn't fit the user's budget or location.
Search-backed and memory-based answers create different routes to visibility. Memory-based answers rely on patterns encoded during training and later model updates. They name established entities without consulting a current webpage. Search-backed answers build a fresh shortlist at request time, giving newer or more specialised sources a chance to compete.
Publishers have the most influence over the search-backed route. Make pages crawlable, keep important information in accessible text, use stable URLs, clarify entities, publish evidence and write passages that answer specific questions without requiring the reader to interpret the rest of the page.
The most useful passages have a clear subject, a direct claim and enough context to stand on their own. "Our platform improves marketing performance" offers little evidence. Stating what was measured, over what period and against which baseline gives the retrieval layer something concrete to work with.
Independent sources back up your own claims. When your company's claims align with trusted reviews, directories, research, customer evidence and editorial coverage, several independent sources point in the same direction. That helps with entity resolution and gives the answer engine more options when selecting evidence.
Some parts of the process remain outside publisher control. Platforms don't publish their complete reranking formulas. Search gets invoked differently across products, accounts and prompts. Location, language, account history and connected data alter the available context. Citation interfaces vary too, so identical evidence receives different treatment on different platforms.
If you want somewhere to start:
- Confirm that the page is indexable and that its main content renders as accessible HTML.
- Place a direct answer near the start of each important section.
- Use consistent names for the company, person, product and category across owned and third-party pages.
- Support important claims with dates, numbers, methodology and specific examples.
- Measure mentions, citations and recommendations as separate outcomes.
These changes improve your page's chances of entering the shortlist and surviving evidence selection. The rest of this guide looks more closely at each stage.
What Actually Chooses the Sources? The Five-Layer Visibility Model
I use a five-layer visibility model to diagnose where a company or page is losing ground. Each layer represents a separate question in the path from model knowledge to a visible recommendation.
| Layer | Core question | Main influences | Publisher control |
|---|---|---|---|
| Training visibility | Does the model have an established representation of the entity? | Historic mentions, consistent associations and training data coverage | Low |
| Retrieval visibility | Can the system find the page or passage when answering? | Crawling, indexing, rendering, relevance and freshness | High |
| Selection visibility | Does the retrieved material become part of the evidence set? | Passage quality, relevance, authority, corroboration and clarity | Medium to high |
| Citation visibility | Does the page receive a visible source link? | Claim attribution, passage specificity and citation interface | Medium to low |
| Recommendation visibility | Is the entity presented as a suitable choice? | Fit, evidence, reputation, availability, safety and user context | Medium |
Take a consultancy with a well-indexed guide. The model already recognises the company from older web references, giving it training visibility. The guide enters the shortlist for a relevant query, giving it retrieval visibility.
The page can still lose at the next stage. A competitor has a clearer definition, stronger data or a passage that answers the prompt more precisely. That competitor's passage gets selected as evidence.
Even when the consultancy's guide contributes to the answer, another source receives the visible citation because its wording maps more cleanly to the final claim. Recommendations introduce another set of considerations: location, pricing, reputation and suitability for the user's situation.
This is what makes the framework practical: each layer points to a different fix. A crawl problem needs a technical fix. Weak selection calls for better evidence and passage structure. Low recommendation visibility needs clearer positioning, proof and third-party validation rather than another round of on-page optimisation.
In my experience, retrieval visibility is where most publishers should start. AEO-Lite lets you check your retrieval and citation readiness against this model in under a minute.
When Does an AI System Search? Model Memory vs. Live Retrieval
AI products answer from model knowledge, retrieve current information or blend both.
Model knowledge works well for stable concepts and common facts. "What is customer lifetime value?" usually gets answered without checking the web. A query about today's exchange rate or a newly released product needs current information.
The decision also depends on the product. ChatGPT chooses web search based on the request, and the user can invoke search directly. Search-backed responses include links to relevant web sources. (openai.com)
Blended answers combine background knowledge with recently retrieved evidence. The model already understands the topic. The search layer supplies current details, specific examples or supporting sources.
