How Long Does Information Stay in AI Answers — Crisis Management and the Time Axis

How long does information about corporate misconduct or an online backlash remain in the answers generated by AI? Within the scope of this review, we could not confirm any public data that answers this question directly. This article therefore does not offer a figure for how many days it takes to disappear, nor a claim that it never disappears.

But this is not a discussion that ends at "we do not know." The structures that govern persistence can be confirmed from large-scale empirical observation and from the providers' official documentation.

First, the stability of citation sources differs greatly from platform to platform. Google AI Mode and ChatGPT Search show large week-over-week turnover, while in Google AI Overviews the citation sources stayed fixed across 17 weeks for more than half of the prompts. Neither "AI answers are unstable" nor "AI answers are stable" holds as a blanket statement.

Second, the generation of an AI answer involves the retrieval of external web information, but how trained knowledge and retrieved results are integrated in any individual answer is not published.

Third, within the scope of this review, we could not confirm a general right for a legal person to have true but unfavorable information about itself removed from AI answers. The separation of scope and subject under each legal instrument is handled in the companion article "Does the right to be forgotten reach AI answers?"

There is a further point that is easily missed. Data that measures the turnover of citation sources is not data that measures whether information has disappeared. The same study shows that the body of the answer can change even when the citation sources have been fixed for 17 weeks. The stability of citation sources and the stability of the answer text are separate metrics, and neither can be uniquely inferred from the other.

The conclusion narrows to one point. When it disappears cannot be predicted. And because it cannot be predicted, the alternative to waiting for it to disappear is continuous observation.

• Why the assumption that "time solves it," on which crisis response has been built, may not hold for AI answers
• Why unfavorable information that is true does not fall within correction premised on being wrong
• The four structures that govern persistence, and the platform differences in the stability of citation sources
• Why data showing that "citation sources turn over" must not be read as evidence that "information disappears"
• How to build AI observation into a crisis response timeline (what to observe, what to record, what not to do)
• A list of what could not be confirmed in this review, and where to send the questions about legal instruments

Public relations and crisis management practitioners, and the executives accountable for them. It assumes organizations that have already experienced an incident, as well as organizations that want to understand how a past incident they consider closed is being treated in AI answers.

AIO, GEO and LLMO are practitioner terms for dealing with citation, mention, display and impression in AI search; they are not official standards. Google states in its official documentation that there are no additional requirements for appearing in AI Overviews or AI Mode and that no special optimization is needed, and it lists the same technical fundamentals as ordinary search. For an organization of the terms, see What is AIO and What is LLMO.

The central question of this article is: how long does information about corporate misconduct or an online backlash remain in AI answers?

We could not confirm data that answers this question directly. More precisely, within the scope of this review, we could not confirm any public research or study that repeatedly retrieved generative AI answers over time under fixed conditions, for a specific instance of misconduct at a specific company, and measured the decay period of the mention rate for that fact.

Descriptions citing a number of days circulate in the trade press, but we could not confirm any for which the composition of the sample, the measurement procedure and the method of reproduction are published. This article does not adopt those figures.

Nor does this article set out to solve the problem from the side of legal instruments. Whether existing removal instruments reach AI answers is handled by the companion article "Does the right to be forgotten reach AI answers?" This article confines itself to the time axis and observation.

On that basis, the article proceeds in this order: a review of the conventional assumption (§1), the treatment of information that is true (§2), an organization of the structures that govern persistence (§3), the boundary lines needed to avoid misreading the existing data (§4), and the practical design (§5–§7).

"We do not know" is not the same as "nothing can be said." The period is unknown, but the structures that govern the period are known.

Public relations and crisis management practice has long been built on an assumption of decay.

In conventional crisis response, the calming of coverage volume and of reactions on social platforms can serve as one element in judging that an incident has closed. The design of "if the initial response is not botched, time is on your side" rests on that observation. A view is formed of how long it will take to settle, a response structure is assembled, and closure is declared. Many crisis management manuals contain this time axis, whether explicitly or not.

