What is AI citation decay?

AI citation decay is the gradual loss of citations a page or a brand earns in AI answers, measured over repeated runs of the same questions. A page enters the answer, holds a place in it for some weeks, appears less often, and then stops appearing. Nothing visible happens on the page itself. The engine simply stops choosing it.

Decay is worth separating from two things it gets confused with. A citation drop is an event: the page was there on Tuesday and gone on Wednesday, usually because the engine re-ran retrieval and something else won. Answer variance is noise: ask the same question twice in the same hour and the citation list differs, which is a property of how these systems sample rather than a change in your standing. Decay is the trend underneath both, and you can only see a trend by sampling repeatedly over weeks.

The market has settled on three words for parts of the same thing. Citation decay describes the downward trend. Citation drift describes the substitution of one credible source for another across runs. Citation half-life borrows from physics and puts a number on the trend: the time it takes for citation frequency to fall by half from its peak. All three describe a system that rotates its sources rather than ranking them once and leaving them.

The practical reason to measure decay rather than presence is that presence is not durable. A screenshot of a good answer is a record of one sample. Our note on why AI engines give a different answer every time you ask covers the sampling behaviour that makes a single check so weak as evidence.

How long does an AI citation actually last?

Most AI citations do not last at all. Trakkr Research's study of citation decay logged 108,650 citation URLs across 10,991 brands and 8 tracked models, running identical prompts daily and recording a first-seen and last-seen timestamp for every URL. It found that 73.5% of citations appeared once and never came back, and put the half-life of brand mentions at roughly 31 days.

New pages show the same shape from the other end. Semrush published 81 test pages and tracked their citations for 30 days in research released in December 2025. Google AI Mode cited 29 of those pages, 36% of them, within a day of publication. By day 30, only 21 pages were still being cited. The pages that won early were not the pages that held.

Repeat-run data puts a floor under how much consistency to expect. AirOps, in its research on how citations and mentions affect visibility, reported that only 30% of brands stayed visible from one AI answer to the next, and 20% across five consecutive runs. It also found that brands earning both a citation and a mention were 40% more likely to resurface than brands earning a citation alone.

Read those numbers as a distribution rather than a countdown on your own page. They come from vendor datasets across mixed categories, and a page answering a stable question in a slow market will outlive a page answering a news-shaped one. What the measurements agree on is that a citation has to be re-won constantly. The decision to cite you is taken again every time somebody asks the question.

Why do published citation half-life figures disagree, 11 days or 31 days?

The published figures disagree because they are not measuring the same thing. Every published half-life answers two questions before it gets anywhere near yours: whether it counted pages or brands, and whether the clock started at the peak or at the first citation. Each combination returns a different number from the same data, so a half-life quoted without its unit and its starting point cannot be compared with anything, your own log included.

The clearest case is one study that published two of them. Profound tracked 883,000 pages across seven AI engines for twelve months and measured each page separately on every engine that cited it, which produced 1.19 million page-and-engine lifecycles. The median page fell to half its peak citation share within 11 days of peaking, while the median gap between a page’s first citation and that same half-of-peak point was 42 days (Profound, “The half-life of an AI citation is 11 days”, checked 16 October 2026; this environment could not open the post directly, so both figures were taken from two independent readings of it). Same pages, same corpus, two numbers a month apart, and the only thing that moved was the starting line.

The median also hides a split that matters more than the median does. In the same Profound study, 78% of cited pages dropped to half their peak share inside a fortnight of peaking, and the 22% that got past the fortnight had a median half-life of 27 to 29 days instead (same source and date as above). Most pages are gone almost at once, and the ones that survive two weeks behave like a separate population with a benchmark of their own. A page still holding at three weeks should be read against 27 to 29 days rather than against 11.

Trakkr Research’s 31 days, quoted in the section above and widely requoted elsewhere, is a brand figure rather than a page figure, which is why it outlasts both of Profound’s. Brand mentions aggregate every page you own, so one of your pages losing its slot to another of your pages leaves the brand number where it was. That is the domain-level and URL-level gap the section on citations disappearing without a ranking change works through, arriving one level up.

