79% of ChatGPT responses draw from training weights rather than live web search. That figure, from Metricus's April 2026 analysis of ChatGPT's response patterns, is the most important number in engine-specific AEO strategy — and most practitioners have never heard it.
It means that for 8 in 10 ChatGPT queries, the content you publish today is invisible. The model answers from what it learned during training. Your new blog post, your updated schema, your refreshed FAQ — none of it exists in the ChatGPT response unless the model's training data included it or the 21% browse-mode trigger fires. For Perplexity, the opposite is true. Every query triggers live web retrieval. A page you published three hours ago with correct schema is immediately eligible for a Perplexity citation.
These are not differences in quality or preference. They are differences in architecture. Parametric memory (knowledge stored in model weights) and non-parametric memory (knowledge retrieved from external sources at query time) produce different citation behaviours, require different optimisation strategies, and operate on different timescales. Understanding which engine uses which memory type — and in what proportion — is the foundation of an engine-specific AEO programme.
What Is Parametric Memory in an LLM and Why Does It Produce Citation Lag?
Parametric memory is knowledge encoded in the model's weights during pretraining. The learning happens during training, through billions of gradient updates that adjust the model's internal parameters to predict text accurately across a vast corpus. That knowledge is then frozen — baked in to the weights. When the model answers a question from parametric memory, it is not searching anything. It is doing feedforward computation: the input tokens activate patterns in the weights, and the output emerges from those activations.
Parametric memory has three defining characteristics that directly affect AEO strategy. First, it has a knowledge cutoff. The model knows nothing that was published after its training cutoff, regardless of how important or accurate that content is. GPT-5.4 (the latest ChatGPT model as of mid-2026) carries an August 2025 knowledge cutoff. Any brand information, product feature, or market data published after that date does not exist in GPT-5.4's parametric memory (Metricus, April 2026). Second, parametric memory is expensive to update. Retraining a frontier model costs upwards of $100 million and requires months of compute across thousands of GPUs. OpenAI releases new models every three to six months, meaning training data lags real-world events by at least that long. Third, parametric memory is accessed silently — the model does not announce when it is drawing from training data versus live search. You cannot observe which memory type produced a specific answer without controlled testing.
The AEO consequence of parametric memory is citation lag. Content published after the training cutoff cannot earn parametric citations until the next model update — which may be months away. This is why new products, new features, and new studies do not appear in ChatGPT's base responses: the parametric weights have not been updated to include them. The 13-week content decay pattern documented in the content decay guide is partly a parametric memory effect: content published shortly before a training cutoff gets encoded, but content published between training updates is invisible to parametric memory until the next update.
What Is Non-Parametric Memory and How Does It Determine Perplexity's Citation Behaviour?
Non-parametric memory is knowledge that exists outside the model's weights and is retrieved at query time. It is stored in external systems — web indexes, vector databases, document stores — and fetched into the model's context window when relevant to a query. The model then reasons over the retrieved content to generate an answer. This is the RAG architecture covered in the previous article in this series.
Perplexity is the closest thing to pure non-parametric retrieval in major AI search engines. Every query triggers real-time web search via Perplexity's Sonar model and its proprietary web index covering 200+ billion URLs. The model generates answers grounded in retrieved content, not in pretraining weights, for virtually all factual queries. This is why Perplexity's citations update within hours of a page being re-crawled — the retrieval system fetches current web content on every query rather than drawing from stale training weights.
Pure non-parametric retrieval has its own failure mode: retrieval quality dependency. If the retrieval system returns poor-quality passages — thin content, low-information-density chunks, unchunkable paragraphs — the generation quality degrades regardless of how good the underlying model is. Perplexity's citation selection is therefore almost entirely a function of content structure and chunking quality rather than historical training data presence. A page published yesterday with clean BLUF structure, FAQPage schema, and correct AI crawler access competes on equal terms with a page published three years ago with the same structural quality. Perplexity does not give historical pages a systematic advantage beyond freshness weighting — which actually disadvantages old, unmaintained pages rather than helping them.
How Does ChatGPT's Hybrid Architecture Require Dual-Track AEO?
