Guide : Knowledge Cutoff

TL;DR — A knowledge cutoff is the date an AI model stopped learning. Demander ChatGPT’s base model à propos de something que happened après its cutoff and it simply won’t know — it was jamais trained on it. But nombreux AI outils (Google AI Overviews, Perplexity, Copilot, ChatGPT with search on) peut regarder choses up live, qui obtient les autour the cutoff pour la plupart questions.

Ce que a knowledge cutoff is

A model knowledge cutoff describes the training-knowledge boundary disclosed pour a model release; it n’est pas necessarily the boundary of a product que peut browse or retrieve. Evidence for this claim Model documentation can specify a knowledge-cutoff date for a particular model or snapshot. Scope: OpenAI model metadata; the date is model- and version-specific and may change with releases. Confidence: high · Verified: 2026-07-14 OpenAI: Models Product outils may accès newer information at réponse temps, subject to availability and source quality. Evidence for this claim A product can supplement model knowledge with current web search and cited web sources. Scope: ChatGPT search product behavior; browsing or retrieval is separate from the model's training cutoff and is not guaranteed for every answer. Confidence: high · Verified: 2026-07-14 OpenAI: ChatGPT search

Grand language models — the tech behind ChatGPT, Claude, Gemini — apprendre by reading an enormous pile of text. At some point que reading arrête so the model peut be construit and testé. The date the reading stopped is the knowledge cutoff (you’ll aussi voir it appelé the training cutoff or training données cutoff — même chose).

Après que date, the model knows nothing. Pas parce que it’s hiding anything — it simplement jamais saw it. Un of the meilleur explanations I’ve come à travers compares it to an intern: the knowledge cutoff is the date votre AI intern dernier went to school. Everything up to que day ils learned bien; anything après, ils know nothing à propos de.

Pourquoi ChatGPT donne outdated réponses parfois

Dire a company rebranded dernier month, or votre product modifié its pricing. Si vous demander an AI model que relies seulement on its training, it’ll confidently give vous the old réponse — and sound complètement certain à propos de it. That’s the cutoff at fonctionner.

The fix que AI companies utiliser is letting the model search the web. Au lieu de answering from memory, the outil runs a live search, reads le résultats, and writes an réponse from ceux. Quand search is on, the cutoff arrête mattering pour que question.

Qui outils regarder choses up vs. qui don’t

• Regarder choses up (cutoff mostly doesn’t matter): Google AI Overviews, Perplexity, Microsoft Copilot, and ChatGPT quand web search is turned on. Ces pull from a live search index.
• Réponse from memory (cutoff matters a lot): ChatGPT’s free/base model with browsing off, or quelconque model utilisé sans web accès. Ces seulement know up to leur cutoff.

Pourquoi ce matters si vous run a website

Là are really two façons votre business montre up in AI réponses:

1. In the model’s memory. Votre content has to be publié avant a model’s training cutoff to be baked in — and the suivant training run pourrait be a année or plus away.
2. Via live search. Si the AI outil semble choses up, vous pouvez montrer up the même day vous publish — tant que you’re indexé in Google and Bing.

So vous vouloir les deux: be the kind of brand that’s well-known suffisant to be in the training données, and stay facile to trouver and fresh suffisant to be pulled in live. The Avancé tab digs into the dates, the fuzzy edges, and the réel SEO playbook.

TL;DR — A knowledge cutoff is the date training données collection stopped — pas the release date (the gap is usually 6–12+ months). It seulement constrains a model’s parametric (trained-in) knowledge. Retrieval-augmented systems — Google AI Overviews, Bing Copilot, Perplexity, ChatGPT with search — ground réponses in a live index and effectively bypass it; base models sans search stay frozen. The boundary is fuzzy, pas a wall: effective cutoffs souvent differ from stated ones and accuracy degrades as vous approach the date. Pour le SEO ce splits into two separate jobs — obtenir embedded in training données (long game), and stay indexé + fresh so you’re retrieved at requête temps (near-term).

