Skip to content

Q&A[2026-11-16]

How Data Teams Make Better Data Retention Decisions

IMInformatics MagazineNovember 16, 2026

Keeping every record forever can create more risk than value. This article shares expert insights on matching retention to real business use, protecting privacy, and preserving the data that matters most. Learn how clear policies, tiered storage, and timely deletion help data teams make smarter retention decisions.

§ How this was made

The perspectives in this article were written by the named contributors and submitted through Connectively. Connectively's automated system selected them and used AI to write the headline, introduction, section headings, summary sections, and illustration, and to make minor edits for clarity. Articles are assembled and published automatically. Report errors to corrections@connectively.us. How we make our content

Fig. 1 · Q&A · 2026-11-16

Illustration generated with AI

FKFKRAJHZBOMOKJB
In [*] · 25 contributors
FKFahad KhanDigital Marketing Manager, Ubuy Sweden

Protect Irreplaceable Events, Put Privacy First

I ran data platforms for nine years, and retention policy was always the thing we'd deal with next quarter.

The decision I'd describe is one we got wrong in the cautious direction. We deleted a few years of event data because storage costs were climbing and nobody could name a use for it. Defensible at the time.

Eighteen months later we needed to understand a seasonal pattern that required more history than we had, and the analysis simply couldn't be done.

What I learned is that the question isn't what the data is worth now. It's whether it's reconstructible. Anything you can regenerate from a source system can go cheaply. Anything that only exists because you captured it at the time is irreplaceable, and that distinction matters far more than volume.

Privacy cuts the other way and has to come first. If you shouldn't be holding it, reconstructibility is irrelevant.

FKFaizan KhanPR and Content Marketing Specialist, Ubuy Singapore

Match Retention to Actual Use

I determine how long data is stored based on what decision the data supports and when that decision expires. If there is no known use, the data receives minimal retention, since keeping the data incurs cost and poses privacy risks without providing any value.

The decision that taught me the biggest lesson was one concerning raw clickstream data. We kept the raw data for five years "just in case," and it accumulated to 90 TB, costing approximately $2,700 per month to store. When we conducted an audit as to how frequently people asked questions using those logs, 94 percent of all questions covered only the last 13 months.

We instituted a new policy that we kept detailed information for 13 months and a summary after that. Storage costs were reduced by 71 percent, privacy review times were cut in half, and no analysis required the deleted detailed information.

Lesson: retention of data should be determined by its use rather than by fear of deleting it.

Define One Current Window

Raw rows and decision rows get different lifespans, and only one window is allowed to be current.

When we onboard a client we write a month-one baseline from Search Console and freeze it. The file states its own job: it is "the frozen Month-1 state every later report (Month 4, Month 6) compares against; it is never edited, a re-run writes a new version." Small file, derived summary, kept as long as the client exists.

Underneath it sit three pulls with three purposes. Current work uses the "last 14 days: 2026-08-30 to 2026-09-12", which came to 24239 query by page by country rows. Stability uses the last 28 days, 33977 rows. Trend, which I never use for a ranking claim, uses the "last 16 months by day", 480 days with data.

The retention call follows the purpose. The 14 day and 28 day pulls can be re-pulled whenever anyone needs them, so old copies earn nothing. The trend pull is the one worth storing, because no API hands those days back once they age out of the interface. The summary is kept permanently, because it is the only thing a month-six report has to compare against.

The lesson I now apply well beyond analytics came from the window rule: one window is current, the rest are context. Teams that keep everything end up holding several numbers for the same question and no way to say which one counts. Writing down which pull answers which question did more for us than any storage policy.

Privacy pushes the same direction. Query rows tied to a page and a country are useful in aggregate and uninteresting per person, so we keep the aggregate and drop the detail instead of debating how long to hold material nobody analyses.

One habit that costs nothing: the baseline names the script that wrote it and the end date of its data. A kept file with no provenance is a liability, because a year later nobody can tell whether it is still comparable to what you are measuring now.

