How AI models learn to cite new sources - and drop old ones

AI models change their citation patterns at three layers: discovery, where crawlers first find your content; training, where periodic retraining reweights which sources look authoritative; and retrieval, which selects sources at query time. Each layer moves on its own cycle, which is why sites gain and lose citations without changing anything themselves.

Key takeaway: AI engines do not cite pages the way search engines rank them. There is no single algorithm making a citation decision in real time. Instead, the process runs across three distinct layers: what the model architecture allows to be indexed, what the training process decides to weight, and what the retrieval system selects at query time. Each layer is governed by different signals and requires different optimisation strategies.

Layer 1 - Discovery: how AI engines find your content

Before any citation decision is made, a model must first be aware that your content exists. Discovery is the process by which AI systems become aware of web content in the first place.

Retrieval-based engines like Perplexity, Google AI Overviews, and Bing Copilot discover content through web crawls, sitemaps, RSS feeds, and API integrations with search engines. These systems run continuously, re-crawling known sources and discovering new ones. If your site is not accessible to these crawlers, or if your content lacks clear structural signals that make it machine-readable, you are invisible at the discovery layer before a single citation decision is made.

Training-based models like ChatGPT, Claude, and Gemini access content differently. Their knowledge comes from training runs in which vast quantities of web content are processed and encoded into model weights. A page that existed before the training cutoff but was never re-processed in a subsequent training run can disappear from a model’s knowledge entirely, even if it remains live on the web. This is the primary reason that citations can drop without any change on your part.

The practical implication: discovery requires both technical accessibility and temporal relevance. Your site must be crawlable, and your content must appear in training data that the model actually uses.

Layer 2 - Extraction: what AI models lift from your content

Once a model is aware of your content, it must decide what to extract and store. This is the extraction layer: the process by which a model identifies which passages, claims, and data points from a source are worth retaining.

AI models extract content that is self-contained and directly answerable. A paragraph that begins with a clear statement of fact, defines a term, or provides a specific figure is far more likely to be extracted and stored than a paragraph that builds an argument across multiple sentences. The ideal extraction target is a passage that answers a specific question without requiring context from surrounding paragraphs.

Consider the difference between these two passages on the same topic. Passage A: “When evaluating software options, teams should consider total cost of ownership, implementation complexity, vendor lock-in risk, and support quality. Each factor interacts with the others in ways that vary by organisation size and industry.” Passage B: “Total cost of ownership for SaaS project management tools typically ranges from GBP 8 to GBP 45 per user per month, depending on feature tier and contract length. Most vendors offer annual plans with a 15-20% discount versus monthly billing.”

A model processing these passages for a query about project management tool pricing will extract Passage B almost every time. It contains a specific, quotable, directly answerable statement. Passage A provides context and framing but gives the model nothing concrete to cite.

Layer 3 - Reinforcement: why some sources get cited repeatedly

Extraction determines what a model can cite. Reinforcement determines what it actually does cite, especially in recurring or high-stakes queries.

Reinforcement is partly a training effect. When multiple users ask variants of the same question and a model consistently cites a particular source, that citation pattern gets reinforced through feedback loops in the model’s use. Human preference data, where raters indicate which responses they find more helpful, can amplify existing citation patterns even when competing sources are equally valid.

The result is a self-reinforcing cycle: established sources get cited more, which amplifies their presence in training data, which makes them more likely to be cited again. New entrants face a cold-start problem where they must overcome the existing citation bias before their content is regularly retrieved.

Authority and recency interact differently at the reinforcement layer. For factual queries with stable answers, older authoritative sources tend to dominate. For queries about current events, emerging products, or rapidly evolving topics, recency outweighs legacy authority. A 2023 article from a niche trade publication on a breaking regulatory development will consistently beat the Wikipedia citation from 2021 for that specific query.

The signals that govern citation selection

Across all three layers, certain signals consistently influence whether a source gets cited. These are not secret formulas but documented patterns observable through repeated querying and analysis.

Entity prominence: Models show measurable preference for sources associated with well-known entities. A claim from a company’s official blog about their own product pricing will typically be preferred over the same claim on a third-party review site. Entity prominence is why brand-owned content often outperforms comparison content for brand-specific queries.

Structural clarity: Question-format headings, numbered lists, comparison tables, and FAQ schema all signal to models that content is structured for extraction. These formats reduce the model’s interpretive workload when deciding what a page is about and what it answers.

Citation density: Pages that are themselves heavily cited by other sources across the web tend to receive higher citation rates from AI models. This creates an indirect SEO-to-AI-visibility pathway: traditional link authority influences AI citation patterns even without direct model training on link graphs.

Training corpus presence: For training-based models, whether your content was included in the most recent training run matters enormously. Content that appeared in GPT-4’s training data but was not included in the GPT-4o update may have effectively disappeared from ChatGPT’s active knowledge, even if it remains findable via web search in Perplexity.

How to diagnose which layer is failing you

Before optimising, diagnose which layer is causing your citation gap. The approach differs depending on which engine you are targeting.

1. Test retrieval-based engines first. Query Perplexity and Google AI Overviews directly with your target queries. If your brand or content does not appear in these results, the issue is likely at the discovery or extraction layer: your content is not being found, or it is being found but not extracted as a cited passage.
2. Check training-based engines separately. Test ChatGPT, Claude, and Gemini for the same queries. If these engines do not mention your brand but retrieval engines do, your content is being discovered but is missing from training data. The fix is different: you need to ensure your content appears in sources that models explicitly crawl, or publish data that is distinctive enough to be collected by training pipelines.
3. Audit your extraction signals. For the pages you want cited, examine whether the content following each heading begins with a direct, self-contained answer. If you find yourself leading with context, background, or narrative framing, you are making the extraction layer work harder than necessary. Restructure with answer-first paragraphs and measure the effect on citation rate over four to six weeks.

Measuring citation change over time

Citation visibility is not static. Even when you make no changes to your site, your citation rate can shift due to competitor activity, model updates, and seasonal shifts in query volume. The only way to know whether your optimisation is working is to track citations systematically.

At minimum, run monthly queries across your five to ten most important target queries and record which sources are cited. Look for three signals: whether your brand appears at all, your share of citations within your category, and whether your citation share is growing or shrinking relative to competitors.

Automated tracking tools like SearchScore Tracker run these checks weekly across ChatGPT, Claude, Gemini, Perplexity, Grok, and DeepSeek, and alert you when citation share drops below a threshold you define. Given that models can retrain and shift citation patterns within a matter of weeks, monthly manual checks are a minimum; weekly automated checks are the reliable approach.