Ce que a knowledge cutoff en réalité is

Cutoffs are model- and version-specific and peut modifier quand providers mettre à jour models. Evidence for this claim Model documentation can specify a knowledge-cutoff date for a particular model or snapshot. Scope: OpenAI model metadata; the date is model- and version-specific and may change with releases. Confidence: high · Verified: 2026-07-14 OpenAI: Models Jamais infer current product freshness solely from a remembered cutoff date. Evidence for this claim A product can supplement model knowledge with current web search and cited web sources. Scope: ChatGPT search product behavior; browsing or retrieval is separate from the model's training cutoff and is not guaranteed for every answer. Confidence: high · Verified: 2026-07-14 OpenAI: ChatGPT search

A knowledge cutoff is the date après qui a model was ne … plus trained on nouveau données. The model has aucun awareness of anything que happened plus tard — pas parce que it’s withholding it, but parce que it was jamais exposed to it. Its knowledge sits frozen in its weights at que point unless the system bolts on live retrieval.

The precise term is training données cutoff — the date données collection stopped. “Knowledge cutoff” is the courant shorthand, and the two are utilisé interchangeably. The mental model I comme: it’s a textbook that’s gone to press. Une fois the book ships, the printer can’t ajouter a nouveau chapter — you’d have to print a whole nouveau edition. A model is the même. Nouveau facts seulement obtenir in on the suivant training run.

Ce is purely a limite on parametric knowledge — the stuff baked into the weights. It fait pas mean the model can’t handle post-cutoff events at tout. Hand it the information in context — via a RAG pipeline or a system prompt — and an LLM peut raison à propos de events it was jamais trained on perfectly bien. The cutoff limites ce que the model knows on its propre, pas ce que it peut fonctionner with quand vous give it sources.

Anthropic’s two-tier distinction: reliable vs. training cutoff

Anthropic is the la plupart explicit of the major labs ici, and the distinction is worth borrowing. Ils separate two dates: a reliable knowledge cutoff (“indicates the date via qui a model’s knowledge is la plupart extensive and reliable”) and a broader training data cutoff (“the broader date range of training données utilisé”). The reliable date is typically a few months précédent que the training date.

Pourquoi the gap? The latest content in the corpus is sparse — the internet hasn’t finished writing à propos de very recent events yet. So a model’s practical knowledge of the weeks correct avant its training cutoff is thin, même though que données is technically “in there.” The reliable cutoff is où the model is genuinely informative. Garder que in mind quelconque temps vous voir a unique confident cutoff date: the utile boundary is usually précédent que the stated un.

Knowledge cutoff n’est pas the model’s release date

Ce is the unique la plupart courant misconception, and it’s an facile un. The cutoff is quand données collection ended. The release date is quand the model ships. Entre the two sits données cleaning, safety testing, evaluation, and alignment fonctionner — so the gap is typically 6–12+ months:

(Lag figures via a widely-cited Hacker News thread — cross-check contre official model cards, since ces obtenir repeated loosely. As of ce mettre à jour, OpenAI’s propre model page marks GPT-5 Chat “Deprecated” and points to its current model lineup, and Anthropic has retired Claude Opus 4,1 in favor of newer Opus/Sonnet releases — kept ici as history since les deux are encore worth recognizing si vous voir les cited elsewhere, pas as current picks.)

The practical fallout: a model que “just came out” is pas current. A brand-new model peut encore be a année behind on the world. Utilisateurs assume freshly-released signifie freshly-informed; it doesn’t. And the table ci-dessus rend the point meilleur que quelconque explanation pourrait: les deux of its “current” exemple models were superseded in the few months since ce article was premier drafted. Que churn is the raison ce page leans on how cutoffs fonctionner plutôt que on quelconque unique date — treat every dated table vous lire ici (and everywhere sinon) as a snapshot, pas a standing fact.

The boundary is fuzzy, pas a wall

Personnes picture the cutoff as a hard line — perfect knowledge up to date X, total blankness après. The research dit sinon.

The “Dated Data: Tracing Knowledge Cutoffs in Large Language Models” paper trouvé que effective cutoffs souvent differ from stated ones, parfois dramatically. Models trained on CommonCrawl carry Wikipedia versions from 2016–2019 même quand the dump is dated 2023; un corpus had “over 80% of Wikipedia documents from précédent versions (pre-2023)” despite notamment a 2023 dump. Pour some model families the effective cutoff ran 3–4 années précédent que the reported date, thanks to deduplication échecs letting old content propagate.