However, within the scope of this review, we could not confirm grounds for transferring that time axis directly to AI answers. This article deliberately gives no specific number of days precisely so as not to leave this point vague.

1-1. The world of search had mechanisms for handling time

In the world of search there were, at least for some cases, explicit means of controlling display status, together with a period.

Google provides a removal tool that lets a site owner temporarily hide pages on their own site from search results, and its official documentation states that the effect lasts about six months. The same documentation states that permanently preventing display requires one of the following: deleting or updating the page itself, password protection, or applying noindex; and it states that blocking with robots.txt is not an appropriate method for this purpose.

Scope matters here. This means is available only for URLs the company controls, and does not extend to news articles or third-party posts. Even so, for URLs under a company's own control there were two states — temporarily hidden and permanently removed — with corresponding procedures and an indication of duration. Within that scope, time could be built into the design.

1-2. For AI answers, it is unclear whether that assumption holds

The problem is that whether this assumption also holds for AI answers cannot be confirmed from the outside.

An AI answer is not a list of URLs but generated prose. What can be confirmed from Google's official material goes as far as the fact that AI Overviews and AI Mode retrieve external web information through Search and query fan-out; how trained knowledge and retrieved results are integrated in an individual answer is not published. If decay occurs, where does it occur, and at what speed? What this article addresses is the content of that "we do not know."

This is the decisive difference between this article and the series of articles that deal with misinformation.

The problem of AI stating incorrect information about a company is handled in AI misinformation countermeasures (the types and structure of erroneous statements), correction and removal of AI misinformation (means of correction and the available instruments), and AI misinformation and legal liability (legal means). This article carries those conclusions forward as its premise.

What this article addresses is information that is not wrong.

2-1. Most corporate misconduct is fact

An administrative sanction was imposed. A recall was carried out. A lawsuit was filed. A director resigned. A post on a social platform drew criticism and an apology was published. In many cases these are facts: they are reported on the basis of published material, and often the company itself has formally acknowledged them.

Because they are facts, the following hold at the same time.

Unfavorable information that is true does not fall within correction premised on the information being wrong. That said, even for accurate information, erasure, cessation of use or removal may be granted for natural persons on other legal bases, such as privacy, personal data protection, or the removal of search results.

These two are also designed separately as a matter of legal instruments. "Correction of error" and "erasure or cessation of use on other grounds" are placed in different provisions in both the EU and Japan (for detail, see the companion article "Does the right to be forgotten reach AI answers?"). The accurate understanding is not "nothing can be done because it is true," but "on the premise that it is true, judge case by case which basis is available."

2-2. Which is why time is what remains

Where information is misinformation, a route exists in principle: point out the error and have it fixed (how far that route actually functions is handled by the article on correction and removal). At the very least, what is being asked for can be defined.

For unfavorable information that is true, that definition is hard to establish. The principal crisis management variables this article addresses are time and context.

• Time: waiting for the frequency with which the information is discussed to fall
• Context: accumulating the record of what was done afterwards, and aiming for a state in which it is set alongside the description of the incident

That said, this does not exhaust the available means. Where legal removal or cessation of use is established on another basis, lawful updating or deletion of the source information, de-indexing from search results, platform-specific reporting channels, and requests for correction to third-party outlets may remain available depending on the situation.

2-3. "So it cannot be removed legally?" — that question is handled in a separate article

Every reader who has come this far will ask the next question. If correction premised on error is unavailable, can the information not be removed legally?

It is a legitimate question. And this article does not answer it.

The reason is that the examination required to answer it sits on a different axis from this article. The right to erasure under Article 17 of the EU GDPR, the Google Spain judgment, the 2017 decision of the Supreme Court of Japan, the analyses by CNIL and the EDPB of exercising rights against AI models, the distinction in Japan's Act on the Protection of Personal Information (APPI) between "correction, addition or deletion" and "cessation of use or erasure," the standing of Article 50 of the EU AI Act — all of these concern legal instruments and rights, and sit on a different axis from the time axis and observation that this article addresses. Touching on them halfway would leave both imprecise.

