Updating your blog posts does not fix stale data. We scanned 6,751 data claims across 938 SaaS blog posts from 45 domains, ran every flagged claim through a temporal verifier with a post-intent classifier, and looked again at the question the April study asked at one-seventh the sample size: does refreshing a post reduce the staleness of the data inside it. Posts that had been updated since publication carried 5.3% of their claims at verified staleness. Posts never touched: 3.9%. A 1.4-point gap — running in the wrong direction. Refreshed posts are statistically more stale than untouched ones (z = 2.45, p = 0.011).
The finding lands harder than the April study did. If refreshes restored freshness, the updated cohort should match or beat the untouched one. It doesn't. The most likely explanation is selection: teams revisit older, higher-traffic posts that carry more stale claims to begin with — and the refresh doesn't reach the claims inside. The April study used calendar heuristics; any claim referencing data more than six months old was flagged. This one keeps the calendar check, adds a temporal-verifier pass with post-level intent, and surfaces a third verdict (needs_review) for the 12.5% of disputed claims where anchoring evidence was weak. The 4.8% headline reflects "verified stale," not "old-looking timestamp." See the methodology change disclosure below.
The trend across post age remains highly significant (Cochran-Armitage z = 8.31, p < 0.0001). Claim type predicts decay (chi-square p < 0.0001). And 4.7% of every cited source URL in the corpus returns a 404 or 410. The picture is sharper than April reported: refreshes touch the wrapper, claims age underneath, the cohort that gets refreshed ends up carrying more stale claims than the cohort no one touches, and most teams have no way to see it happen.
Refreshing a Post Doesn't Reach the Claims Inside It
Five-and-a-half times the sample of the April study, and the gap between refreshed and untouched posts now runs in the opposite direction from what a refresh playbook would predict.
The 413 posts in the v3.1 dataset that had been updated since publication carried 5.3% verified-stale claims [95% CI: 4.5%, 6.1%]. The 355 posts never touched carried 3.9% [95% CI: 3.2%, 4.7%]. The two-proportion z-test returns z = 2.45, p = 0.011 — updated posts are statistically more stale than untouched ones, with a Cohen's h of 0.067. Small but real. Selection is the most plausible mechanism: teams refresh the older, riskier posts. The refresh doesn't reach the claims inside, so the underlying staleness shows up in the cohort that gets revisited.
A gap that runs the wrong direction is freshness theater with the volume turned up. The refresh reached the page wrapper. The claims inside aged at the same rate as the cohort no one touched — and then some.
A refresh rewrites the introduction. Swaps a screenshot. Revises the CTA. Stamps a date on the title. The data claims sit in the middle of the post, in sentences that read fine on a skim. A percentage from 2023. A benchmark that shifted. A competitor feature that no longer exists. These survive every refresh cycle because no one scans for them.
Teams measure refresh cadence. They track "last updated" dates. They report on posts touched per quarter. Those metrics measure editorial activity. Whether the numbers inside the post are still true requires a different kind of audit entirely.
The 413 updated posts came from teams that cared enough to revisit them. The claims inside carried a higher staleness rate than the posts no one had touched.
Stale Data Holds Steady for Two Years, Then Doubles
For the first two years, staleness creeps. It accelerates in year three.
We segmented 938 posts into four age bands and scored every claim. The decay curve has the same shape as the April study and now passes a pre-registered trend test.
Band D (0–12 months): 2.6% verified-stale claims [2.0%, 3.3%]. Band A (12–18 months): 4.9% [3.9%, 6.2%]. Band B (18–24 months): 4.2% [3.3%, 5.4%]. Band C (24–36 months): 8.9% [7.6%, 10.4%]. The Cochran-Armitage trend test across the four ordered bands returns z = 8.31, p < 0.0001 — a clean monotonic increase with a statistically meaningful slope.
The 24-month mark is still the inflection point. The rate roughly doubles between Band B and Band C, and it does so in every domain category we ran the test on.
The 24-Month Cliff
Two forces converge at the 24-month mark.
The underlying data sources have had two full years to shift. Industry benchmarks reset. Vendor pricing changes. Software features get deprecated or overhauled. Meanwhile, the post has fallen far enough down the editorial priority list that even teams with refresh cycles have stopped revisiting it. The claims were already aging. The oversight disappears at the same moment.