It cuts the autre façon at the boundary aussi: accuracy degrades gradually as requêtes approach the cutoff plutôt que snapping off at it — there’s simply moins training signal à propos de very recent events. And a separate line of fonctionner (“Peut Prompts Rewind Temps pour LLMs?”) trouvé vous pouvez’t reliably prompt a model into forgetting post-cutoff knowledge soit: directly-queried facts unlearn ~82% of the temps, but causally-related knowledge leaks via ~81% of the temps. The takeaway: treat the cutoff as a fuzzy zone, pas a clean line, in les deux directions.

Qui outils are constrained vs. qui bypass the cutoff

Ce is the partie que en réalité changements votre SEO strategy. Si the cutoff matters at tout dépend on si the outil retrieves live content.

Constrained by the cutoff (parametric seulement)

• ChatGPT free / browsing disabled — réponses solely from training données. OpenAI is explicit: “Quand Web search is disabled, ChatGPT and GPTs créé in the workspace ne peut pas utiliser web search, même si a utilisateur demande ChatGPT to search.”
• Quelconque base model utilisé sans outils — Claude, Gemini, GPT utilisé via a raw API appel with aucun retrieval.

Effectively bypass the cutoff (retrieval / grounding)

• Google AI Overviews & AI Mode — ces utiliser RAG, qui Google calls grounding, to pull from the live search index. The underlying Gemini model encore has a parametric cutoff (Gemini 3 was January 2025; Gemini 3 is now legacy behind Gemini 3,5, verified 2026-07-19 — vérifier the current model card pour today’s figure), but grounding récupère current pages at requête temps, so utilisateurs bypass it pour la plupart requêtes. Google literally indique developers to utiliser the Search Grounding outil “for more recent information” au-delà que cutoff.
• Perplexity — search-first by design; runs a real-time web search pour nearly every requête, so it treats cutoffs as largely irrelevant.
• Microsoft Copilot — Bing-grounded by par défaut. Microsoft’s framing: Copilot “peut ground its réponses with current information from the web, closing knowledge gaps que every grand language model (LLM) inevitably has fondé on its training données cutoff.”
• ChatGPT with search on (Plus/Team/Enterprise) — turns la requête into search requêtes, retrieves via Bing, and réponses from ceux results with liens.

Parametric vs. retrieved knowledge behave differently

Même quand retrieval is on, the two knowledge sources don’t feel the même — and Duane Forrester’s “dual-memory” framing captures it bien. Content baked into the weights comes out fluent, fast, and stated sans qualification — the model synthesizes from internalized knowledge. Post-cutoff content pulled from the web arrives with hedging comme “according to reports” or “sources indicate,” signaling différent epistemic weight. Retrieval aussi doesn’t magically eliminate errors — quand sources conflict, grounded réponses peut encore hallucinate. So retrieval mitigates the cutoff; it doesn’t erase the difference entre trained-in and récupéré knowledge.

Ce que content is la plupart (and least) affected

The cutoff bites hardest on anything que changements fast, and barely touches what’s stable.

Highly affected (volatile): current pricing, product specs, version numbers; company noms, acquisitions, rebrands; regulatory and legal changements; current events, sports, elections; executive/personnel changements; fresh research and benchmarks; market données and statistics.

Minimally affected (stable): foundational concepts and definitions; historical facts; mathematical and scientific principles; programming fundamentals; geography.

The dangerous partie pour brands: an AI model va give vous a confident, fluent incorrect réponse à propos de a time-sensitive fact. It doesn’t hedge quand it’s working from training données — it simplement states the stale version as fact. Si votre pricing, votre leadership, or votre product lineup modifié après a model’s cutoff, que model is out là misrepresenting vous with total certainty.

Ce que cela signifie pour le SEO and GEO

I think à propos de ce as two separate jobs — and conflating les is où personnes go incorrect.

