You can refresh old content with AI in an afternoon and publish a page that is worse than the one you started with. The model updates a price from memory, smooths a customer quote into something the customer never said, and rewrites a paragraph that was ranking fine. Each of those failures starts with the same request: the whole page plus a vague instruction such as “update this for 2026”.

This guide shows how to set an evidence bar and a human review step for every refreshed page, with five prompts that keep AI on the jobs it does well and a rule for honest “Last updated” dates. Choosing which pages deserve a refresh in the first place is a separate decision, covered in Prickly Blog’s content refresh workflow for AI-assisted blogs.

Summary

AI earns its place in a refresh on three jobs: listing claims that may have aged, redrafting them from source text you paste in, and summarising every change. The URL, the search intent, quotes, first-hand material and partner links stay frozen. Every changed claim must clear the evidence bar for its type (a price needs the vendor’s pricing page, a statistic needs the study itself), and every refreshed page goes through the same human review step, the refresh sign-off, before it goes live.

What should AI do in a refresh, and what stays with a person?

Google’s guidance on using generative AI content gives the reason in one line: generative models predict a likely sequence of words from their training data instead of retrieving facts, so their output can contain errors. Google asks site owners to fact-check AI output by hand before publishing, and it names the places that review has to reach besides the body text: title elements, meta descriptions, structured data and image alt text.

The split looks like this.

Refresh taskWhat AI doesWhat a person does
Find dated claimsLists every price, version, plan name and statistic with its sentenceConfirms the list is complete
Gather sourcesNothing: the model’s memory is exactly what you are replacingOpens each source page on the refresh day
Redraft changed sentencesRewrites from pasted source text onlyEdits for voice and checks every number
Spot gaps against current searchesCompares headings with the queries you paste from Search ConsoleDecides which gaps belong on this page
Change summaryLists every difference between old and newReads the list next to the comparison
PublishNeverSigns off and presses Update

Which parts of an old page must AI never touch?

Freeze these before the first prompt. Wrap them in [FROZEN] markers in the text you paste, or leave them out of the model’s input altogether.

  • The URL and the promise of the H1. A refresh that changes what the page is about puts a new page on an old address, and every link pointing at it now promises something else.
  • Direct quotes. Models tidy quotes, and a tidied quote is a misquote.
  • First-hand material: your own screenshots, figures from your own accounts, photos, results credited to a named person. AI cannot add to them and must not reword them.
  • Author and reviewer lines. Credentials belong to people, and a model will happily upgrade a job title.
  • Partner links and their anchor text. Those sentences were agreed with another site. A changed anchor or a dropped link breaks the agreement.
  • Legal, medical and financial statements that a qualified person approved.
  • Structured data types and the canonical tag. A person updates the values inside them after the text is final.

One more limit concerns intent. Google’s spam policies describe scaled content abuse as many pages generated to manipulate rankings instead of helping people, and list automated transformations such as synonymizing among the examples. An AI pass that rewords a page without adding a single new fact gives readers nothing, so the prompts below forbid rewording outside the sentences you mark.

How do you prepare a page before the first prompt?

  1. Copy the live text into a Google Doc named with the URL and the date. This is the old file for the comparison at the end.
  2. Write down the SEO title and meta description as they are now. They live in your SEO plugin, outside the article text you paste.
  3. Mark the sentences. Tag every sentence that needs checking with [UPDATE] and every frozen passage with [FROZEN]. The aged-claim prompt below does most of the [UPDATE] tagging for you.
  4. Collect sources. For every tagged claim, open the page that clears the evidence bar for its type (the table further down) and paste the relevant paragraph into a sources file with its URL and the date you opened it.
  5. Export current queries. In Search Console, filter the Performance report to the page and copy the top 20 queries. They show what searchers ask today.

Budget 20 to 40 minutes for a 2,000-word post. This step decides whether the model works from facts or from memory.

Which prompts work when you refresh old content with AI?

Each prompt does one job and refuses to guess. Two habits from Anthropic’s guide to reducing hallucinations run through all five: give the model explicit permission to say it does not know, and make it back every claim with a quote from the material you supplied, deleting any claim it cannot back. Paste them into whichever chat model your team uses.

1. Aged-claim finder

“Below is a published article inside <old> tags. List every claim that contains a date, price, plan name, version number, statistic, feature, regulation or named product. Return a table with four columns: the exact sentence, the claim, why it may have changed since publication, and the type of source that would confirm it today. Do not rewrite anything. If you cannot tell whether a claim can age, write UNKNOWN in the third column.”