ChatGPT is neither pure parametric nor pure non-parametric. It uses a hybrid architecture: parametric weights for the majority of responses, with browse-mode RAG retrieval triggered for a minority of queries. Metricus's April 2026 analysis quantified the split: 79% parametric, 21% retrieval-augmented. The trigger for browse mode is query type — ChatGPT activates web search most reliably for explicit time-sensitive queries ("what happened yesterday"), for queries containing the year ("best AEO tools 2026"), and for queries where the model's parametric confidence is low.
This hybrid architecture creates a dual-track AEO requirement. To be cited in the 79% of ChatGPT responses from parametric memory, you need long-term entity presence: your brand, product, and key claims need to have been indexed and prominent before the training cutoff. This takes months of consistent entity building — Wikipedia presence, Wikidata entries, coverage in high-authority publications, and consistent mention across community platforms that feed training corpora. The entity authority guide covers this track in full.
To be cited in the 21% of ChatGPT responses from live retrieval, you need the same technical AEO infrastructure as for Perplexity: AI crawler access, schema, BLUF structure, clean chunking. ChatGPT retrieves via Bing (OAI-SearchBot is the live retrieval crawler; GPTBot handles training data collection). Your page needs to be indexed in Bing to enter ChatGPT's RAG retrieval pool. This is a separate indexing track from Google, and many Google-focused SEO teams have never submitted their content to Bing Webmaster Tools.
The practical dual-track schedule: run Bing Webmaster Tools setup and IndexNow configuration this week (retrieval track). Run your entity-building programme — brand mentions, publication placement, community participation — over the next six months (parametric track). Measure both channels separately in your Prompt Tracker. A rising citation rate on Perplexity before ChatGPT responds indicates the retrieval track is working. A rising ChatGPT citation rate months later indicates the parametric track is compounding.
How Does Claude's Predominant Parametric Mode Change Its AEO Optimisation?
Claude in its default configuration answers from training data without any web retrieval. The model does not browse the web when deployed on claude.ai without search enabled. This makes Claude's default citation pattern almost entirely a function of parametric memory — what Claude learned during training — rather than live retrieval.
Claude's training corpus is Anthropic's proprietary dataset, with a confirmed knowledge cutoff typically six to twelve months behind the latest Claude release. Claude uses Brave Search for live retrieval when web search is enabled — a separate retrieval track, distinct from Claude's default parametric mode. Profound's 2025 analysis found 86.7% overlap between Claude-with-search citations and Brave Search results.
For AEO purposes, Claude's predominant parametric mode means long-term entity building matters more for Claude than for any other major engine. A brand that has been consistently mentioned in high-quality publications, community discussions, and authoritative third-party sources over the past 12 to 18 months has a stronger parametric presence in Claude than a brand that published excellent content three months ago but has no prior training-data presence. The training data window favours persistent signals over recent spikes.
Claude also applies Constitutional AI values during response generation that make it more conservative than ChatGPT about citing sources with thin entity signals. A brand with a consistent, verifiable presence across Wikidata, LinkedIn, and authoritative publication coverage earns Claude citations at higher confidence than an equivalent brand with strong on-site content but no third-party corroboration. The brand hallucination guide covers how to audit and correct Claude's parametric memory about your brand when the training data contains errors.
What Does the Memory Architecture Difference Mean for llms.txt?
The llms.txt file — a Markdown file at your domain root listing your most important pages with descriptions — serves different functions for parametric versus retrieval-based engines.
For retrieval-based engines (Perplexity, ChatGPT in browse mode, Claude in search mode), llms.txt provides a curated navigation map that helps AI crawlers discover and prioritise your most important content. An AI agent doing agentic evaluation of your site reads llms.txt before navigating your full structure — it is the equivalent of giving a researcher an executive summary. Bing has explicitly confirmed llms.txt as a crawl hint for AI retrieval, making it directly relevant for ChatGPT's Bing-based retrieval track.
For parametric engines (ChatGPT's 79% parametric mode, Claude without search), llms.txt has no direct runtime effect — the model is not reading your website when generating a parametric response. The indirect benefit is that a well-structured llms.txt improves crawl prioritisation during training data collection cycles, increasing the likelihood that your most important pages are included in the training corpus with high representation.