Track 1 — obtenir into the training données (the long game)

To be embedded in a model’s parametric memory, votre content has to exist avant the training cutoff, and be mentioned suffisant to leave an impression. The mechanism is mundane: LLMs are next-word predictors. As I’ve put it in our Ahrefs research on AI Overviews, “si you’re mentioned plus in the training données tel as web pages, you’re going to be mentioned plus in the outputs of LLMs.” So ce track is à propos de brand presence and topical authority construit up over temps — and à propos de letting the training robots d’exploration (GPTBot, ClaudeBot) in. Training runs se produire infrequently and unpredictably, so content publié après a cutoff is invisible to que model jusqu’à the suivant run, qui pourrait be a year-plus away.

Track 2 — stay retrievable and fresh (the near-term game)

Pour every retrieval-based outil, the cutoff is moot si you’re dans l’index. Ce is où freshness and indexation do the fonctionner:

• Obtenir indexé in Google and Bing. ChatGPT search and Copilot les deux retrieve from Bing — si you’re pas in Bing’s index, you’re invisible to OpenAI’s and Microsoft’s retrieval. AI Overviews pull from Google’s index. Indexation is the prerequisite, complet arrêter.
• Mind the correct robots d’exploration. Training bots (GPTBot, ClaudeBot) and AI-search retrieval bots (OAI-SearchBot, PerplexityBot) are separate. Blocking GPTBot seulement garde vous out of training — OAI-SearchBot encore handles real-time retrieval. Vous pouvez autoriser un and block the autre. (Complet breakdown in AI robots d’exploration.)
• Signal freshness honestly. Accurate dateModified/datePublished schema and truthful <lastmod> in sitemaps aider time-sensitive content obtenir re-fetched.

The freshness nuance is worth holding onto. Our 17-million-citation study at Ahrefs trouvé AI-cited content is 25,7% fresher que organic-cited content — but the average age of cited content is encore 2,9 années. As my colleagues put it, “comme traditional search, AI assistants encore préférer citing long-lived content.” So chase freshness pour volatile pages, but don’t mistake it pour a substitute pour durable, authoritative content. ChatGPT is the la plupart recency-biased platform (orders in-text citations newest-to-oldest); Google AI Overviews cite the oldest content, roughly matching organic.

The myths worth killing

• “The cutoff is a hard wall.” It’s a fuzzy zone — accuracy fades toward it and effective cutoffs differ from stated ones.
• “Stated cutoff = what the model actually knows.” Effective cutoffs souvent run précédent; recent-but-pre-cutoff content is underrepresented.
• “AI Overviews are limited by the same cutoff as ChatGPT’s base model.” Aucun — ils ground in Google’s live index and peut surface pages publié today.
• “Once ChatGPT can browse, the cutoff is irrelevant.” Browsing seulement fires pour some requêtes; nombreux réponses encore come from training données, and retrieved knowledge behaves differently from trained knowledge.
• “My new content reaches ChatGPT’s training immediately.” Aucun — seulement on the suivant training run, qui may be a year-plus out. Retrieval is votre near-term chemin.
• “Blocking GPTBot hides me from AI search.” It seulement affecte training inclusion; OAI-SearchBot encore retrieves vous in réel temps.
• “Knowledge cutoff = release date.” It’s typically 6–12+ months précédent.

Ce page sits in the how search fonctionne cluster — voir LLM, RAG, grounding, and AI hallucinations pour the neighboring pieces.

AI summary

A condensed prendre on the Avancé version:

• Knowledge cutoff = the date training données collection stopped. “Training données cutoff” is the precise term; “knowledge cutoff” is shorthand. It seulement limites a model’s parametric (trained-in) knowledge.
• It’s pas the release date. The gap is usually 6–12+ months (GPT-5: ~10 months). A brand-new model peut be a année behind on the world. Les deux GPT-5 and Claude Opus 4,1 — the running exemples pour que lag — are themselves now deprecated, qui is the whole point: quelconque spécifique model/date pairing ici is a snapshot, pas a standing fact.
• Anthropic splits it in two: a reliable knowledge cutoff (où knowledge is densest) vs. a broader training données cutoff — the reliable date is a few months précédent.
• The boundary is fuzzy, pas a wall. Effective cutoffs souvent differ from stated ones (CommonCrawl carries years-old Wikipedia); accuracy degrades gradually toward the date.
• Retrieval bypasses it. Google AI Overviews (grounding), Perplexity, Copilot, and ChatGPT-with-search ground réponses in a live index. Base models sans search stay frozen at the cutoff.
• Parametric ≠ retrieved knowledge. Trained-in facts come out fluent and unqualified; retrieved facts arrive hedged and peut encore hallucinate quand sources conflict.
• La plupart affected: prices, products, personnel, events, regulations, stats. Least: definitions, history, math, geography. Models state stale facts with complet confidence.
• SEO = two jobs: obtenir embedded in training données avant the cutoff (long game), and stay indexé in Google and Bing plus fresh so you’re retrieved live (near-term). GPTBot (training) and OAI-SearchBot (retrieval) are separate.