This question is addressed head-on in the companion article "Does the right to be forgotten reach AI answers?" There, the scope and the subject are separated instrument by instrument, and the conclusions are shown to diverge between legal persons and natural persons. To state one point in advance: within the scope of this review, we could not confirm a general right for a legal person to have true but unfavorable information about itself removed from AI answers. For the grounds and the exceptions, please see the companion article.

The separate question of whether information that is wrong can be contested legally is handled by AI misinformation and legal liability.

From here, this article proceeds on the premise that legal instruments cannot be relied on, treating the matter as a question of time and observation. On time, no outlook can be established. Context is handled in §5.

2-4. Handling cases where error is mixed in

In practice, fact and error are mixed. The fact of a sanction is correct, but its content, timing or scope is described incorrectly. In such cases, handle them separately. Deal with the incorrect portion within the misinformation framework (see AI misinformation countermeasures), and treat the correct portion as a question of the time axis. If they are left mixed together and the aim becomes "remove all of it," neither response can be designed.

The period is unknown. But what governs the period can be organized from empirical evidence and official documentation.

3-1. Structure ① The stability of citation sources differs greatly by platform

A study published in May 2026 by SISTRIX, a provider of SEO tools, observed 82,619 prompts and 1,548,213 snapshots across six countries, three platforms and 17 weeks (17 December 2025 to 8 April 2026, weekly reference dates). It measured, at the domain level, how far citation sources turn over from week to week.

The main finding of the study is not "turnover is severe" but "the structure differs entirely from one platform to another."

In AI Overviews, for 53% of prompts not a single citation source changed across the 17 weeks. A further 28% saw only partial change, and the remaining 19% turned over at a rate comparable to AI Mode (churn-in 46%). The structure is polarized between being in and being out.

In AI Mode, 86.5% of prompts have a stable core of one to five domains, while everything outside that core turns over at 89% per week. In ChatGPT Search a core structure is itself rare: the median prompt contains not a single domain that appears in all 17 weeks. The figure of 74% is the German-language value, and country differences are also reported, with 60% in the United Kingdom and 42% in France.

On top of that, for AI Mode no convergence over time has been observed. A consistent weekly churn-in rate of 54–59% continues across the six countries, and the study concludes that no trend toward stabilization is visible across the 17 weeks. At URL level the figure is larger still, at 85% per week.

"AI answers are unstable" must not be used as a blanket statement. Which platform and which structure the questions relevant to your company sit in cannot be known without observing.

* This study is a first-party study based on data from the provider's own tool, and the figures need to be read separately for each platform. That the agreement of citation sources is limited even when the same query is re-run on the same day is handled in PR and the citation sources AI uses.

3-2. Structure ② No correction process for legal persons can be confirmed

Within the scope of this review, we could not confirm, as a feature common to the major AI providers, a process by which a legal person can file for correction of statements about itself with a guaranteed processing deadline and outcome (for detail, see correction and removal of AI misinformation). From the standpoint of the persistence period, this means that no route for shortening the period by means other than waiting can be confirmed as an established process.

3-3. Structure ③ Rewriting inside the model is not stable

For techniques that locally rewrite the internal knowledge of a trained model (knowledge editing), problems have been reported in which the rewrite is not stable and ripples out into neighboring statements (for detail, see correction and removal of AI misinformation). From the standpoint of the persistence period, this means that the premise that "it is enough to remove a particular statement from inside the model" cannot be relied upon.

3-4. Structure ④ Trained knowledge and search results are mixed together

This is the structure that makes the outlook on the period difficult, but the layers of evidence need to be read separately.

Layer one: what can be confirmed officially. Google states officially that AI Overviews and AI Mode may use "query fan-out" — a technique that issues multiple related searches across subtopics and data sources — and that, while generating a response, this identifies more pages to reference and can surface a broader and more diverse set of links than ordinary web search. In other words, one question can be expanded internally into several derived questions. There is a route by which material about a company is gathered even from questions that do not search on the company name.