Citation visibility depends heavily on this decision. Memory-based responses have no newly retrieved webpage to cite. Search-backed responses create an opportunity for your page to be found, selected and linked.
Exact trigger logic varies by product. Requests for current sources, prices, availability or recent developments give a strong reason to search. Stable questions, casual conversation and familiar concepts often stay within model knowledge.
How a Prompt Becomes a Retrieval Query
People write prompts as requests, explanations and conversations. Search systems work better with focused queries, so the product interprets the prompt before retrieval begins.
This interpretation usually covers four tasks:
- identifying the main intent
- recognising people, companies, products and places
- resolving ambiguous terms
- deciding which facts are needed to complete the answer
Complex prompts get split into several searches. Google describes this as query fan-out in AI Mode, where related searches run across subtopics and data sources before the results are combined. (blog.google)
Take this prompt:
Should a Series A SaaS company with a small marketing team use Customer.io or Klaviyo?
An illustrative set of retrieval queries could include:
- Customer.io versus Klaviyo for SaaS
- Customer.io pricing for a small team
- Klaviyo B2B SaaS features
- Customer.io implementation requirements
- Klaviyo CRM and ecommerce focus
The actual queries remain hidden, but the example shows why exact keyword matching has limited reach. A useful page covers the entities, criteria and decisions behind the request, not only the original wording.
Clear product categories, audience descriptions, pricing information, comparison criteria and limitations all create more opportunities to match the generated subqueries.
How Candidate Pages and Passages Get Retrieved
Retrieval creates the initial pool of possible evidence. Most systems combine two broad approaches.
Lexical retrieval looks for overlap between the words in the query and the words in a document. BM25 is a common method. It rewards relevant term matches and reduces the value of words that appear on almost every page.
Exact product names, technical terminology and specific phrases benefit most. A page that clearly names Customer.io, lifecycle marketing and SaaS onboarding sends strong lexical signals for those concepts.
Semantic retrieval looks at meaning. Text is represented as numerical vectors called embeddings. Passages with similar meaning sit close together in this vector space, even when they use different wording.
A page about reducing subscriber loss matches a query about churn prevention without repeating that exact phrase.
Many systems use hybrid retrieval, combining lexical and semantic results. Exact terms identify the right entity. Semantic matching broadens the pool to passages that answer the same underlying question in different language.
Retrieval often works at passage level. A long guide gets broken into smaller sections or chunks, and only one of those passages enters the answer context. Clear headings and self-contained explanations help each section retain meaning after extraction.
Technical accessibility comes first. The search layer needs a stable page, crawl permission, index eligibility and readable text. Important claims hidden inside images, complex interactions or unreliable client-side rendering create unnecessary retrieval risk.
For the full technical eligibility checklist, see the AEO Growth Playbook.
How Sources and Passages Are Reranked
The first retrieval pass aims for coverage. Reranking narrows the pool to the material most useful for the specific answer.
Each candidate gets scored on:
- relevance to the full question
- clarity of the passage
- fit with the required language and location
- freshness for time-sensitive queries
- authority and source reputation
- support from independent sources
- diversity across the evidence set
Freshness has different value depending on the question. Current pricing and regulations favour recently updated material. A durable explanation of statistical significance doesn't become better simply because it was published yesterday.
Authority works in context too. An official product page is a strong source for current features and pricing. An independent comparison is more useful for assessing usability or trade-offs. A forum reveals practical problems that the vendor's own documentation doesn't cover.
Source diversity prevents an answer from repeating one publisher's viewpoint. Three pages that copy the same original claim provide less independent support than three sources with separate evidence.
Organic search performance influences which pages enter the shortlist, especially when the AI product relies on an established search index. The final evidence set still reflects the needs of the generated answer. A lower-ranking page with a precise passage receives the citation ahead of a broader page ranking above it.
In my view, this is the part most AEO advice gets wrong. I've watched a mid-ranking guide on mortgage repayments get the citation while stronger organic pages got ignored, simply because its passage answered the repayment question in one clear block.