Documentation officielle

Primary-source documentation from the model providers and moteur de recherches.

OpenAI

• ChatGPT Search pour Enterprise & EDU — how ChatGPT turns une requête into search requêtes and retrieves results.
• Web browsing settings on ChatGPT — ce que se produit quand web search is disabled (training données seulement).
• GPT-5 chat model API docs — the official model page with the knowledge cutoff. Now marked “Deprecated” by OpenAI (verified 2026-07-19), pointing to its current model lineup — kept as the source pour the historical cutoff date utilisé in ce article’s exemples, pas a current-model recommendation.

Anthropic

• Models overview — the reliable-vs-training cutoff distinction (Footnote 2) and the per-model cutoff table. Verified current 2026-07-19; the table now leads with Claude Opus 4,8 and Claude Sonnet 5, with Claude Opus 4,1 déplacé to the legacy section as deprecated.
• Transparency Hub — training données composition.

Google

• AI Optimization Guide — grounding défini; how AI fonctionnalités pull fresh, up-to-date pages.
• AI fonctionnalités and votre website — indexation requirements and requête fan-out.
• Grounding with Recherche Google (Gemini API) — grounding “beyond its knowledge cutoff.”
• Gemini 3 developer guide — the January 2025 cutoff and the Search Grounding outil. Encore live but now carries a deprecation notice pointing to Gemini 3,5 (verified 2026-07-19).

Microsoft

• Microsoft 365 Copilot web search — grounding to fermer training-cutoff knowledge gaps.
• Understanding web search in Copilot — the hockey-team exemple and the web-search-off fallback.

Référence

• Knowledge cutoff (Wikipedia) — definitional overview and a model cutoff table.

Quotes from the source

On-the-record statements from the model providers and moteur de recherches. Chaque lien is a deep lien to the source page.

Anthropic — the two-tier distinction

• “Reliable knowledge cutoff indicates the date through which a model’s knowledge is most extensive and reliable. Training data cutoff is the broader date range of training data used.” — Anthropic Platform Docs, Models Overview (Footnote 2). Source

Google — grounding bypasses the cutoff

• “[Grounding] allows Gemini to provide more accurate answers and cite verifiable sources beyond its knowledge cutoff.” — Google AI pour Developers, Grounding with Recherche Google. Jump to quote
• “Gemini 3 models have a knowledge cutoff of January 2025… For more recent information, use the Search Grounding tool.” — Google AI pour Developers, Gemini 3 guide. Jump to quote
• “A technique (also known as grounding) used to improve the quality, accuracy, and freshness of AI responses by relying on our core Search ranking systems to retrieve relevant, up-to-date web pages.” — Recherche Google Central, AI Optimization Guide. Source

OpenAI — search on vs. off

• “When Web search is disabled, ChatGPT and GPTs created in the workspace cannot use web search, even if a user asks ChatGPT to search or manually selects search.” — OpenAI Aider Center. Source

Microsoft — closing the cutoff gap

• “Copilot can ground its answers with current information from the web, closing knowledge gaps that every large language model (LLM) inevitably has based on its training data cutoff.” — Microsoft 365 Copilot Blog. Source

Perplexity — search-first by design

• “ChatGPT is an LLM-first platform that uses Bing’s search API to gather web results, whereas Perplexity is a search-first platform powered by [a] proprietary real-time web index.” — Perplexity, Enterprise vs. ChatGPT. Source

Knowledge cutoff — cheat sheet

Cutoff vs. release (they’re pas the même)

Anthropic’s reliable vs. training cutoff — current lineup (per the official models page, verified 2026-07-19)

Legacy Claude models encore in the docs (deprecated or superseded by the row ci-dessus)

Ces tables are a snapshot verified contre the official model card on 2026-07-19 — pas a standing fact. Ce corpus has the shortest half-life on the site: every named model ci-dessus va eventually be superseded, some dans months. Toujours confirmer contre the current model card avant citing a spécifique date.