What can be confirmed officially ends there. Google Search Central states that external web information is retrieved, but it does not publish how trained knowledge and retrieved results are integrated in an individual answer.

Layer two: what has been confirmed in general LLM research. The study by Xie et al. presented at ICLR 2024 (arXiv:2305.13300) set a model's parametric memory against external evidence that contradicts it under controlled conditions, and reported two behaviors that appear contradictory. Where the external evidence is coherent and convincing, the model shows high receptiveness even to evidence that contradicts its parametric memory. On the other hand, where the external evidence contains information consistent with the parametric memory, the model shows strong confirmation bias even when contradictory evidence is presented at the same time.

This result cannot be converted directly into a persistence period in commercial AI search. A controlled experiment and a product in live operation are different things. Even so, it does suggest that a simple model of "new information equals overwrite" cannot be taken for granted.

3-5. There is also a published range for how fast changes are reflected

There is no data on the period itself, but there is a range that providers state officially regarding the time it takes for changes to be reflected.

Google states that where display does not change after snippet controls have been implemented, one response is to wait for the page to be recrawled and the change processed, and puts the time required at several days to several months. This concerns the speed of reflection rather than removal, but the fact that no upper bound is given weighs heavily on anyone designing a time axis.

OpenAI puts the time for a robots.txt update to be reflected on the search side at about 24 hours, while stating explicitly that a site that has opted out of OAI-SearchBot will not appear in ChatGPT's search answers but may still appear as a navigation link. Blocking does not necessarily eliminate every path to display.

3-6. A note: what remains outside your own site

Even after a company corrects its own site, information remains on third-party surfaces such as news coverage, message boards, reviews and video (the supply structure is handled by PR and the citation sources AI uses). In the context of this article, it is enough to note that the surfaces a company can operate on and the surfaces where information remains do not coincide.

The data in §3-1 is powerful, but it is easily misread. Here we draw the boundary lines.

4-1. News articles are unlikely to persist as citation sources

The SISTRIX study classified citation sources in Google AI Mode by domain type and reported, for each type, the proportion that enters the stable core (core rate).

The core rate for news media is 1.4%, the lowest among the categories. The study describes this as news being a one-way ticket.

To be precise, however, 1.4% is the proportion that remains in the stable core; it is not a value that directly measures the citation lifespan of an individual news article. It cannot be read as "everything other than that 1.4% disappears the following week." What it shows is a tendency for news domains not to settle in as permanent citation sources.

Information about misconduct comes into the world, in most cases, as news articles. It is tempting to reason from this figure that "if it stops being cited as time passes, surely the information disappears too." That inference does not hold.

4-2. The stability of citation sources and the stability of the answer text are separate metrics

The same study confirms the following about AI Overviews. Among the prompts whose citation sources did not change at all across 17 weeks, in 87% the body of the generated answer changed from week to week. That is, different prose was written each week from the same eight sources.

This single point sits at the center of this article. What has been demonstrated is one direction only, as follows.

Even where citation sources are fixed, the body of the answer can change.

The reverse direction — that if citation sources turn over the content of the answer also changes — was not tested in this study. Nor did the study test whether the two are statistically independent. The accurate formulation is therefore this. The stability of citation sources and the stability of the answer text are separate metrics, and neither can be uniquely inferred from the other. The churn-in rate of citation sources is a metric for which sites keep being cited, not a metric for what keeps being said.

4-3. What the absence of public data actually means

Putting the foregoing together, the state of the public data is as follows.

What is measured is the supply side (which sites are cited), not the content as seen from the demand side (what is said). This asymmetry is the structural reason why the question of when it disappears cannot be answered.

4-4. In addition, agreement between platforms is low

The same study reports that, for the same prompt, the domains cited by AI Overviews and by AI Mode differed in 83% of cases, with a Jaccard index of 0.17 indicating the overlap. Between AI Mode and ChatGPT Search it was lower still, at 0.125. Even two features from the same provider barely overlap in their citation sources.

