Content freshness and AI answers: how often models and indexes actually update

"How long until our new pricing shows up in ChatGPT?" is one of the most common questions in GEO — and it has no single answer, because there is no single system. The answer a buyer sees is assembled from layers that update on completely different schedules: the model's frozen training knowledge, a retrieval index refreshed on search-engine timescales, and crawlers hitting your site around the clock.

Teams that don't distinguish the layers make expensive mistakes in both directions. They panic-rewrite pages because an offline model repeated something from 2024 (which no page edit can fix), or they shrug at a wrong price in a cited answer (which a page edit fixes within weeks). Once you can name the layer, you can predict the lag — and decide whether the right response is a content sprint, a robots.txt fix, or patience.

The three clocks

Every generative answer about your brand draws on some mix of three sources, each with its own refresh rate.

Clock 1 — Training cutoffs: months to years

A model's parametric knowledge is frozen at its training cutoff. What it "knows" about your brand offline — your category, reputation, flagship products — was formed from the web as it existed months or years ago, and doesn't budge until a new model ships. Training corpora are dominated by broad web crawls accumulated over years, which means brand associations are built from accumulated, repeated descriptions across many sites — not from your latest release notes. This clock rewards patience: consistent positioning, sustained third-party corroboration, the same claims echoed for quarters. It also means a rebrand or repositioning will haunt you in offline answers for a model generation or more.

Clock 2 — Retrieval indexes: days to weeks

When ChatGPT searches the web, Perplexity answers with citations, or Google composes an AI Overview, the model is fed passages from a search index built on recent crawls.² These indexes refresh the way search indexes always have: important, frequently-changing pages get recrawled often; obscure ones can wait weeks. This is the layer where a content change actually lands quickly — publish a sharper, better-evidenced page today and it can be feeding answers within days.

Clock 3 — The crawlers: continuous

Underneath both sits the plumbing: the bots. OpenAI operates GPTBot (training data), OAI-SearchBot (search index) and ChatGPT-User (live fetches on behalf of a user's request);¹ Anthropic runs ClaudeBot;³ Google's crawlers feed both search and, via the Google-Extended token, AI training.² Cloudflare's network data shows AI-crawler traffic has grown into a significant share of all bot activity.⁴ The practical point: the pipes are always on. If your server logs show GPTBot and ClaudeBot visiting weekly, your freshness bottleneck is not the crawl — it's the index update and, further up, the model itself.

What the clocks mean for strategy

The two loops reward different work, and confusing them wastes quarters:

Most teams over-invest in one loop. Pure content shops churn out fresh pages but never earn the repeated third-party descriptions that shape the next model's priors. Pure PR shops build reputation that models trained last year reflect beautifully — while their pricing page misleads every engine that browses.

A quick diagnostic for which clock is your bottleneck: ask the same buyer question in an engine with browsing on, then in a mode or engine that answers from memory. If the browsing answer gets you right and the offline answer doesn't, your problem is the slow clock — keep publishing and corroborating, and time will catch up. If even the browsing answer misses or misquotes you, the fast loop is broken somewhere you can actually fix this quarter: the page is thin, the fact isn't crawlable, or the crawler never visits. Check in that order — content, rendering, logs — because each is cheaper to rule out than the next.

Date your content honestly

Freshness only works if engines can tell how fresh you are. Two mechanisms, both cheap:

• MentionBeat dates. A "Last updated" line near the top of the page, in plain text. Retrieval systems and re-rankers use recency signals, and answer engines increasingly surface dates to users.
• Structured dates. datePublished and dateModified in your Article/Product schema, kept truthful.⁵ This is the machine-readable version of the same promise.

The temptation to game this is real and self-defeating. Bumping dateModified without changing content is the freshness equivalent of keyword stuffing: detectable (diff two crawls and nothing changed), and corrosive — you're training engines to distrust your dates. Re-date when you actually revise: numbers rechecked, prices confirmed, sections rewritten. A page that says "reviewed January 2026" and means it is a better retrieval candidate than one stamped yesterday with a 2023 screenshot in it.

A refresh cadence that matches the clocks

Given the update speeds above, a sane default cadence looks like this:

1. Quarterly — money pages. Product, pricing, comparison and top category pages get a real review every quarter: verify every number, price, spec and screenshot; update the changelog; bump dateModified. Quarterly matches the fast loop's memory — old crawls age out of answers within weeks of the refresh.
2. Semi-annually — evergreen guides. Your explainers and how-tos need less churn, but a twice-yearly pass keeps examples, versions and statistics current enough to stay quotable.
3. Immediately — anything that invalidates an answer. Price changes, discontinued SKUs, renamed products. Ship the page update the same week as the change, because every engine that browses is now a channel repeating whatever your site said last crawl.
4. Always-on — corroboration. The slow loop doesn't have a refresh date. Third-party mentions, reviews and consistent descriptions accumulate into the next training run whether you're paying attention or not.

Stale answers: why they hurt and how to catch them

The nastiest freshness failures aren't invisibility — they're confidently wrong answers. A buyer asks for pricing and gets your 2024 tiers. An assistant recommends a product line you discontinued in the fall. A comparison quotes a spec you improved two releases ago, making you look worse than you are. Because AI answers strip the "as of" context a webpage carries implicitly, a stale fact sounds exactly as authoritative as a fresh one — and with users clicking through to sources on only about 1% of AI-summary visits, almost nobody checks.⁶

You can't fix what you never observe. Detection is a monitoring problem:

• Include prompts in your test suite that elicit facts, not just mentions: "how much does X cost," "what's the accuracy of X," "does X still sell Y."
• Re-run them on a schedule — stale facts reappear as indexes and models update at different times, so a fact you fixed in ChatGPT can persist in Gemini for weeks.
• When a stale fact shows up, trace the layer: if the engine cites a source, it's the fast loop — fix or update the cited page (often not yours). If there's no citation and no browsing, it's parametric — you can't patch the model, but you can ensure every retrievable page states the current fact clearly, so browsing-enabled answers correct it.

This kind of scheduled, fact-checking sampling is tedious by hand — it's exactly the loop a platform like MentionBeat automates — but even a monthly spreadsheet run beats finding out from a prospect.

Frequently asked questions