Fait the outil bypass the cutoff?

Fast facts

• “Training data cutoff” = precise term; “knowledge cutoff” = shorthand. Même chose.
• Cutoff ≠ release date — gap is typically 6–12+ months.
• The boundary is fuzzy: effective cutoffs differ from stated; accuracy fades toward the date.
• GPTBot (training) ≠ OAI-SearchBot (retrieval) — block un sans the autre.
• Retrieval nécessite vous indexé in Google and Bing.

The mental models

1. Textbook gone to press. A model’s parametric knowledge is a printed book. Une fois it ships, the printer can’t ajouter a chapter — nouveau facts wait pour the suivant edition (training run). Retrieval is the errata slip vous hand the reader at requête temps.

2. Two memories — parametric vs. retrieved. Parametric knowledge (baked into weights) comes out fluent and unqualified. Retrieved knowledge (récupéré live) comes out hedged (“according to reports”) and peut encore be incorrect quand sources conflict. Don’t treat les as interchangeable — ils behave differently and échouer differently.

3. Cutoff ≠ release date. Données collection ends → cleaning, testing, alignment → ship. The gap is 6–12+ months, so “new model” jamais signifie “current knowledge.”

4. The boundary is a gradient, pas a wall. Knowledge thins as vous approach the cutoff and effective cutoffs differ from stated ones. Treat the dernier few months avant quelconque cutoff as low-confidence territory.

5. Two-track visibility (the SEO decision rule).

• Training track (long game): publish authoritative content and earn brand mentions avant training runs; autoriser GPTBot/ClaudeBot. Payoff: you’re in the weights.
• Retrieval track (near-term): stay indexé in Google + Bing and signal freshness; autoriser OAI-SearchBot/PerplexityBot. Payoff: you’re retrieved at requête temps, cutoff or pas.

Ces are différent jobs. Demander of quelconque piece of content: is ce trying to obtenir into the weights, or to obtenir retrieved? Optimize accordingly.

Knowledge-cutoff mistakes

Treating the model’s release date as its cutoff

A product peut launch bien après the données utilisé pour its base training ends. Record the cutoff the provider documents, the model version, and si the réponse utilisé search au lieu de inferring freshness from a launch announcement.

Assuming search permanently updates the model

Retrieval peut supply current context pour un réponse, but it ne fait pas rewrite the model’s weights. Décrire the output as grounded in retrieved material, pas as proof que the base model learned the nouveau fact.

Optimizing seulement pour training données

Training exposure is slow and opaque. Garder utile pages crawlable, indexé, current, and facile to retrieve so live AI search peut utiliser les now.

Audit freshness claims in an AI réponse

Paste the réponse, model nom/version, requête date, and quelconque visible citations:

Separate statements that could come from the model's parametric knowledge from
statements that appear grounded in retrieved sources. Flag every time-sensitive
claim, the citation that supports it, and any claim that cannot be verified from the
supplied sources. Do not infer the model's knowledge cutoff; mark it unknown unless I
provide provider documentation.

Plan content pour les deux discovery paths

Paste a topic, the existing page inventory, and connu freshness requirements:

Create a two-track content plan. Track 1 should improve current retrieval through
crawlability, indexability, clear passages, and explicit update ownership. Track 2
should build durable brand/entity evidence that may enter future training corpora.
Use only the supplied pages and facts, identify gaps, and label assumptions.