2. Source-bound rewrite

“Rewrite only the sentences marked [UPDATE], using only the facts inside <sources>. After each rewritten sentence, quote the source sentence that supports it, with its URL. If no source sentence supports a claim, delete the claim and leave [] where it was. Keep every passage marked [FROZEN] exactly as it is. Match the tone of the surrounding text.”

3. Query gap check

“Here are the article’s H2 headings and the top 20 search queries that lead to it. List the queries the article does not answer yet. For each one, say whether it fits this article or belongs to a different topic. Do not draft any text.”

4. Opening rewrite

“Rewrite the first paragraph so that its first two sentences answer the question in the H1 directly. Use only facts already in the article or inside <sources>. Keep it under 70 words.”

5. Change summary

“Compare <old> and <new>. List every difference in a table: location, old text, new text, type (fact, link, wording, new section, removed section). Flag any change to a passage marked [FROZEN] and any new claim that has no source quote.”

Run prompt 5 in a fresh chat, so the model reads both versions cold instead of defending the rewrite it just produced.

What is the evidence bar for every updated claim?

The evidence bar is the rule a claim must clear before it goes into the refreshed page. Write it down once and apply it to every page. Each claim type has one kind of source that counts.

Claim typeClears the barFails the bar
Price or planThe vendor’s own pricing page, opened on the refresh day and linked next to the numberA review site, a comparison post, the model’s memory
StatisticThe study or report itself, with publisher, year and sample sizeA blog post that quotes the study without linking it
Product featureThe vendor’s documentation or changelogA reseller’s feature list or a two-year-old launch post
Regulation or legal ruleThe official text from the legislator or regulator, in its current versionA law firm’s summary, a news article about the draft
QuoteThe original interview, talk or post, word for wordA quote collection site, a paraphrase in someone else’s article
Google behaviourGoogle Search Central or Search Console HelpAn SEO forum thread
General claim without a numberYour own reasoning, written as adviceA sentence that sounds like a fact and cites nobody

Add one fallback rule: a claim that finds no qualifying source within ten minutes gets turned into plain advice or removed. Keeping it “because it was there last time” is how a page ends up quoting a 2023 price in 2026.

Why so strict with AI output? OpenAI’s research post Why language models hallucinate (5 September 2025) argues that standard training and evaluation reward guessing over admitting uncertainty. Its own SimpleQA example shows the trade: gpt-5-thinking-mini declined to answer 52% of questions and got 26% wrong, while OpenAI o4-mini declined 1% and got 75% wrong, with similar accuracy (22% and 24%). A model that may answer UNKNOWN, as in the prompts above, hands your reviewer far fewer confident errors to catch.

What does the human review step look like for every refreshed page?

Every refreshed page goes through the same named step, the refresh sign-off, whether AI changed three sentences or thirty. The reviewer works from a comparison of old and new, because a clean new version hides what moved.

  1. Open the comparison. In Google Docs, Tools > Compare documents turns every difference between the old and new file into suggested edits in a third document. In WordPress, the revisions screen compares any two saved versions of a post.
  2. Check every changed claim against the evidence bar. Click each source link, find the number or sentence on that page, and confirm the wording matches.
  3. Confirm the frozen parts are untouched: quotes, partner links, author lines, first-hand figures.
  4. Review what Google names besides body text: the title, meta description, structured data and image alt text.
  5. Click every link in the changed sections. Sources move too, and an updated claim that links to a 404 has lost its evidence.
  6. Read the first screen aloud. If it no longer answers the H1 within two sentences, it goes back to the editor.
  7. Decide the date with the rule in the next section.
  8. Sign off in the refresh log with initials, date and the number of claims checked.

Example time budget for a 1,500-word post where AI changed 12 claims: 12 claims at 3 minutes each is 36 minutes, plus 10 minutes on the comparison, 5 on metadata and 5 on links, so 56 minutes. If that feels slow, set it against the cost of a wrong price on a page that sells.

Reviewer roles and a fact-check log for a whole SEO team are covered in human review for AI SEO articles.

When does a refreshed page get a new “Last updated” date?

The rule fits on one line: move the “Last updated” date only when a claim, a price or a section really changed. A fixed typo or a swapped image keeps the old date.

