If you are asking how do you refresh stale content for AI visibility, you do not need another publishing calendar. You need a pass that replaces the sentence an answer engine can still lift. Evergreen pages explain why some URLs should stay alive. This page is the SOP: audit the extractable unit, update the evidence, restructure the answer block, then recheck citations.
Stale content is not the same as a dead URL. A post can still rank, still get traffic, and still feed ChatGPT, Perplexity, or an AI Overview last year’s number. That is the gap. Search Console will not always show it. Prompt logs will. Treat refresh as living content work, not as “write a new blog post because the old one feels tired.”
What “stale” means when AI answers are the goal
For traditional SEO, decay often looks like falling impressions, a slipping average position, or a title that no longer matches the SERP. For AI visibility, decay looks like a model quoting a claim that is no longer true, or skipping your URL because a competitor’s sentence is newer and easier to lift. You can have both problems on the same page. You can also have only the second.
A useful test: copy the one paragraph you would be willing to see cited tomorrow. Check the date, the product name, the statistic, and the process. If any of those would embarrass you in a sales call, the page is stale for AI even if Google still likes the URL.
This is different from living content versus evergreen content. That comparison is about the operating model. This SOP assumes you already decided the URL should live. Now you have to keep the citeable unit honest.
When to refresh versus when to leave the URL alone
Refresh when the question is the same and the answer aged. Leave it when the extractable unit is still true and the page still matches intent. Rewrite or merge when the question changed, the product changed, or two URLs now fight for the same prompt.
- Refresh: the FAQ still matches GSC phrasing, but the number, vendor, or step list is old.
- Leave: definition is still accurate, examples are still representative, and prompt logs still show your current sentence.
- New URL: the searcher wants a different job (formats versus citations, evergreen versus refresh SOP).
- Merge: two posts answer the same prompt with two wordings and split the quote.
Do not refresh because a quarter ended. Do not skip refresh because the page “still ranks.” Rankings and citations have different half-lives. A 2024 sample size in a 2026 answer is a trust problem, not a traffic problem.
The refresh SOP, in order
Use this sequence on one URL at a time. Parallel refreshes feel productive and make measurement impossible. Pick the pages that already get AI impressions, already get cited with an old claim, or sit on a living-content hub with zero CTR for a how-to query.
- Audit the extractable unit and the prompts that currently hit the URL.
- Update evidence: dates, numbers, product names, screenshots of process (as text, not as the only proof).
- Restructure the answer block so the new claim can be lifted without the intro.
- Align FAQ, table, and body so they do not contradict each other.
- Recheck the same prompt set after a real window, then decide whether to stop or do a second pass.
That is the whole job. Everything below is how to do each step without turning it into a synonym article. If a step does not change HTML or the recheck log, it is optional commentary. Keep the five steps intact even when the team wants to skip to “add more FAQs.” Extra FAQs on top of a stale number still quote the stale number.
Step 1: Audit what an answer engine can actually lift
Open the live HTML, not the Google Doc. Models do not cite your outline. They cite what is in the page. Highlight definition sentences, FAQ answers, table cells, numbered steps, and any statistic. Those are candidates. Narrative case studies and motivational intros usually are not.
Then list the prompts. Use the query that already ranks on this URL, plus two close variants. If the page is the evergreen visibility post, the stale-refresh query is a mismatch. That is why this spoke exists. If the page is a decay monitor, the prompt might be “how do we know content is decaying,” not “how do we refresh for AI.”
Write down the sentence that currently appears in AI answers, if you have a log. If you do not, write the sentence you think they would take. That guess is your working unit. The refresh has to replace that unit in HTML, not merely add a new section underneath it.
What to capture in the audit sheet
Keep a small sheet: URL, primary prompt, current extractable sentence, date on the claim, last refresh, AI cited Y/N, Google position if you have it. You do not need a platform to start. You need one owner and a date column. If your team already runs a content decay monitoring workflow, add an AI-unit column there instead of inventing a second tracker.
Step 2: Update the evidence, not the vibe
Most “refreshes” rewrite the lede and leave the old number in section four. Answer engines still find section four. Put the new evidence next to the claim. If a statistic cannot be verified, remove it. Paraphrase a public source if you need a fact. Do not invent precision.