Whether something has disappeared from AI answers therefore cannot be settled by checking a single engine. It has to be viewed across multiple engines.

Since the period cannot be predicted, practical design has to move from estimating to keeping watch.

5-1. Professional bodies have begun to address monitoring itself

Before turning to design, it is worth confirming how far professional bodies have addressed this area. The movement divides in two.

Discipline for those who use AI. On 10 June 2026 the Chartered Institute of Public Relations (CIPR) in the United Kingdom published a best practice guide on the responsible use of AI in PR practice. It translates the seven principles of the Global Alliance's Venice Pledge into practice, setting out checklists for prior evaluation of tools, handling of data, verification of outputs, and items to confirm before publication. This concerns how practitioners use AI.

Practice for those who observe AI. On the other side, on 9 July 2026 PRsay, the blog of the Public Relations Society of America (PRSA), published an article dealing with the design of reputation monitoring on language models. It organizes reputation into three dimensions — awareness, attitude and attribution — and discusses monitoring that systematically collects, analyzes and evaluates what language models say about a company. That it lists typical questions such as quality, reliability, corporate misconduct, sustainability and attractiveness as an employer overlaps with the concerns of this article.

In other words, the practice of reputation monitoring on LLMs has begun to be addressed around the professional bodies.

Even so, within the scope of this review, we could not confirm any standard or guidance from a professional body indicating the persistence period of information about a specific instance of misconduct. The treatment reaches the level of "you should observe" and does not extend to "how long it remains." The design of observation therefore has to be assembled by each company itself. What follows is a proposal for that design.

5-2. Building it into the crisis response timeline

AI observation should be built into each phase of the existing crisis response timeline rather than run as a standalone initiative.

The frequency after closure is decided by the nature of the incident and the possibility of recurrence. Watching at the same frequency forever is not realistic. Define a period of stability and the recurrence conditions (similar incidents, related regulatory developments, progress in pending proceedings, and so on), then move to ordinary periodic monitoring.

The most important element is the record from before the occurrence. Without something to compare against, it is impossible to tell whether the post-incident state is attributable to the incident. Observation in normal times is what gives observation in a crisis its meaning.

5-3. What to observe

Whether something is displayed is not enough. In a crisis management context, the content has to be examined as well.

The last row is the subject of this article. From a one-off snapshot, it is impossible to distinguish a change from mere fluctuation.

5-4. What to record

AI answers can differ from execution to execution even for the same question. Without records, nothing can be verified afterwards.

Stateless, under identical conditions, multiple times. Vaipm measures AI-space perception through a total of 25 stateless queries across multiple AI engines. This is not an optimum derived from research; it is Vaipm's operational design. The point of fixing the number and the conditions lies less in the values themselves than in maintaining a state in which comparison with the previous observation is possible.

A caution applies here. Where a crisis incident involves natural persons such as officers or employees, storing the full text of AI answers over the long term can itself constitute the handling of personal information. The requirement to retain full text for verifiability collides head-on with the requirement not to accumulate personal data unnecessarily. The measurement design must therefore always include three elements: access rights (who may view the records; restrict this to crisis management and legal), retention period (do not hold records beyond the period needed for verification), and a deletion policy (who decides and by what procedure, and how descriptions identifying an individual are treated).

Do not accumulate without limit on the grounds that it is "for measurement." The subjects of rights concerning personal data are natural persons (see the companion article); this is a condition to be built into the measurement design rather than a detail.

5-5. Accumulating context

Since the information is true, eliminating the description is unlikely to be a workable goal. Rather than placing the goal on removal alone, observe for a state in which, when the incident is mentioned, the subsequent course of events is set alongside it. The measures for supplying that information are as follows.

• Publish the content of the response, remediation and recurrence prevention on the company's own official pages as text (Google lists making important content available in text form as a fundamental for AI features)
• Place the description of the incident and the description of the response on the same page, without separating them
• State dates and points in time explicitly (descriptions from which it is impossible to tell what period is being referred to tend to generate errors about timing)
• Provide the information in a form that third parties can confirm as fact

This article does not hold measurements showing that these measures raise the probability of such joint description. As seen in §3-4, new material may also work in the direction of reinforcing an old understanding. Execute them as hypotheses, therefore, and confirm by observing after the measures are taken. If measures are stacked up without observation, whether they worked will never be known.