Two Google pages back it up. The helpful content guide includes, among its self-assessment questions, whether you change page dates to make pages seem fresh when the content has not really changed. Google’s byline date documentation asks for a visible date labelled “Published” or “Last updated”, and for that date to match the datePublished and dateModified values in the page’s structured data.

So the reviewer checks three things at sign-off: the change summary contains at least one fact, price or section row; the visible date and dateModified show the same day; and nothing on the page claims a date in the future. Sitemaps, recrawl requests and recheck dates belong to the Prickly Blog workflow linked in the introduction.

What goes wrong when AI refreshes a page unsupervised?

Each failure below has a check in the process above that catches it.

What happensHow it looks on the pageThe check that catches it
Invented “updated” statistics“Recent data shows a sharp rise” with no linkEvidence bar: no study, no number
Price from the wrong plan or billing periodA monthly price shown as the yearly oneReviewer finds the number on the vendor page
Quotes rewordedA customer line that reads smoother than the originalFrozen parts check in the comparison
Synonym swapsParagraphs reworded with no new factsChange summary shows only wording rows
Lost linksA partner or internal link missing after the rewriteLink count before and after
Date moved, facts unchangedA new “Last updated” line on the same old claimsDate rule at sign-off
Scope creepA new section that answers a sibling page’s questionQuery gap check marked it a different topic

Links deserve their own pass during a refresh. The checks for AI-suggested internal links are in internal linking for AI content.

Worked example: one AI-assisted refresh, start to finish

Example: a 1,900-word comparison post from March 2025 that lists eight tools with prices. The figures are illustrative; the arithmetic is the point.

StepWhat happenedTime
PrepOld text saved, frozen parts marked, top 20 queries exported25 min
Prompt 117 claims listed: 8 prices, 4 plan limits, 3 features, 2 statistics3 min
Sources13 pages reopened (8 pricing pages, 3 help centres, 2 studies) at about 4 minutes each52 min
Prompt 29 sentences redrafted; one statistic had no current source and came back as []5 min
EditsEditor rewrote 3 of the 9 for voice and cut the unsupported statistic15 min
Prompts 3 and 52 new FAQ questions from current queries; change summary of 14 rows10 min
Refresh sign-offComparison, evidence bar, metadata, links, date decision40 min
Total17 claims checked, 8 updated, 1 removed, new date shown150 min

Two and a half hours (25 + 3 + 52 + 5 + 15 + 10 + 40 minutes), and every changed claim now has a source opened that day and a reviewer’s initials next to it. Notice where the time went: more than a third of it on reopening sources, which is the one job AI cannot do for you.

A refresh borrows from review, editing and linking, and each of those has its own guide.

Related

human review for AI SEO articles: reviewer roles and a fact-check log for the whole team

Related

AI SEO editing checklist: the line-by-line edit once the facts are settled

Related

internal linking for AI content: how to check the links an AI tool suggests during the refresh

Related

AI content quality score: score the refreshed page against the same rubric as a new one

FAQ

Can I just ask a chat model to “update this article for 2026”?

That request leaves the model to fill gaps from memory. Google’s generative AI guidance notes that models predict likely words instead of retrieving facts, so a blanket update produces confident numbers with nothing behind them. Use the narrow prompts on this page instead, with sources pasted in.

Should a refreshed page say that AI helped update it?

Google’s generative AI guidance suggests giving readers context on how content was created where that context helps them. A short update note listing what changed, when, and who reviewed it covers that and tells readers more than a bare date does.

Does rewording an old post with AI count as a refresh?

Rewording alone adds nothing new for readers, and Google’s spam policies list automated synonymizing among the examples of scaled content abuse. A refresh changes facts, examples, links or coverage; any wording changes ride along with those.

What if the AI cannot find a source for a claim?

Then the system is working. The source-bound prompt deletes the claim and leaves [] in its place, a method taken from Anthropic’s hallucination guide. The editor decides whether to find a qualifying source by hand, turn the sentence into plain advice, or drop it.

Which parts of an AI refresh can run without a person?

Finding aged claims, drafting the change summary and checking that links still load can run unattended. Choosing sources, approving changed claims, deciding the date and pressing Update stay with a named reviewer on every page.

Which page should you refresh first?

Pick one page that already earns clicks and quotes prices or statistics. Run the aged-claim finder on it today, open the sources it lists, and put the page through the refresh sign-off before you touch a second one. The first run shows how long your evidence bar really takes.