How Selected Evidence Becomes an Answer
After reranking, the product assembles a working context from the selected passages. The model receives extracts rather than complete pages, and that context has limited space. Passages get shortened, deduplicated or compressed before generation, so repeated information is reduced while useful distinctions and supporting facts are retained.
Conflicting sources require a decision. Suppose one page lists a product at €99 per month and another lists €129. The product favours the official pricing page, mentions that pricing varies by plan, or avoids stating an exact figure when the evidence can't be reconciled.
The model then writes a coherent response from this mixed evidence. It combines a product definition from one page, pricing from another and user feedback from a third. The final wording is usually a synthesis rather than a direct extract.
Writing for retrieval is different from writing a sentence you expect the model to repeat. The strongest contribution is often a clear fact, definition, comparison or piece of evidence that survives paraphrasing.
Why Some Used Sources Are Never Cited
Evidence selection and visible attribution happen at different points in the pipeline.
One source helps the product understand terminology, frame a comparison or confirm a fact. The final sentence then links to whichever source supports its most specific claim. The other contributing pages stay invisible.
Sources also support only part of a sentence. An answer combines a pricing fact, a market observation and an opinion, then displays one citation beside the entire line. The link doesn't prove every part of the statement.
Citation interfaces create further differences. Some products place links beside individual sentences. Others show source cards or a reference panel. The number of exposed links is often smaller than the number of pages used during synthesis.
Location, language, account settings and connected sources alter the evidence pool. Two people asking the same question receive different sources because the product has different context about each request.
Citation tracking is valuable, but read it as visible attribution rather than a complete record of everything that influenced the answer.
Platform Comparison: How ChatGPT, Perplexity, Google, Gemini, Copilot, and Claude Differ
The main products differ in when they search, what context they use and how they expose sources.
| Platform | Retrieval pattern | Citation behaviour | Practical publisher implication |
|---|---|---|---|
| ChatGPT | Search is invoked when the request benefits from current web information or when the user selects search | Search-backed answers show source links, while many knowledge-based answers have none | Build both clear entity recognition and strong search-time eligibility |
| Perplexity | Web retrieval is central to the standard answer experience | Inline citations are a core part of the interface | Publish current, source-like pages with specific and attributable passages |
| Google AI Overviews and AI Mode | Built on Google Search systems, with query fan-out for complex requests | Links appear within or beside generated responses, depending on the surface | Google indexing remains foundational, while passage selection can differ from organic ranking order |
| Gemini | Behaviour varies across the Gemini app, Google services and connected contexts | Sources depend on the product mode and available grounding | Evaluate the specific Gemini surface instead of treating it as identical to Google Search |
| Microsoft Copilot | Public web answers can use Bing grounding, while workplace products can also use Microsoft Graph | Public sources are commonly shown in grounded answers | Bing eligibility matters for public content, while enterprise visibility depends on organisational access and permissions |
| Claude | Web search is used when enabled and relevant to the request | Web-backed answers include direct citations | Create pages with clear evidence and enough context to support progressive research |
OpenAI documents selective search and linked web sources. Google documents query fan-out and the use of Search information systems. Anthropic documents current web search with direct citations. (anthropic.com) (openai.com)
The table describes product behaviour, not fixed provider relationships. Search infrastructure, interfaces and model routing change over time. No two products share the same shortlist or selection logic.
A page that performs well in Perplexity receives fewer links in ChatGPT when ChatGPT answers from model knowledge. A page cited in Google AI Mode doesn't surface in Gemini when the request is grounded in connected account data. Platform-specific testing beats any universal AI ranking score.
What Increases Selection and Citation Probability
Technical eligibility
Your source needs to be accessible before its quality matters.
Check that important pages can be crawled, indexed and rendered as text. Keep canonical URLs consistent, submit accurate sitemaps and remove accidental indexing blocks. Confirm visibility in both Google and Bing. Coverage in one index doesn't guarantee coverage in the other.