Published figure	What was counted	Where the clock starts	Read it as
11 days (Profound)	One page’s citation share on one engine, smoothed on a two-week rolling average	That page’s own peak	How fast a win erodes once the page tops out
42 days (Profound, same study)	The same pages and the same share	The page’s first citation	How long a page earns anything at all
27 to 29 days (Profound, same study)	The 22% of those pages still above half their peak share after a fortnight	Peak	The benchmark that applies once a page has held two weeks
About 31 days (Trakkr Research)	Brand mentions across 108,650 citation URLs, 10,991 brands and 8 tracked models	Peak mentions	What the brand sees while its own pages rotate underneath it

Which half-life anchor should your own log use, the peak or the first citation?

Anchor your own half-life on the peak, and record a citation rate for every round so the peak can be found at all. A log that keeps only a first-seen and a last-seen date can produce the 42-day kind of number and never the 11-day kind, because the peak is a rate and not an event. Then hold that one anchor for the life of the series, since the comparison that decides anything is against your own earlier rounds and not against somebody’s published median. The calculation further down this page anchors on the peak, which is the stricter of the two and the one that tells you when to act.

Breadth and engine both change the odds for the pages that do last, which is the other argument for keeping the series separate. In the Profound study, among pages still cited beyond two weeks, a citation on Claude was 2.8 times as likely as one on ChatGPT to still be relevant at eight weeks, in the study’s own wording, and pages cited by five or more engines were 5.5 times as likely as pages cited by a single engine to be above half their peak share at that point (same source and date as above). Pages tied to a date or an event faded fastest of all.

An 11-day median also puts a floor under sampling cadence. Weekly rounds leave roughly one reading between a page’s peak and the point where half its share is gone, so the shape of the curve survives but the peak itself stays an estimate. A monthly round will usually record the decline and never record the high, which is how a page that worked for a fortnight gets written off as a page that never worked. Weekly stays the right default for commercial prompts, and a page you are watching after a specific change is the case where daily rounds pay for themselves.

Why do your citations disappear when your rankings have not changed?

Citations disappear without a ranking change because AI retrieval is stable at the domain level and volatile at the URL level. BrightEdge's 2026 analysis of week-to-week citation changes found that 96.8% of cited domains and 97.2% of mentioned brands showed no change week over week, and that the top two citation positions held steady 99.4% of the time.

Set that beside the 73.5% one-and-done rate the Trakkr study found for individual URLs and the picture resolves. The engines keep going back to the same trusted domains and keep changing which page from those domains they use, because the question arrives in slightly different wording each time and a different page fits best. Your standing has not moved. Which of your pages holds the slot has.

BrightEdge also found that movement, when it happens, is mostly downward. Among the roughly 3% of domains that did shift week over week, 87% were declines, and more than half of all citation volume was attached to domains that were losing ground, not gaining it. A monitoring program that only alerts on large swings will miss almost all of this, because almost all of it is small and one-directional.

Two things follow for how you report. Measure at both levels, because a domain-level dashboard will tell you everything is fine while the pages you invested in stop earning anything, and a URL-level dashboard alone will make ordinary rotation look like a crisis. And treat a lost citation on a page that still ranks as a content question, not an SEO one. Our guide to how AI engines choose which sources to cite covers what changes the answer at that level.

Which metrics should you actually track for citation decay?

Four metrics cover decay, and each one answers a question the others cannot. Citation half-life tells you how fast a win erodes. The one-and-done rate tells you whether you are earning tenancies or accidents. Reappearance rate tells you how dependable you look across repeated runs of the same prompt. Citation share tells you whether you are losing ground or the whole category is rotating.

Metric	What it measures	Published reference point	How to sample it
Citation half-life	Days until citation frequency for a URL or brand falls by half from its peak	About 31 days for brand mentions (Trakkr Research)	Weekly runs over at least 12 weeks, plotted per URL
One-and-done rate	Share of your cited URLs that appear in exactly one run and never again	73.5% of all citations studied (Trakkr Research)	Same prompt set, first-seen and last-seen dates per URL
Reappearance rate	How often you are cited again on the next run of the same prompt	30% of brands visible run to run, 20% across five runs (AirOps)	Back-to-back runs, same wording, same day
Citation share	Your share of all citations in the answers to your prompt set	96.8% of domains unchanged week over week (BrightEdge)	Weekly, scored against the competitors named alongside you

Sampling cadence decides whether any of the four is readable. Weekly is the right default for commercial prompts: frequent enough that a 31-day half-life shows up as a curve rather than a cliff, and cheap enough to keep running for a year. Daily sampling earns its cost only while you are testing a specific change and want to see when it lands. The cadence we run inside an engagement, and what each run is allowed to conclude on its own, is set out in the SIGNALS methodology.