Evidence that usually needs a date: market sizes, model names, UI labels, pricing, “as of” process steps, sample sizes from your own tests. Evidence that can stay: stable definitions, decision criteria, failure modes that are still true.
In practice, we see teams “refresh” a page by adding 2026 in the title and leaving a 2024 table. The title is for humans in the tab. The table is for the model. Update the table. Then decide whether the title still describes the job.
Original measurements
If you measured something, isolate the method next to the number: window, sample, engine, definition of “cited.” If you did not measure it, do not dress a hunch as a percentage. One honest method note beats three vague “we’re seeing more citations.”
Step 3: Restructure the answer block
After the fact is current, make it extractable. A new number buried in a 90-word paragraph with a hedge and a CTA will get paraphrased into mush. A short definition, a FAQ answer, a table row, or a numbered step can be quoted.
Lead the page with the answer to the primary question. For this topic, the answer is: refresh the extractable unit, update evidence, restructure the block, recheck citations. Then support with the SOP. Do not hide the SOP under a history of AI search.
If the CMS makes tables ugly, still use a real HTML table for comparisons. Screenshots of Notion are weak citation units. FAQ answers should be two or three paragraphs that stand alone, not one-liners. Steps should be actions a practitioner can do this week.
Step 4: Align FAQ, body, and internal links
Contradiction kills trust. If the body says “refresh the unit” and the FAQ says “republish every 90 days,” the model will pick one and you will not control which. Write the FAQ from the same SOP. Use the searcher’s grammar, including awkward phrasing from GSC when that is the query.
Internal links should send people to the related job, not restated copies. Link evergreen visibility for cadence. Link living versus evergreen for the operating model. Link decay monitoring for the detection loop. Link how to improve old posts that stopped ranking when the problem is Google decay, titles, and links, not the AI unit.
Do not rewrite those URLs here. If you find yourself explaining what evergreen means for three paragraphs, you drifted. Link and return to the refresh pass.
Step 5: Recheck citations on a fixed prompt set
After you ship, wait. Answer engines do not recrawl on your editorial calendar. A useful window is longer than a day and shorter than a quarter. Recheck the same prompts, same engines, same wording. Note whether your new sentence appears and whether the old number disappeared.
If the old claim is still quoted, you did not replace the unit. Search the live page for the old number. If it is still there, you know why. If it is gone and still quoted, the model may be using a cached snapshot or a different URL of yours. That is a follow-up, not a reason to publish a synonym post.
Wins look boring: the current FAQ sentence shows up, the date is right, the URL is the one you intended. Brand mentions in a roundup are a weaker signal. Track them separately if you must. Do not call them a successful refresh.
A worked example: stale statistic on a living URL
A page explains evergreen content and still ranks for a refresh how-to query. The extractable unit is a 2024 sample: “FAQ-first pages appeared in 11 of 40 answers.” Prompt logs in 2026 still quote that sentence. Google position is stable. The team wants to “write a new AI refresh post.”
The right move on that URL is not a clone. It is: keep the evergreen URL on the evergreen job, and put the refresh SOP on its own URL when the query is different (this page). On the evergreen URL, either update the sample with a new window and method note, or remove the number if you cannot rerun it. Then recheck the prompts that were quoting 2024.
If you instead publish three new posts that each mention “refresh,” you split the how-to intent and leave the stale sentence live. That is how teams stay busy and still lose the quote.
Refresh versus the stopped-ranking playbook
Old posts that stopped ranking need titles, internal links, intent match, and sometimes consolidation. That playbook is still valid. AI refresh is a narrower cut of the same living-content system. You can do both on one URL. You should not pretend they are the same pass.
Stopped ranking: the page is hard to find. Stale for AI: the page is easy to quote and the quote is wrong. Fixing titles will not remove last year’s pricing from a table. Updating a table will not fix a title that no longer matches the SERP. Assign the owner by the failure mode, not by who “owns content ops” in Slack.
When both are true, do the evidence and extractable unit first, then the SEO pass. Models and crawlers both benefit from accurate HTML. Humans in the SERP still need a title that tells the truth.
Who should own the pass
The person who can change the HTML should own the refresh. A strategist can mark the unit. An analyst can pull prompts. If the ticket dies in “we’ll get to it in the next content sprint,” the stale sentence stays in production. Living content fails in the queue, not in the framework.