Check your crawler rules too. Some providers distinguish between bots used for model training and bots used for search or answer grounding. Where the platform supports it, you can restrict training access and keep content available for search-backed answers.
Stable pages are easier to retrieve and cite. Frequent URL changes, duplicate versions and conflicting canonical tags weaken the connection between a claim and its source.
Entity clarity
Answer engines need to recognise that mentions across your pages refer to the same company, person or product.
Use consistent names and descriptions on the homepage, About page, author profiles, product pages and third-party listings. Make the category, market, geography and audience explicit.
Describing yourself as a growth consultancy on one page, a performance agency on another and an AI transformation studio elsewhere creates an avoidable entity-resolution problem. The descriptions can differ, but they should point to a coherent identity.
Structured data supports machine understanding when it accurately reflects visible content. Organization, Person, Article, Product and LocalBusiness markup clarify relationships and attributes. Schema should document the page, not make claims that users can't see.
Passage and content design
Write each passage to answer one clear question and carry enough context to survive outside the page.
Start important sections with the answer. Follow with evidence, explanation or examples. Use descriptive headings and keep paragraphs focused.
Evidence improves a passage's value. Include named sources, dates, methodology, original figures, product specifications and clear comparison criteria. Distinguish measured results from interpretation.
Tables work well when the reader needs to compare several options. Checklists help with procedures. Definitions should be concise enough to quote and accurate enough to trust.
Don't force every section into the same template. A natural article still contains strong extraction points without reading like a collection of featured snippets.
External authority and corroboration
Your website provides first-party information, while independent sources show how that information holds up outside your own presentation. Trade publications, recognised directories, research, review platforms, customers, partners and professional profiles all help, and the quality and relevance of the source matter more than raw volume.
External references also help resolve entities. When reputable sources consistently connect the same company name with the same category, people and products, the answer engine has a clearer body of evidence.
Link building supports this, but the goal is broader than acquiring a high link count. The stronger outcome is an ecosystem of independent pages that confirm your organisation's identity, expertise and claims.
Commercial and recommendation readiness
Recommendation queries include practical constraints. Your product needs to fit the user's situation, not simply appear relevant to the category.
Make audience, price range, geography, availability, limitations and main use cases easy to find. If you run a consultancy, explain which companies you serve, the problems you handle and the type of engagement you offer.
Proof helps the product make a safer recommendation. Case studies, reviews, recognised customers and specific results provide more support than broad statements about quality.
Clear constraints help too. A product that openly states who it serves and where it falls short becomes a better candidate for the right prompt because the product matches it with greater confidence.
What Does Not Have Strong Evidence Yet
| Claim | Evidence status | More accurate interpretation |
|---|---|---|
| Training inclusion guarantees future mentions | Unsupported | Training may create background recognition, while current answers can still depend on retrieval, context and product rules |
| AI visibility is simply SEO with a new name | Incomplete | SEO supports technical eligibility and authority, while answer synthesis, passage selection and citation add further stages |
| Schema directly increases AI citations | Unsupported as a universal ranking factor | Accurate schema helps machines interpret visible content and entities |
| llms.txt is an AI ranking factor | Unsupported | It may serve as a voluntary content map, but major platforms have not documented citation or ranking benefits |
| Page-one Google rankings guarantee AI citations | False | Organic ranking can influence the candidate pool without determining the final evidence set |
| More backlinks automatically create more AI mentions | Oversimplified | Trusted links and independent references support authority and corroboration, while raw counts reveal little on their own |
| A visible citation proves complete provenance | False | A link may support only part of the adjacent claim, and additional uncited sources may have contributed |
The standard worth holding to: describe platform-documented behaviour as documented, experimental findings as experimental and reasonable interpretation as interpretation. Confidence should follow the evidence. I've sat with page-one results that never once showed up as a citation, while a quieter page further down the SERP kept getting the link. Ranking helps you get into the room. It doesn't decide who speaks.
How to Measure AI Citation Visibility
In my experience, a single visibility score hides too much. Separate the observable outcomes instead.