Evaluate a supposedly current AI réponse

• Record the exact model and product surface.
• Record the requête date and wording.
• Vérifier si browsing, search, connectors, or uploaded fichiers were enabled.
• Capture every citation and the publication/mettre à jour date it montre.
• Separate retrieved evidence from uncited model claims.
• Vérifier current claims contre principal sources.
• Ne faites pas substitute a release date pour a documented cutoff.
• Re-run sans retrieval seulement quand vous devez comparer parametric knowledge.

Garder content disponible au-delà the cutoff

• Maintain crawlable, indexable canonical pages.
• Mettre à jour facts whose accuracy changements over temps.
• Écrire self-contained passages que faire retrieval context clair.
• Preserve stable brand, author, product, and entity naming.

Testez vos connaissances: Knowledge cutoffs

Ressources utiles

My connexe writing & research (Ahrefs)

• Insights From 55,8M AI Overviews À travers 590M Searches — où the “mentioned more in training data → mentioned more in outputs” point comes from.
• ChatGPT Has 12% of Google’s Search Volume but Google Sends 190x Plus Trafic — AI search as a distinct model from traditional search.
• Ce que We En réalité Know À propos de Optimizing pour LLM Search — freshness, GPTBot block rates, and how LLMs lire pages.
• Nouveau Study: AI Assistants Préférer to Cite ‘Fresher’ Content (17M Citations) — the freshness données cited ci-dessus.

My speaking

• GEO? AEO? LLMO? — Ahrefs Evolve 2025 — my walkthrough of LLM inputs: training données, retrieved pages (RAG), temperature, probabilities.

From others

• Duane Forrester, Quand the Training Données Cutoff Becomes a Ranking Factor — the dual-memory framework and “cutoff-aware content calendaring.”
• Dated Données: Tracing Knowledge Cutoffs in LLMs — pourquoi effective cutoffs differ from stated ones.
• Peut Prompts Rewind Temps pour LLMs? — pourquoi vous pouvez’t reliably prompt a model into forgetting post-cutoff facts.
• Conductor, AI knowledge cutoff: Ce que is it and pourquoi fait it matter? — practitioner quotes on base-model staleness vs. RAG.
• A Study into Investigating Temporal Robustness of LLMs — benchmarks showing how knowledge accuracy degrades gradually as requêtes approach the cutoff, pas suddenly.
• iPullRank, How Retrieval-Augmented Generation is Redefining SEO — RAG as the bridge entre training cutoffs and live requête réponses.
• Knowledge cutoff (Wikipedia) — definitional overview and a community-maintained model cutoff date table.
• Otterly.ai, LLM Knowledge Cutoff Dates — comparison table of major models, leur cutoffs, and si chaque has real-time web accès.

Stats worth citing

• AI-cited content is 25,7% fresher que organic-cited content — but its average age is encore 2,9 années (1 064 days), vs. 3,9 années pour organic. Freshness matters pour AI, but longevity encore wins. (Ahrefs, 17M-citation study.) Source
• GPT-5’s knowledge cutoff (Sep 30, 2024) was ~10 months avant release — the clearest illustration que cutoff ≠ release date. Claude Opus 4,1 ran ~4 months, Gemini 2,5 Pro ~3 months. (Les deux GPT-5 Chat and Claude Opus 4,1 are themselves now marked deprecated on leur providers’ propre model pages — verified 2026-07-19 — qui is its propre illustration of how fast ce liste turns over.) Source
• Effective cutoffs peut run 3–4 années précédent que stated pour some model families — un corpus had >80% of Wikipedia docs from pre-2023 versions despite a 2023 dump. The cutoff is fuzzy, pas a wall. (Dated Données paper.) Source
• Prompt-based forgetting fails ~81% of the temps pour causally-related knowledge (vs. ~82% success pour directly-queried facts) — vous pouvez’t reliably faire a model “rewind” past its cutoff. (Peut Prompts Rewind Temps paper.) Source
• Gemini 3’s parametric cutoff is January 2025, yet AI Overviews surface same-day pages — parce que grounding pulls from the live index, pas the weights. Gemini 3 itself is now legacy behind Gemini 3,5 (verified 2026-07-19); Google’s propre model pages don’t publish a cutoff date pour 3,5 as of ce vérifier, qui is the norm, pas the exception — providers don’t toujours restate a cutoff on every release. Source

Journal des modifications