JHJens HagelFounder & Owner, kjello

Set Policies Before Urgency

When I went back through three of our recent infrastructure analyses at mid-sized firms, none of them had a written deletion policy. A creative agency merging into a larger group wanted to export all its SharePoint data to an 8 TB disk and leave it in the server room. We advised keeping the data reachable and backed up instead, and we don't cancel licences on our own, because cancelling is a decision. My rule is to set retention per type of data before anyone urgently needs it.

Favor Aggregate Metrics, Purge Digital Noise

Managing over $100M in ad spend across 200+ companies has taught me to separate aggregated performance signals from granular user-level logs. I retain macro attribution and lifetime value trends long-term to train machine learning models, but cycle out granular session identifiers to respect user privacy and eliminate data bloat.

A major turning point came when bloated historical session data in Google Analytics began skewing conversion reporting without offering any predictive value. We purged the granular user logs and shifted our focus to syncing only qualified bottom-of-the-funnel milestone data into HubSpot.

That decision taught me that hoarding raw digital breadcrumbs never drives revenue--it just creates noise. Today, we anchor data retention strictly to measurable attribution and actionable KPIs, purging anything that no longer informs campaign performance.

OMOscar MoncadaCo-founder and CEO, Stratus10

Tier Older Records by Access Needs

Retention starts with what the business or its customers are required to keep. In regulated industries, the framework often decides for you: financial records, healthcare data, and audit logs can carry 5, 10, or even 30-year requirements. Where there's no regulatory driver, a lot of data doesn't need to be kept at all if nobody will use it.

The question we ask is how much of the data will be queried later, and what it costs to keep all of it in fast, expensive storage indefinitely.

The decision that taught us the most was choosing tiering over hard cutoffs. Instead of deleting data on a fixed date or keeping everything at full cost, we move aging data into progressively cheaper storage tiers. It stays available for later analysis without burning budget, and keeping less data in active storage shrinks the blast radius if something goes wrong. I now apply that thinking broadly: decide what needs to be fast, what needs to be kept, and what can go, instead of treating all data the same.

Classify Sensitivity, Expire Identifiers

As CEO of Netsurit supporting IT and cybersecurity for hundreds of client organizations globally, I balance analytics and privacy by cataloging and classifying data assets strictly by sensitivity. This ensures we only store high-risk records for the exact duration needed to meet compliance frameworks like GDPR or HIPAA.

A key lesson came from managing hybrid infrastructure backups, where holding personal information longer than required created compliance liability without delivering actionable analytical insight. We established a strict rule never to retain personal data beyond its defined operational purpose and implemented automated controls using Microsoft System Center to manage our lifecycle and backup policies.

We now separate anonymized operational data for long-term analytics while routinely destroying sensitive personal identifiers once their specific purpose expires. Conducting regular risk assessments to purge stale data keeps cloud costs down and prevents unnecessary exposure.

Prioritize Strategic Signals Over Operational Noise

I work at the intersection of AI, marketing systems, and operations for home service contractors, so data questions aren't abstract for me--they directly affect how we build and advise on tech stacks for real businesses.

The lesson that changed how I think about this: we were holding onto granular campaign-level data far longer than it was actually useful. After a certain point, hyper-specific click data from two or three years ago tells you almost nothing actionable about today's buyer behavior--especially now that AI-driven search has fundamentally shifted how customers find contractors. Old data wasn't an asset; it was noise with a storage bill.

What I apply broadly now is a simple filter: does this data connect to a decision I'll realistically make? For us, that means retaining aggregated performance trends long-term while letting granular personal interaction data age out much sooner. When we rebuilt our internal systems at CI Web Group, we separated "strategic signal" data from "operational noise" and treated them with completely different retention logic.

The privacy side is where most small business owners get complacent. Holding customer contact and behavioral data indefinitely because "it might be useful" is exactly the mindset that creates liability without value. Build your retention policy around what decision it enables--if you can't name that decision, you probably don't need the data.

LKLokesh KaranamStaff Software Engineer, RR Donnelley

Limit Versions, Tier Excess Logs

We keep customer data for one primary goal. Analytics and machine learning models and predictions. That is what the data is for.