Give each URL a single next action: update fact, move fact into FAQ, remove unverified number, or leave. A list of twenty “consider refreshing” rows is not an SOP. Five URLs with a date and an owner is.
- Analyst: prompt set, current citation capture, date of last check.
- Editor: extractable unit rewrite, FAQ alignment, contradiction pass.
- SEO: title, slug stay/go, internal links, cannibalization check.
- Owner: ship date and the recheck date, not a vague “monitor.”
What not to do
Do not add “Updated 2026” in the first line and leave the body. Do not swap synonyms in every H2. Do not publish a second URL that repeats this SOP with a tool vendor’s name in the title. Do not gate the citeable claim behind a form. Do not refresh ten URLs the night before a report and call the report a test.
Do not chase every LLM-flavored prompt that adds “as a content engineer.” The job is still: find the stale unit, replace it, prove it. Role-play queries are noise around a simple operational loop.
Do not treat AI Overview volatility as a reason to skip measurement. Volatility is why you use a fixed prompt set. If answers change day to day, you still need a log. You do not need a new blog post for every swing. That monitoring job is a different spoke.
Refresh theater
Refresh theater is activity that looks like living content: new hero image, new intro, same table. It burns editor time and does not change what a model can lift. If the audit cannot point to the sentence that changed, it was theater. Ship fewer URLs. Change the unit.
How this fits a living content system
Living content is measure, improve, repurpose. This SOP is the improve step for pages that must stay true in AI answers. Measurement is the decay sheet plus prompt log. Repurpose is turning a validated FAQ into a sales one-pager, not spinning a duplicate blog.
If your system only publishes, you will keep asking how to refresh and never staff it. If your system only monitors decay in Google, you will miss the wrong quote in Perplexity. Put AI-unit health on the same review as content decay, with a lighter cadence if you have to, but the same URL list.
Click Laboratory uses this pass when a living-content URL is answering the wrong job, or when a cited sentence is older than the product. We would rather fix five units than ship five essays about “the future of AI search.”
A 90-minute pass you can run this week
You do not need a quarter-long program to start. Block 90 minutes on one URL that already shows up in AI answers or that ranks for a refresh-style query with the wrong page type.
- Paste the live page into a doc. Highlight every number, product name, and “as of” process.
- Run five prompts you actually care about. Save the cited sentence, if any.
- Replace or remove anything you cannot defend in a customer call today.
- Move the main answer into a FAQ or short definition at the top if it was buried.
- Ship. Put a recheck date on the sheet. Do not open a second URL until this one is done.
That pass is ugly and effective. The polished version is the same steps with a nicer sheet. Teams fail by skipping to URL six before URL one has a recheck.
Need a refresh pass on the URLs AI already quotes?
If evergreen and decay pages explain the system, but AI answers still quote last year’s number, the gap is the extractable unit on the live URL. We can map which pages already get impressions, which sentences are stale, and which refresh actions are worth staffing.
Request a living content / content engineering audit and we will run this SOP on the URLs you cannot afford to have misquoted, without turning it into another cadence calendar.
Stale content refresh questions for AI visibility
Use these when the evergreen page still ranks for a refresh how-to query, or when AI answers quote a number your team would not defend in a sales call. They follow the same SOP as the article: audit the unit, update evidence, restructure the answer block, then recheck citations. Cadence and merge decisions belong on those posts, not as a second refresh article.
How do you refresh stale content for AI visibility without rewriting the whole URL?
Start with the extractable unit, not the intro. Find the sentence, table row, or FAQ answer that should be cited, then update the evidence and date next to it. If that unit is still true, keep the URL. If the question the page answers has changed, restructure the first screen so the new answer is obvious. A full rewrite is for pages that no longer match intent, not for pages that only have an old statistic.
Use the evergreen visibility hub for cadence questions, and this SOP for the actual pass. After you ship, recheck the same prompt set. If the old claim still appears in AI answers, the refresh did not replace the unit the model was using. That is a HTML problem, not a calendar problem. Keep the slug, keep the primary intent, and change the claim the engine can lift.
If two URLs now answer the same prompt, merge instead of polishing both. Refreshing a duplicate does not recover the citation. It splits it. One owner, one URL, one recheck log.
What makes content stale for AI answers versus Google rankings?