5-6. What not to do

We state explicitly what cannot be written. Within the scope of this review, we could not confirm any of the following.

The fact that a figure is presented does not mean that the figure is verifiable.

Following the argument of this article, the conclusion necessarily comes down to one.

1. Within the scope of this review, we could not confirm public research or a study measuring the decay period of the mention rate for corporate misconduct at a specific company (§0, §6)
2. The stability of citation sources differs greatly by platform: in AI Mode no convergence is visible even across 17 weeks, while in AI Overviews citation sources stayed fixed for more than half of the prompts (§3-1)
3. The stability of citation sources and the stability of the answer text are separate metrics, and neither can be uniquely inferred from the other (§4-2)
4. That the generation of AI answers involves the retrieval of external web information can be confirmed officially, but how this is integrated with trained knowledge is not published (§3-4)
5. Most misconduct is fact, so the framework of correction premised on the information being wrong is unavailable from the outset (§2)
6. And we could not confirm a general right for a legal person to have true but unfavorable information removed from AI answers either (legal instruments are in the companion article "Does the right to be forgotten reach AI answers?")

There is nowhere to find material for predicting the period.

The method for dealing with what cannot be predicted is the same as in other fields. Keep measuring, and detect change. Structurally it is no different from taking quality metrics continuously, or from tracking the incidence of workplace accidents.

Vaipm calls the approach of treating AI-space perception as an object of continuous management, rather than a one-off check, AI Perception Management (AIPM). AIPM is not an industry standard, but the framework Vaipm uses (What is AIPM).

The practical implication of this article comes down to a single line.

Add "we have observed and recorded how the incident is treated in AI answers" to the conditions for declaring closure.

That coverage has stopped does not mean that AI answers have changed. And whether AI answers have changed cannot be known without looking.

Q1. When will information from an online backlash disappear from AI answers?

Within the scope of this review, we could not confirm public data that answers this question. More precisely, we have not confirmed public research or a study that, for a specific instance of misconduct at a specific company, repeated answers over time under fixed conditions and measured the decay period of the mention rate. Descriptions citing a number of days circulate in the trade press, but none publishes the sample, the measurement procedure or the method of reproduction. Rather than estimating a period, we recommend switching to a design that observes and confirms.

Q2. If we have the original article deleted, will it disappear from AI answers too?

Not necessarily. First, how trained knowledge and retrieved results are integrated in an individual answer is not published, so disappearing from the search side does not necessarily mean disappearing from the answer. Second, Google states that recrawling and processing can take several days to several months, so the time until reflection varies. In addition, information remains on third-party surfaces such as news coverage, message boards, reviews and video.

Q3. Can unfavorable information that is true not be removed legally?

This article does not address that question, because examining legal instruments and rights is separate work from the time axis and observation that are this article's axis. The companion article "Does the right to be forgotten reach AI answers?" handles Article 17 of the GDPR, the Google Spain judgment, the 2017 decision of the Supreme Court of Japan, the CNIL analysis, the Act on the Protection of Personal Information and Article 50 of the EU AI Act, separating them instrument by instrument. What can be said within the scope of this article goes as far as one point: within the scope of this review, we could not confirm a general right for a legal person to have true but unfavorable information about itself removed from AI answers. Whether information that is wrong can be contested is handled by AI misinformation and legal liability.

Q4. I hear news articles are unlikely to persist as citation sources. Won't time solve the problem?

In the SISTRIX study, the proportion of news media domains among Google AI Mode citation sources that enter the stable core was 1.4%, the lowest value among the categories. However, this is a rate of persistence in the stable core, not a value that directly measures the citation lifespan of individual articles. It is also a figure for which sites remain as citation sources, not a figure for whether information stops being discussed. The same study shows that even where citation sources did not change across 17 weeks, the body of the answer changed from week to week in 87% of cases. The two are separate metrics, and neither can be uniquely inferred from the other.