Hold the prompt wording fixed, because a rewritten prompt resets the series and you lose the comparison that the whole exercise depends on. Keep a separate list for new prompts. Our guide to tracking AI visibility across engines covers the prompt-set design underneath this, and automating AI visibility reporting covers the plumbing.

Can Google Search Console show you citation decay?

No, and the reason is worth knowing before anyone builds a decay report out of it. Search Console’s generative AI report counts impressions, which it defines as “how many times links to your site were shown to a user in a generative AI feature on Google Search”, and it offers four dimensions: pages, countries, dates and devices. There is no queries dimension, and clicks, click-through rate and average position appear nowhere in its documentation. It covers AI Overviews and AI Mode in Google Search, reports Discover separately, and excludes Search Labs experiments (Google, Search Console Help, generative AI performance report, read at source 28 October 2026).

Two of those limits are fatal to the measurement this page is about and one is merely annoying. An impression is not a citation: a link shown under an AI Overview is the closest free first-party proxy anybody has, and it still counts appearances rather than the engine choosing your page as a source for a particular claim. Worse, with no queries dimension you cannot tie a fall in impressions to the prompt that lost you the appearance, which is the one join a decay log depends on. The annoyance is the pull. The Search Analytics API accepts six values for its search type parameter, discover, googleNews, news, image, video and web, and none of them covers AI Overviews, AI Mode or generative AI features (Google, Search Analytics API query reference, read at source 28 October 2026), so the AI numbers come out through the report’s export button by hand rather than into a scheduled job with everything else.

What Search Console’s generative AI report can and cannot tell you about decay, from Google’s own documentation read on 28 October 2026

Question you are asking of a decay log	Search Console
Were links to this page shown less often in Google’s AI features this month than last?	Yes, by page and by date, impressions only
Which prompt stopped returning us?	No. The report has no queries dimension
Was the page cited as a source, or just listed?	No. It counts impressions of links, not source selection
Did the same thing happen on ChatGPT, Perplexity or Claude?	No. Google surfaces only
Can the series be collected automatically each round?	No. Manual export; the API has no AI search type

So use it as a second witness rather than as the instrument. When your own panel shows a page sliding on Google AI Mode, the AI impressions for that URL falling over the same weeks is corroboration from Google’s own logs that something changed on Google’s side, and a flat impression line under a falling citation rate suggests the engine is still showing you and choosing you less, which is a different problem with a different fix. Keep the export beside the panel, dated, and never let it stand in for the prompt-level series, because the four numbers earlier on this page all need a denominator Search Console does not have.

How do you tell citation decay from ordinary answer variance?

You tell decay from variance by measuring your own noise first, because the day-to-day churn in an AI answer is large enough to hide a real decline inside it. GetMentions put the same 2,398 questions to ChatGPT, Google AI Mode, Perplexity and Gemini every day for a week and logged 530,875 citations, and reported that 69% of the sources behind a typical answer changed from one day to the next, with 84% of the sources cited for a question used by only one of the four engines (GetMentions, “AI Citation Volatility: A 530,875-Citation Study”, checked 28 September 2026). Against turnover at that rate, a page missing from today’s answer is one sample and not a finding.

Variance starts before a day has passed, which is the part most logs never capture. Parse re-ran more than 16,000 prompts about 22 times each across ChatGPT and Google AI Overviews in March and April 2026 and found that repeat answers to the same prompt shared only about a fifth of their cited sources on ChatGPT, and about 31% on Google AI Overviews (Parse, “AI citation volatility by industry”, checked 12 October 2026). Two runs of the same question minutes apart can therefore disagree about your page, so a round that consists of one run per prompt measures the draw rather than the position.

How much noise you carry depends on how often you are cited at all. BrightEdge’s citation volatility analysis found that frequently cited domains moved about 0.7% week to week while sporadically cited domains swung by more than 50%, and located the inflection near 50 citations, where weekly volatility drops from roughly 50% to 8% (BrightEdge, “AI Search Engine Citation Volatility Analysis”, checked 28 September 2026). A page above that line and a page below it need different thresholds, and the same five-point drop means opposite things on the two of them.

Baseline volatility also differs by engine, which is why a single number pooled across engines is the least readable figure you can produce. The same analysis reported these weekly averages, and found that high-authority domains held under 10% volatility on every engine it tracked.