Entity visibility
Track whether your brand, product or person appears in the answer, how accurately it is described and which competitors appear beside it.
Mention rate is especially useful across non-branded category and problem queries. Branded prompts mainly test whether the product describes an entity that the user has already named.
Citation visibility
Record whether your domain is cited, which URL receives the link and where the citation appears. Review whether the page fully supports the claim, supports part of it or has been cited inaccurately.
Separate first-party and third-party citations. You might have low citation share on your own domain but appear frequently through review sites and editorial coverage.
Recommendation visibility
Track whether your entity is presented as an option, whether it appears prominently and which conditions shape the recommendation.
A passing mention in a list differs from a direct recommendation for a defined audience.
Those outcomes have different commercial value.
Prompt-cluster coverage
Build a stable set of prompts covering:
- branded questions
- category discovery
- problem-based searches
- comparisons
- recommendations
- local intent
- expert discovery
- current or time-sensitive information
Run the same clusters across platforms. Keep the wording stable enough to compare results. Add a smaller set of natural variations to see how sensitive the answer is to phrasing.
Commercial effect
AI visibility influences a decision without producing an immediate click. Combine referral traffic with branded search growth, lead-source questions, CRM notes and assisted-conversion analysis. More than once, a lead has told me they first found the company in ChatGPT, then searched the brand later. The click never appeared in analytics. The conversation still started there.
A simple tracking table can include:
| Field | Example |
|---|---|
| Date | 15 July 2026 |
| Platform and surface | ChatGPT Search |
| Location and language | Netherlands, English |
| Account state | Logged in, memory enabled |
| Prompt category | Comparison |
| Exact prompt | Customer.io or Klaviyo for a small SaaS team? |
| Web search used | Yes |
| Brand mentioned | Yes |
| Recommendation present | No |
| Citation URLs | Recorded URLs |
| Competitors mentioned | Klaviyo, HubSpot |
| Notes | Correct category, outdated pricing |
Repeat tests over time. Generated answers vary between runs, models and accounts. One result is an observation. A repeated pattern is a signal.
Publisher Checklist
Technical
- Confirm indexing in Google and Bing.
- Check canonical tags, sitemap entries and rendered HTML.
- Review training and search crawler rules separately where supported.
- Keep important pages stable and accessible.
Entity
- Use consistent company, product and author names.
- State category, audience, market and geography clearly.
- Align the website with major external profiles.
- Keep structured data accurate and visible-content based.
Content
- Answer the main question early in each section.
- Create passages that make sense when extracted.
- Add dates, numbers, methods and sources.
- Use tables and checklists where they improve understanding.
- Update claims that depend on current products, prices or rules.
Corroboration
- Build independent references around important claims.
- Prioritise relevant and credible sources over volume.
- Keep owned and external descriptions consistent.
- Track which third-party pages receive citations for your brand.
For the full 30-day implementation plan, see the AEO Growth Playbook.
Not sure where your own page stands on any of this?
AEO-Lite checks retrieval and citation readiness, including crawlability, entity clarity, passage structure and evidence signals, against the mechanics explained above. It flags structural and evidence gaps. It does not predict whether any specific AI system will cite you.
Run a free check →FAQs
What is the difference between being mentioned and being cited in an AI answer?
A mention means an entity appears in the answer. It may come from model knowledge, a retrieved page or another source that discusses the entity. A citation is a visible link connecting a claim to a webpage. The two outcomes can happen separately. A brand may be mentioned with no link, while a report can be cited without its publisher being named. Track mention rate to understand recognition and citation rate to understand visible source attribution.
Does ChatGPT always search the web?
ChatGPT can answer from model knowledge or use web search. OpenAI says ChatGPT chooses to search based on the request, while users can also activate search directly. Questions involving recent events, current prices or explicit source requests are natural candidates for search. Stable conceptual questions may be answered without it. Search-backed answers can include links to web sources. (openai.com)
Can I block AI training and remain visible in AI search?