Any post older than 24 months is high-risk for stale data. The hidden cost of outdated charts compounds with every month past that threshold.
Source Citations Decay Faster Than Statistical Claims
Claim type predicts decay speed more than domain, category, or post length. The chi-square test of independence across the four production claim types returns p < 0.0001.
A post states "73% of marketers report using AI for content creation." Statistical claim. Another says "As of Q2 2024, HubSpot offers a free CRM tier." Temporal claim. They go stale on different clocks.
Source citations decay fastest. At Band D, 7.3% of cited-source claims are already verified stale. At Band C, that climbs to 18.0%. Statistical claims move from 2.3% to 7.2% across the same bands — slower at the start, but a 3.1x climb. Temporal claims jump from 0.5% in Band D to 12.1% in Band C, the steepest relative curve in the set. Comparative claims stay flatter than the others; they age, but they don't accelerate the same way.
That maps directly to the cost of stale blog data. A post dense with statistical or source-citation claims carries more maintenance liability per word than a post built on comparative arguments. Triage by claim type. Source citations first, because they fail loudly when the link breaks. Statistical claims second. Temporal claims third, with the caveat that most "Q2 2024" references are recoverable by simply updating the year, while statistical claims may require re-running the underlying analysis.
Cited Sources Are Going Dark Faster Than the Claims Themselves
While we had every claim and every source URL in a database, we ran a second check the April study didn't: what proportion of the cited sources are still reachable?
We sent HTTP HEAD requests against every unique source URL in the corpus — 1,935 of them, across 787 hostnames — and classified each by reachability. Then we fetched the alive ones for publish and modified dates.
Results, claim-weighted:
- 4.7% of cited source URLs return 404 or 410 [4.1%, 5.3%]. The link is gone. The claim cites a page that no longer exists.
- 27.7% of claims in the corpus have neither a named source nor a URL [26.6%, 28.7%]. Not an unverified claim — an unverifiable one.
- Among the alive sources where the page exposed dates: 21.8% pointed to data that is itself 24+ months old (the cited source is stale, even if the citing post is fresh).
Pair those numbers with the staleness curve and the failure mode becomes a system. A post cites a source. The source page disappears. The claim still reads fine in the body, because the prose around it doesn't care that the link 404s. The refresh cycle never visits the link. The post continues to rank, continues to convert, continues to make the claim — pointing at a URL that returned a server error six months ago.
That is the kind of thing that breaks downstream. Someone clicks the citation in a board deck. Someone runs an AI-summarized search that follows the citation. Someone files an editorial fact-check. The brand pays for the broken link, not the publisher of the source that went dark.
Stale Data Costs $1,000 to $160,000 Per Year
The cost range hasn't moved much, but the math underneath has tightened again. v3.2b's post-intent classifier pulled another 76 claims out of the stale bucket — 39 flipped to fresh, 21 went to a new needs_review surface, and the rest dropped via v3.2a's methodology short-circuit. The cost per correction is unchanged, and the bulk of the work is still finding the claims, not fixing them.
The cost model assumes 0.5 hours per claim for review and correction at the auditor's hourly rate. That estimate is conservative. Posts with unattributed claims (27.7% of the corpus) take longer because there is no source to check against — the auditor has to find the original data point from scratch.
The Variables That Move the Number
Blog size sets the floor. Average post age determines the staleness multiplier. Audit frequency swings the widest, and it is the variable most teams have never formalized.
The typical scenario: a 50-post SaaS blog with an average post age of 18 months, audited annually at $75 per hour. Roughly $2,000 per year in claim-maintenance debt under v3.2b's classifier-tightened rate. The April study estimated $6,000 for the same configuration; the v3.1 verifier brought that to $2,400; v3.2b's post-intent classifier shaved another $400 off. The gap is measurement, not improvement. Each tightening is significant enough to justify a line item. None of them are large enough to trigger one.
44% of posts in the dataset had been updated since publication. Their staleness rate did not improve. The question is what happens during that update.
The answer determines whether the cost model below applies to your team or overstates your correction rate.