Engine	Reported average weekly citation volatility	What that means for your threshold
Perplexity	72.8%	Expect most of the citation list to turn over weekly. Only a sustained multi-round slide is readable here
Google AI Mode	62.4%	Similar. Compare rolling averages rather than this round against the last one
Google AI Overviews	41.8%	The steadiest of the three, so a drop here is worth investigating sooner
Any engine, frequently cited domain	Under 10%	Tight enough that a single lost citation is a signal rather than noise

ChatGPT was tracked in the same analysis and the pattern held, though we are not quoting a baseline figure for it because we could not confirm one. Treat the four numbers above as the published expectation and your own two baseline runs as the threshold that actually applies, which is what the SIGNALS methodology schedules a week apart before any fix goes live. Two rules fall out of it. Keep one series per engine, because a corpus with 84% single-engine sources is four different measurements wearing one name. And call a move real when it clears your own baseline spread or runs the same direction across three consecutive rounds, whichever comes first. How many runs each round needs before the number has the resolution to show a slide at all is worked out in how many runs a visibility measurement needs.

How do you calculate a decay rate between two reporting periods?

A decay rate between two reporting periods is the later period's citation rate minus the earlier one's, divided by the earlier one's, read as a percentage. A page cited in 30% of runs last month and 18% of runs this month has decayed by 40%, because twelve divided by thirty is 0.4. Work in rates rather than raw citation counts on both sides of that subtraction: a count also falls when you shorten the prompt set, and a smaller denominator is not decay. It is the same subtraction the SIGNALS methodology applies between two rounds, and the rate is the only figure in the log that is comparable across them.

The subtraction only carries information if five things are identical on both sides of it. The engine, one series each and never pooled. The prompt set and its exact wording. The number of runs per prompt. The geography and account state the runs were made in. And the rule that decides what counts as a citation, written down before the first round rather than settled afterwards. Change any one of the five and the number you have produced is a measurement of your own setup, which is why the change log belongs beside the data instead of in somebody's notebook.

Then ask where the loss sits before acting on it, because one site-wide decay rate describes two unrelated problems. It can be a page that lost nearly all of its citations while the rest of the site held, or a site whose pages each gave up roughly the same share of theirs. The first is a page to fix. The second is usually a retrieval or indexing change that no amount of editing one page will answer.

One decay rate, two different problems. The figures are arithmetic on a worked example rather than a measurement of any real site.

Shape of the loss	Period 1	Period 2	Decay rate	What it points at
Concentrated	50 citations, 46 of them on one page	30 citations, 6 of them on that page	40% site-wide, 87% on the page	That page. Check whether it still answers the prompt it used to win, and whether a competitor published something better
Spread	50 citations across 18 pages	30 citations across 17 pages	40% site-wide, roughly 40% on most pages	Not the pages. Look at crawler access, indexing, and whether the engine changed how it retrieves for this category

Report the per-page rate and the site-wide rate together for that reason, and keep the page count beside both. A site-wide figure on its own is the one number in this whole measurement that can move without anything you control having changed.

How long a window do you need before a decline counts as decay?

The window you need before a decline counts as decay is three to four weeks of daily runs, and ten days is the floor rather than the target. Julius Schulte, Malte Bleeker and Philipp Kaufmann, in Don't Measure Once: Measuring Visibility in AI Search (GEO), arXiv 2604.07585, 8 April 2026, ran a rolling-window analysis over 1,726 per-brand series spanning 40 to 46 days each and found that the standard error of a detection rate estimated over a window of d days falls below 0.10 at ten days, below 0.05 at twenty-four, and below 0.02 at thirty-four. We read those figures off the preprint itself on 19 October 2026 rather than off coverage of it.

Window of daily runs	Standard error of the rate	What you can say with it
1 day	Not estimable from one observation	Nothing. One round is a draw
10 days	Just under 0.10	A large drop is visible. A ten-point move is not
24 days	Just under 0.05	A sustained decline of the size worth acting on
34 days	Just under 0.02	Fine enough to compare one month against the next

Two findings in the same paper explain why the floor sits that high. The sources an engine cited overlapped from one day to the next by only 34% to 42% across its four campaigns, aggregated over 4,044 consecutive-day pairs, while the brands it named overlapped by 45% to 59%, so the source list is the more volatile of the two series and a cited-URL trend inherits that volatility. And within a single day, the authors found the standard error of a brand's detection rate drops below 0.10 at seven runs per prompt and that source coverage needs eight, with the 95% interval at seven runs still wide enough, on the authors' own reading of it, to detect a large difference but not to rank brands of similar visibility. Both figures were read at source on 19 October 2026.