Some platforms use separate crawlers or controls for training and search. This allows publishers to restrict training access while keeping content available for search-backed answers, where supported. The exact setup differs by provider and can change, so crawler documentation should be reviewed directly. Blocking a training bot doesn't automatically remove a page from every search or grounding product.
Does schema markup improve AI citations?
Schema helps search systems understand entities, page types and attributes when it accurately matches the visible content. There's no established cross-platform rule showing that adding schema directly increases citations. Use it to clarify the Organization, author, article, product or local business represented on the page. Treat it as part of technical and entity quality rather than a citation shortcut.
Why do different AI platforms cite different sources for the same question?
Each product builds its own shortlist and runs its own retrieval, reranking and citation process. Search indexes, query rewriting, account context, location and interface rules all differ. One platform searches; another answers from model knowledge. Even when both search, they generate different subqueries or prefer different passages. So the same topic gets cited by Perplexity, omitted by ChatGPT and surfaced in Google AI Mode.
Do backlinks matter for AI visibility?
Trusted links improve search visibility, authority and independent corroboration. All three support AI retrieval and selection. Raw backlink count is a poor standalone measure. A relevant editorial reference or recognised industry listing provides more useful evidence than many unrelated links. Focus on links and mentions that help define the entity, confirm its expertise or support a claim that an answer engine needs.
Does llms.txt work?
There's no strong evidence that llms.txt improves rankings, mentions or citations across major AI products. It may provide a convenient map of preferred content for tools that choose to read it, but broad platform adoption and measurable citation benefits remain unproven. It can be a low-cost addition after indexing, internal linking, entity clarity and content quality are in order. It shouldn't displace work with clearer evidence behind it.
Why does AI cite a competitor that ranks below me on Google?
Organic ranking and AI evidence selection solve related but different problems. Google ranks your page as the strongest overall search result. An answer engine finds a competitor's passage more precise for one part of the prompt. The competitor might also have newer data, clearer comparisons or stronger third-party corroboration. Review the exact passage that received the citation rather than comparing only domain positions.
How should AI citation visibility be measured?
Track mentions, citations and recommendations separately across a stable set of prompts. Record the platform, date, location, account state, wording, cited URLs and competitors. Review whether citations actually support the claims beside them. Repeat the tests because generated answers vary. Connect this visibility data with referral traffic, branded search and lead-source information to understand commercial impact.
Can my content influence an AI answer without being cited?
Yes. A passage contributes background, terminology, comparisons or evidence while another page receives the visible link. Model knowledge also reflects previous exposure to an entity without retrieving a current webpage. Citation reports show visible attribution, not the full chain of influence. The effect is difficult to measure directly, which is why citation data should be combined with mention tracking and answer-content analysis.
Does freshness always improve citation chances?
Freshness matters when the answer depends on change. Prices, product features, regulations, news and availability all benefit from current sources. Stable concepts often reward depth, clarity and authority instead. Updating an evergreen page without improving its substance adds little. Use visible review dates and update the facts that have genuinely changed.
Does Wikipedia help a brand appear in AI answers?
Wikipedia provides a clear, widely referenced entity description and often appears in search-backed research. A legitimate article supports recognition and corroboration. It's neither available nor appropriate for every company, and creating one requires independent notability and reliable sources. The broader objective is a consistent, independently documented entity footprint. Wikipedia forms part of that footprint when the subject qualifies.
Methodology and Limitations
This guide combines platform documentation, information-retrieval concepts, academic and industry research, and a practical framework for diagnosing AI visibility.
Platform-confirmed behaviour is presented with greater confidence than conclusions drawn from black-box testing. Observational findings can reveal patterns, but they don't expose the internal weights used by a product. Provider relationships, retrieval systems and citation interfaces also change over time.
The five-layer model I use is a diagnostic framework for publishers and marketers. It separates the points where an entity becomes known, found, selected, cited and recommended. Its purpose is to make testing and optimisation more precise, and still leave room for the parts of the pipeline that remain undisclosed.
Last reviewed: July 15, 2026