Q5. I hear citation sources are fixed in AI Overviews. Is that reassuring?

It is not grounds for reassurance. The SISTRIX study does show that for 53% of AI Overviews prompts not a single citation source changed across 17 weeks, but the same study also shows that in 87% of the prompts whose citation sources were fixed, the body of the answer changed from week to week. That citation sources do not move does not mean that what is said does not move. If anything, a judgment that "it is fixed, so there is no need to look" leads to changes being missed.

Q6. What should we record, and how often?

What should be recorded is the engine and feature name, the model version, the date and time of execution, the language, the region setting, the login state, the full text of the prompt, the number of repetitions, and the full text of the answer. Storing a summary makes later verification impossible. The state before the incident occurs matters especially, since change cannot be identified without something to compare against. As for frequency, a single check is not enough. Even AI Overviews and AI Mode from the same provider differ in their citation sources in 83% of cases (Jaccard index 0.17), and between AI Mode and ChatGPT Search the value is lower still at 0.125. On top of that there is fluctuation from one execution to the next, so a one-off result cannot distinguish a change from a fluctuation.

Q7. If coverage stops, will AI answers settle down too?

Within the scope of this review, we could not confirm data showing that correspondence. Coverage volume and the treatment in AI answers are separate systems, and there are no grounds for inferring one from the other. Using coverage volume alone as the material for judging closure means declaring closure without confirming the state of AI answers.

Q8. On what basis should closure be declared?

In addition to the decay of coverage volume, add the condition that the treatment in AI answers has been observed and recorded. Specifically, observe under identical conditions across multiple engines; record whether the incident is mentioned, the content of the description, the citation sources, and the change since the previous observation; and keep it in a form that can be compared with the baseline state from before the incident. Frequency may be reduced after closure, but the practical approach is to define a period of stability and the recurrence conditions, then move to periodic monitoring.

Q9. If we issue a new announcement, will the old description be overwritten?

There is no guarantee. The study presented at ICLR 2024 (Xie et al., arXiv:2305.13300) reports that, in a controlled experiment, a model shows high receptiveness to coherent external evidence even where it contradicts parametric memory, while showing strong confirmation bias where the external evidence contains information consistent with parametric memory. This result cannot be converted directly into a persistence period in commercial AI search, but it does suggest that a simple model of "new information equals overwrite" cannot be taken for granted. Build a step into the design that observes and confirms after the announcement.

Q10. Is putting out large volumes of content for suppression effective?

We do not recommend it. There are no grounds for applying to AI answers, as is, the simple suppression model of conventional SERPs in which lowering the rank of particular URLs makes them invisible. And as shown in §3-1, at least in AI Mode and ChatGPT Search citation sources turn over week by week, so the structure works against the idea of filling the surface with volume. Manipulating reviews and word of mouth can breach the terms of the respective platforms, and if discovered the damage may exceed that of the original incident.

Q11. If we block AI crawlers, will we stop being displayed?

Not completely. OpenAI states explicitly that a site that has opted out of OAI-SearchBot will not appear in ChatGPT's search answers but may appear as a navigation link. It also states that a robots.txt update takes about 24 hours to be reflected. Furthermore, what can be blocked is only your own site; news coverage and third-party posts are outside its scope. Blocking also lowers the citability of your own site, so the side effects need to be considered.

Q12. Do the PR industry bodies have guidance on this problem?

It has begun to be addressed. CIPR published a best practice guide on the use of AI in PR practice on 10 June 2026, and PRsay, the blog of PRSA, published an article on the design of reputation monitoring on language models on 9 July 2026 (§5-1). Even so, within the scope of this review, we could not confirm any standard or guidance from a professional body indicating the persistence period of information about a specific instance of misconduct. The treatment remains at the level of "you should observe."