A page can still rank in Google with an outdated example and still get skipped or misquoted in AI answers. Stale for AI usually means the citeable claim is old, contradicted, undated, or buried. Rankings can survive on backlinks and history. Answer engines prefer a claim they can lift without inheriting last year’s number.
Watch both. Search Console can show the page is still found. Prompt logs show whether the answer uses your current sentence. Refresh when either signal says the evidence is behind the question people are asking now. A falling position is a search problem. A quoted 2024 sample in a 2026 answer is a trust problem. They can happen on the same URL, and they need different edits.
If you only fix titles for CTR, the model can keep using section four. Put the new fact next to the old unit. Then confirm the old unit is gone from the HTML, not merely contradicted later in the article.
How often should we refresh pages that already get AI impressions?
There is no universal calendar. Review the URLs that already appear in AI answers or that get AI-related impressions on a monthly pass, and refresh when a fact, product, or process changed, not because a quarter ended. Fast-moving topics (pricing, regulations, model behavior) need shorter loops. Stable definitions can wait longer if the extractable unit is still true.
Tie the loop to a decay or citation sheet, not a random blog calendar. If you refresh everything every 90 days, you will ship noise. If you never refresh the pages that already get cited, competitors will replace your sentence. The cadence post on this site covers how often to stay cited. This page covers what to change when you open the URL.
Give the owner a stop rule: after one evidence pass and one recheck window, either the current claim appears or you schedule a second pass. Endless rewrites without a prompt log are not a living system.
Should we create a new URL instead of refreshing the old one?
Create a new URL when the question is different. Refresh when the question is the same and the answer aged. This page exists because “how do you refresh stale content for AI visibility” is not the same as “what is evergreen content.” Cloning a synonym URL for the same intent splits the citations you already have.
If the old URL ranks for the query and the content is wrong, fix the old URL. If the old URL answers a different job, link it and keep this SOP on the refresh job. Architecture mistakes look like “we published more” and still lose the quote. Searchers and models both follow the URL that historically matched the prompt. Do not abandon it for a fresher slug if the job did not change.
Use merge or redirect when two posts fight for one prompt. Use a new spoke when the job is new. That is the same rule as the refresh-versus-merge guide, applied to AI quotes instead of only rankings.
What should we update first: stats, FAQs, or the introduction?
Update the citeable unit first: the dated fact, the FAQ answer, the table cell, or the definition. Introductions rarely get quoted. If you only rewrite the lede, AI answers can keep using the old number from section four. Put the new evidence next to the claim the model already likes.
Then align the FAQ so it does not contradict the body. Then, if needed, tighten the intro so a human sees the same answer. Measurement comes last: recrawl the same prompts after a real window, not the next morning. A prettier hero line does not replace a stale table row. Models lift the densest, dated claim they can isolate.
If the FAQ still lists last year’s step, the FAQ is the unit. If the table is current and the intro is fluffy, leave the intro for last. Order the work by what can be copied into an answer, not by what is easiest to rewrite in the CMS.
How do we know a refresh worked for AI visibility?
Hold a fixed prompt set and compare source URLs and quoted sentences before and after. A win is your current claim appearing, not a brand mention in a roundup. Also check that the old statistic disappeared. If the model still cites last year’s number from your page, the extractable unit was not replaced in HTML.
Give it a window. Answer engines lag. A one-day check will lie. Pair the log with the decay workflow you already use for search, so the same URL does not get two conflicting owners. Record engine, date, prompt, quoted text, and whether the date on the claim matches what you shipped.
If impressions in Search Console rise but the quoted sentence is still old, the SEO refresh worked and the AI refresh did not. Treat those as separate outcomes. Do not call the job done because traffic recovered.
Does refreshing for AI hurt traditional SEO?
Not if you keep the primary intent, the URL, and the internal links, and you improve accuracy instead of stuffing synonyms. Search and AI both reward a page that still answers the query with current evidence. Harm happens when the refresh turns a focused URL into a new topic, or when you publish a duplicate spoke that cannibalizes the original.
Use the stopped-ranking refresh playbook for title, links, and decay. Use this SOP for the answer block. They stack. They should not fight over the same heading. A clearer definition and a dated table can help both SERPs and answer engines. A new H1 that chases a different keyword can hurt both.
If you are tempted to add a second primary keyword to “cover AI,” you are usually about to split intent. Keep one job per URL. Link related jobs. Measure each with the sheet that matches the job.