So the practical rule has two halves and both have to hold. Enough runs per round, seven at the very least and eight if you are trending which page was cited rather than whether the brand was named. And enough elapsed time, with a month compared against a month rather than this week against last. A scheduled job that runs the panel nightly is what makes a twenty-four day window affordable, because the alternative is somebody running prompts by hand for three weeks before the first decision. Anything faster than that is a report of your own sampling noise, presented with a date on it.

How do you calculate citation half-life from your own log?

Citation half-life is the elapsed time multiplied by 0.693, then divided by the natural log of the peak citation rate divided by the current one. The inputs come off the log rather than out of a tool, so the log has to record the denominator as well as the hit. Every round needs the number of prompts run, not just the number of answers you appeared in, because the rate is the only thing comparable between rounds.

One row of the log

Round

Round number and the date it ran

Engine

One engine per row. Never pooled

Prompts

How many prompts were run in the round, worded exactly as the round before

Cited

The URL, and how many of those prompts cited it

Dates

First seen and last seen for that URL

Worked through on one page, over six rounds three weeks apart on a fixed set of 40 prompts against a single engine:

Round	Week	Prompts citing the page	Citation rate
1	0	12 of 40	30.0%
2	3	10 of 40	25.0%
3	6	8 of 40	20.0%
4	9	7 of 40	17.5%
5	12	6 of 40	15.0%
6	15	5 of 40	12.5%

The peak is 30.0% and the latest round is 12.5%, fifteen weeks later. Fifteen multiplied by 0.693, divided by the natural log of 30 divided by 12.5, which is 0.875, gives 11.9 weeks. So this page has a half-life of about twelve weeks, and the table confirms it from the other direction: the rate reached half its peak at round five, week twelve. On that curve the page is cited in about 6% of runs twelve weeks from now and about 3% twelve weeks after that. These numbers are arithmetic on a worked example rather than a measurement of any real page, and they are the same calculation the SIGNALS methodology runs on a measured one.

Read your own number against the published ones rather than instead of them, and there are now two from separate datasets. Trakkr Research put the median brand half-life at roughly 31 days. Profound’s analysis of 240 million citations put the median half-life of a cited source at about 4.5 weeks, which is the same month arrived at from a different corpus and a different unit of analysis (Profound’s 240-million-citation analysis, as reported by Searchbloom and LumenGEO, checked 2 October 2026; we were not able to open Profound’s own write-up). Against either figure a twelve-week half-life is a page decaying slowly for its category, and the decision it supports is when to refresh rather than whether to. The same arithmetic on a page cited twice at its peak returns a number with no information in it, which is what the volatility threshold in the section above is for. Our guide to tracking AI visibility across engines covers how to build the prompt set that produces the denominator.

How do you score citation stability from one run to the next?

Citation stability is the overlap between the set of URLs an engine cited last run and the set it cited this run, scored from 0 to 1. Count the URLs that appear in both runs, divide by the count of URLs that appear in either, and you have it. A score of 1.0 means the engine returned an identical source list; 0.0 means nothing carried over. Turnover is one minus that score, and it is the number most published research reports.

Work it on two rounds of the same prompt set. Round one cited 18 URLs across the answers, round two cited 20, and 7 of them appear in both lists. The two rounds name 31 distinct URLs between them, so stability is 7 divided by 31, or 0.23, and turnover is 77%. Those two figures are arithmetic on a worked example rather than numbers out of a study. Score the whole cited set rather than only your own pages, because the number you want is how much the engine churns on this question. Your own appearances are a subset of that, and judging them without the denominator is how a normal week gets read as a collapse.

The published turnover rates say what to expect before you have two rounds of your own. SISTRIX ran 82,619 qualified prompts through Google AI Mode, Google AI Overviews and ChatGPT Search across six countries over the 17 weeks from 17 December 2025 to 8 April 2026, taking weekly reference dates and logging 1,548,213 snapshots. The study found that Google AI Mode exchanged 56% of its cited sources each week and ChatGPT Search 74%, while AI Overviews stayed completely stable for just over half of the queries tracked (SISTRIX, “AI Citation drift: How stable are sources in AI search results?”, read 2 October 2026).