• Within the scope of this review, we could not confirm public research or a study measuring the decay period of the mention rate for corporate misconduct at a specific company
• Most misconduct is fact, so the framework of correction premised on the information being wrong is unavailable
• Four structures govern persistence: platform differences in the stability of citation sources, the absence of a correction process for legal persons, the instability of rewriting inside the model, and the fact that the integration of training and retrieval is not published
• The stability of citation sources differs greatly by platform. In AI Mode no convergence is visible even across 17 weeks, while in AI Overviews citation sources stayed fixed for more than half of the prompts
• Do not read citation-source churn data as evidence of the disappearance of information. The two are separate metrics, and neither can be uniquely inferred from the other
• We could not confirm a general right for a legal person to have true but unfavorable information removed from AI answers either (legal instruments are in the companion article)
• Move the design of practice from estimating a period to observing continuously, and add the observation and recording of AI answers to the conditions for declaring closure

Primary sources (official documentation)

1. Google Search Central, "AI features and your website" (query fan-out, no additional requirements for AI features, recrawling taking several days to several months, providing content in text form). Last updated 2025-12-10

https://developers.google.com/search/docs/appearance/ai-features

2. Google Search Central, "Remove a page hosted on your site from Google" (the removal tool's effect lasts about six months, methods for permanent measures, application limited to URLs the company controls)

https://developers.google.com/search/docs/crawling-indexing/remove-information

3. OpenAI, "Overview of OpenAI Crawlers" (robots.txt reflected in about 24 hours; a site opted out of OAI-SearchBot may still appear as a navigation link)

https://developers.openai.com/api/docs/bots

4. CIPR (Chartered Institute of Public Relations), "The Responsible Use of AI in PR" best practice guide, published 10 June 2026

https://newsroom.cipr.co.uk/cipr-publishes-new-best-practice-guide-on-responsible-ai-use-in-pr/

5. Advice from CIPR (PR practitioners should not directly edit clients' encyclopedia entries, but provide content and suggestions to the community; removal of vandalism excepted. Restated in January 2026)

Empirical research and large-scale observational data

6. Xie, J., Zhang, K., Chen, J., Lou, R., Su, Y., "Adaptive Chameleon or Stubborn Sloth: Revealing the Behavior of Large Language Models in Knowledge Conflicts," ICLR 2024 (arXiv:2305.13300)

https://arxiv.org/abs/2305.13300

7. SISTRIX, "AI Citation drift: How stable are sources in AI search results?" published May 2026. 82,619 prompts and 1,548,213 snapshots; six countries, three platforms, 17 weeks (2025-12-17 to 2026-04-08)

https://www.sistrix.com/blog/ai-citation-drift-how-stable-are-sources-in-ai-search-results/

Practitioner material and notes

8. PRSA / PRsay, "How to Set Up Reputation Monitoring in Language Models," 9 July 2026 (the three dimensions of awareness, attitude and attribution; the design of reputation monitoring on language models)

https://prsay.prsa.org/2026/07/09/how-to-set-up-reputation-monitoring-in-language-models/

9. The SISTRIX study is a first-party study based on data from the provider's own tool, and the figures need to be read separately by platform. It is aggregated at the domain level; at URL level, larger turnover is reported.
10. Primary sources for the legal instruments mentioned in this article (Article 17 of the EU GDPR, the Google Spain judgment, the Supreme Court decision of 31 January 2017, CNIL / EDPB Opinion 28/2024, Article 50 of the EU AI Act, the Act on the Protection of Personal Information) are collected in the sources list of the companion article "Does the right to be forgotten reach AI answers?"

• PR and the citation sources AI uses — where AI takes public relations information from (the parent article of this one)
• Does the right to be forgotten reach AI answers? — the companion article, from the side of legal instruments and rights
• AI misinformation countermeasures — the types and structure of incorrect information
• Correction and removal of AI misinformation — the question of whether correction is possible
• AI misinformation and legal liability — legal means and where responsibility lies
• What is AIPM — the approach of managing AI-space perception continuously
• What is AIO / What is LLMO — an organization of the terms