The cost model above assumes every claim is found during the audit. That assumption depends entirely on how deep the review goes. If the update cycle stops at structural edits and never reaches the data layer, the budget spent on refreshes produces zero claim corrections. The 14.9% staleness rate in updated posts reflects audits that target the wrong layer.
LiquiChart's content debt estimator runs the same cost model on your numbers.
Enter your numbers. The defaults below reflect the v3.1 scan medians.
If the result is worth a conversation, generate a shareable link and send it to whoever owns your content budget.
Methodology Update
The 4.8% headline in this post replaces the 18.4% from the April version and the 5.9% interim figure from v3.1. That is not a finding about content getting better. It is a finding about measurement getting tighter — twice.
The April study (138 posts, 868 claims) flagged any claim referencing data more than six months old as stale. A calendar heuristic. The v3.1 study (938 posts, 6,751 claims) keeps the same calendar check and adds a temporal-verifier pass: an LLM that asks, given the context around the claim, whether the underlying data is actually outdated. The verifier short-circuits event dates, historical baselines, annual-cadence references, and stable-fact comparisons — the categories the April study was over-flagging.
Across 938 posts, the v3.1 verifier reviewed 656 calendar-flagged claims. It confirmed 380 as actually stale and rejected 276 (42% of its review pool) as false positives. If you back-applied the April study's calendar-only methodology to the v3.1 data, the rate would be ~9.8% — closer to April's 18.4% but still lower, with the residual gap coming from the broader domain mix in v3.1.
v3.2b adds a post-intent classifier on top of the v3.1 verifier. When a post's title contains a past year (e.g., "YouTube Demographics & Data to Know in 2023"), the classifier decides whether the post is an archival snapshot (anchored to that year) or a maintained reference (annual page the publisher refreshes) — and a 28-cell verdict matrix routes the claim accordingly. Anchored claims get treated as fresh. Genuinely stale claims stay stale. Ambiguous claims surface in a new third verdict — needs_review — instead of being forced into one bucket or the other. Across 168 candidate claims from 50 archival-or-maintained posts, 39 flipped to fresh, 21 went to needs_review, and 108 stayed stale. The net effect is a 1.1-point drop in the overall stale-claim rate (5.9% → 4.8%) and a 2.5-point drop in the Band-C rate (11.4% → 8.9%) — the past-year-title posts were concentrated in the older bands.
Five deltas vs the pre-registered v3 methodology are disclosed in full at scripts/staleness-study-v3.1/METHODOLOGY.md, with v3.2a (methodology-metadata short-circuit) and v3.2b (post-intent classifier + verdict matrix + needs_review) addenda alongside it. The most consequential v3-level call: claim types narrowed from six to four (aligning to LiquiChart's production taxonomy), and source URL verification ran as a standalone post-scan pass rather than inline. The §6.2 inter-rater reliability study is single-annotator, not two-rater, and is labeled as such in the validation appendix.
The April study's posts have not been retracted. They have been annotated with a supersession notice pointing here. Future research updates will land as supplements to this post, with the same v3.1+ measurement protocol, so the headline can be tracked across time without redefinition.
The Debt Is Already on the Books
Content debt stays off the quarterly plan. It stays off headcount. It shows up as ad hoc requests, late-night fixes before a sales call, and the slow realization that a third of the claims in your oldest posts contain numbers you would not put in a slide deck today.
80% of SaaS blog posts contain data claims. The average post carries 7.2 of them at the v3.1+ scale. After 24 months, 8.9% are verified stale. Multiply that across a blog of any size and the liability is already running. Add the 4.7% dead-source rate and the 27.7% of claims with no traceable source, and a meaningful chunk of every refresh you ship is pointing readers at links that 404 or sources that don't exist.
Page-level refreshes leave it untouched. The claims need claim-level detection, and that requires living content infrastructure built around the unit that ages: the individual claim.
That is what living content changes. Maintenance targets the claim.
Pick your highest-traffic post and see how many claims are at risk.
Every quarter your content budget ignores claim-level maintenance, the 24-month cliff gets closer for another cohort of posts. The debt is already on the books. The only question is whether you price it before or after it prices you.