The practical implementation guidance in the llms.txt mechanics guide applies to both tracks. The NotioncCue llms.txt Generator builds a spec-compliant file that serves both retrieval-path crawl guidance and training-data-collection prioritisation simultaneously — because the same file is what both GPTBot (training) and OAI-SearchBot (retrieval) use as a navigation signal.
How NotioncCue Helps You Manage Engine-Specific Memory Strategy
Engine-specific AEO is impossible to manage without engine-specific measurement. A tool that reports aggregate AI citation rate across all engines tells you whether your overall programme is working but not which engine track — parametric or retrieval — is driving the result or lagging behind.
The NotioncCue Prompt Tracker runs your target prompts independently on ChatGPT, Perplexity, Claude, Gemini, and Google AI Overviews and reports citation rate separately per engine. This engine-specific data reveals the memory-type pattern in your citation performance. If Perplexity citation rate is high (strong retrieval track) but ChatGPT citation rate is low (weak parametric track), the fix is entity building and training-data presence — not content structure improvements, which you have already done. If both are low, the fix starts with AI crawler access and schema (retrieval track), since that produces faster results before the longer-cycle parametric work compounds. If Perplexity responds immediately to a content change but ChatGPT does not respond for six weeks, you are seeing the difference between retrieval-engine response time and parametric-engine response time — exactly the pattern the memory architecture predicts.
The NotioncCue llms.txt Generator builds a file that serves both memory tracks: it guides retrieval-system crawlers to your highest-priority content (non-parametric track) and provides a structured navigation signal for training-data collection crawlers like GPTBot and ClaudeBot (parametric track). One 20-minute implementation step improves both tracks simultaneously. The generator validates spec compliance, checks that linked pages are crawlable, and formats descriptions to match the query intent of buyers rather than the editorial intent of your content team.
Start your free NotioncCue trial and configure engine-specific prompt tracking from day one. The memory architecture differences between ChatGPT, Perplexity, and Claude are the single most powerful explanation for why different content strategies are required across engines — and engine-specific citation data is the only way to diagnose which track is your current bottleneck.
A quick three-query test reveals whether your brand's parametric memory in ChatGPT is accurate. Run these three prompts in ChatGPT with web browsing disabled: "What is [your brand]?" "What does [your brand] do?" "What are [your brand]'s main features?" Compare the answers to your current website. Any discrepancy — wrong product name, outdated feature list, incorrect pricing model — is a parametric memory gap. The only way to close it is long-term entity signal building, not a page update. The training data that produced the error will persist until the next major model update, typically six to twelve months away. Entity corrections published now enter the next training cycle. Entity corrections published two weeks before a model release are too late.
Frequently Asked Questions About Parametric vs Retrieval Memory and AEO Strategy
How do you know whether a specific ChatGPT response came from parametric memory or live retrieval?
The most reliable indicator is whether the response cites source URLs. Parametric responses typically do not cite URLs — the model is drawing from trained weights, not retrieved documents. Browse-mode RAG responses include citation links. A ChatGPT response without citations on a factual query is almost certainly from parametric memory. You can confirm by asking a follow-up: "Did you search the web for that answer?" — ChatGPT will disclose whether it used search tools in its response generation.
Does Gemini use parametric or retrieval memory for AI Overviews?
Google AI Overviews are retrieval-based — they pull from Google's search index using a RAG architecture similar to Perplexity's but built on Google's proprietary index and reranking system. Base Gemini model responses (in Gemini.google.com) use a hybrid approach similar to ChatGPT: primarily parametric with retrieval triggered for time-sensitive or high-uncertainty queries. For AEO targeting Google AI Overviews specifically, retrieval-track optimisation (content structure, schema, Google indexing) is the primary lever. For Gemini model responses in other contexts, parametric entity building becomes more important.
Can you accelerate parametric memory updates by publishing content more frequently?
Not through publication frequency alone. Parametric memory updates happen during model training runs, which OpenAI, Anthropic, and Google control independently of publication schedules. What you can influence is training data quality and prominence: content that earns more third-party citations, backlinks, and community references before a training cutoff is more likely to be represented in the trained weights with higher confidence. Publishing one well-cited, widely-referenced piece before a training cutoff is more parametrically effective than publishing ten pieces that receive no third-party attention.