Published weekly source turnover, and the stability score it implies

What is being counted	Weekly turnover	Stability score	Source
Cited URLs, Google AI Mode	56%	0.44	SISTRIX, 82,619 prompts over 17 weeks
Cited URLs, ChatGPT Search	74%	0.26	SISTRIX, same dataset
Cited URLs, Google AI Overviews	No change for just over half of queries	1.0 on that half	SISTRIX, same dataset
Cited domains, all engines	3.2%	0.97	BrightEdge, week-to-week citation analysis

The last row is the one that makes the other three readable. At domain level the engines barely move, and at URL level two of the three replace most of what they cited a week ago. So a stability score near 0.25 on ChatGPT Search is the resting state of the engine rather than evidence about your page, and a single missing citation inside it is not a measurement. What does carry information is a stability score that falls well below your own two baseline rounds for the same prompts and the same engine, or a page absent from four consecutive rounds on a prompt whose list is otherwise holding. Keep the score per engine for the reason the variance section gives, and log it beside the citation rate each round so the two can be read together.

What do you do when a page starts to decay?

Diagnose before you rewrite, because three different failures look identical on a decay chart. The page may have been dropped from retrieval, in which case it is absent from every answer instead of losing to a rival. It may still be retrieved and no longer chosen, which is a content problem in the comparison the engine is making. Or the question itself may have moved, and the page now answers a phrasing nobody uses.

Separate the three by reading what replaced you. Run the prompt, list the sources the engine cited instead, and open them. If the replacements are newer pages answering the same question, you have a freshness and depth problem on a question you already own. If they answer a different question, the query has drifted and you need a new section or a new page, not an edit. If nothing replaced you and the answer simply got shorter, the engine found less to work with and the whole category is thinner than it was.

Where the fix is a refresh, change what the evidence rewards. Add the sourced numbers the current version lacks, date them, and give each claim a named source in the same paragraph, which is what makes a sentence quotable in the first place. Tighten the answer under each heading so the first sentence resolves the question. Then re-run the prompt set weekly and watch the curve rather than checking once and declaring it fixed.

Not every decaying page deserves the work. A page losing citations on a question with no commercial value is telling you where not to spend. Rank your decay list by whether the prompt has a buyer behind it, which is the same judgement our AI citation checklist applies at the page level.

How do you recover a lost AI citation, and how long does recovery take?

You recover by winning the question again, not by restoring the citation you had, and for one page on one engine the old citation is unlikely to come back. The reason is in the figures above: Trakkr Research's log of 108,650 citation URLs across 10,991 brands and 8 models found 73.5% of citations appeared once and never returned. A citation that vanishes is the ordinary case rather than an incident, so treat recovery as re-entry into the retrieval set for a question, and judge it on how often the page appears across a run of prompts instead of on whether one URL reappears.

Set the threshold before you start or you will read noise as a recovery. Profound defines decay on its own product as a page's citation share falling to half its peak and holding there for 14 days, and states the condition explicitly: the page only counts as decayed if it stays below the threshold "with no bounce back above the threshold in between" (Profound, citation decay, published 29 September 2026, read 4 November 2026). Run that rule backwards and it gives you a usable test for the other direction. One run back above the line is not a recovery, and a fortnight above it is a reasonable minimum before you record one, which is the same logic as the stability score further up this page.

The one recovery lever with a measured number attached is not on the page. AirOps, in its research on how citations and mentions affect visibility, found brands earning both a citation and a mention were 40% more likely to resurface than brands earning a citation alone. That splits recovery work in two. The page gets the rewrite described in the section above, and the brand needs to be named in places the engine already reads, which is why third-party mentions do more for a decayed page than another internal link will.

On the timeline, the honest answer is that nobody has published a figure worth repeating. Several vendor pages quote recovery windows in weeks with no method, no sample and no definition of recovered, and this page is not going to launder one of them into a benchmark. What is measured is the floor: Semrush's 81 published test pages were picked up by Google AI Mode for 29 of them, 36%, within a day of publication, so crawling and indexing are not where most of the waiting happens. The wait is the comparison. Your rewritten page has to beat the pages that replaced you, and you can read that contest directly by running the prompt and opening whatever the engine cited instead.

One piece of context makes the size of the job clearer. Profound states that half of the content answer engines cite is less than 13 weeks old, in the same post as the decay definition above, and publishes no method or sample for the claim, so take it as the vendor's characterisation of its own corpus rather than a measurement you could check. Even read loosely, it sets the standard a recovering page is held to. A page last revised six months ago is competing for a slot against a body of sources most of which is newer than that.

What else do people ask about AI citation decay?

What else do people ask about measuring decay in practice?

What else do people ask about citation stability and half-life?