You have to be diligent with it. The data must stay secure. And you must delete customer data as fit, on the customer's approval. Right to be forgotten is real. We follow GDPR and CCPA rules.

The problem starts when data grows. You lose track of what is in the dataset. You cannot say if it is up to date or not. And it is not just the data itself. You also keep logs and replicas. In the cloud, every rewrite keeps versions of the same data again and again. That is what drives cost.

In our case we had more than 500 terabytes of data. AWS S3 storage cost us over $12,000. When we looked deeper, most of it was logs and repeated versions of the same data.

The lesson was simple. Keep only the versions you need. We enabled lifecycle policies that keep only the versions we need and move the rest to lower tiers like Glacier. Retrieval is slower, but the data stays with us at a lower cost. That cut our cost by 70%. It also let us delete unnecessary data as we go.

We apply the same rule to our machine learning runs. Once a model run is complete and the results look good, we do not need that input data anymore. We only need the predictions. New data comes with the next run. So we have a policy that deletes the run data in about seven days.

That keeps us privacy compliant and keeps our costs in check. Retention is not about keeping everything. It is about keeping what you need, for as long as you need it, and deleting the rest on a schedule.

KSKruno SulićFounder & SaaS Product Builder, Cliprise

Roll Up Trends, Expire Identities

When future analytical value has to be balanced against privacy risk and storage cost, I start by separating what is truly needed for operations from what is only potentially interesting later. My rule is simple: keep identifiable raw data for the shortest period that still supports the product, support, fraud review, and essential business analysis. After that, convert what I can into aggregate, de-identified, or event-level summaries that still let the business learn without holding unnecessary personal detail indefinitely.

As a founder working on web products and consumer privacy education, I have learned that the most expensive data is often not the data you pay to store. It is the data you have to govern forever. Every extra field creates future obligations around access control, deletion, exports, incident response, and explaining why you kept it in the first place.

One retention decision that taught me a broader lesson was keeping more detailed logs than I actually needed because they might be useful later for product analysis. Some of that data did help in the short term, but over time it became clear that the long-tail value was much lower than expected, while the privacy and operational burden kept growing. The lesson was that "maybe useful someday" is not a retention policy.

What I apply more broadly now is a tiered approach. First, define the real business question the data supports. Second, assign a default expiration date to raw records at the moment they are created, not months later when cleanup becomes difficult. Third, preserve trends instead of identities whenever possible by rolling older data into summarized datasets. And fourth, review retention regularly because products change faster than policies do.

In practice, if I cannot explain in one sentence why a specific category of personal data still needs to exist after a certain date, I treat that as a sign it should be deleted or reduced. That approach usually improves privacy, lowers cost, and forces better analytics discipline at the same time.

Kruno Sulic Founder, LookupPlainly

Trust Source Systems Over Analytics

I keep data for as long as it answers a question I will ask again, and no longer. Most businesses do the opposite. They keep everything because storage is cheap and deleting feels risky, then find the hidden cost later: a pile of personal data they have to protect, and reports so cluttered nobody trusts them. For marketing data the practical split is between raw, identifiable records and the aggregated figures that drive decisions. The first needs a short, written retention period. The second can be kept for years, because monthly revenue by channel holds no names.

The lesson that changed how I work came from analytics audits. Across the 240 GA4 properties audited for my GA4 adoption study, reported conversions diverged from the backend order or CRM system by 14.7% on average (https://visionary-marketing.co.uk/blog/ga4-adoption-statistics-2026). That told me the long history people were reluctant to delete was often the least reliable data they held, and five years of a flawed number is a liability with a nice chart on top.

These days I decide retention by asking which system is the record of truth. The order system or CRM holds what happened, and that is what I protect and keep. Analytics tools are for spotting patterns, so I keep their summaries and let the granular event data age out. It costs less, carries less privacy risk and leaves a cleaner picture for whoever reads it next.

Reduce Workspace Access Before Deletion

Shared operating environments create a retention problem about boundaries, not volume. Data used to coordinate delivery can outlive the permission context in which it was shared, particularly across agency teams, client stakeholders, and temporary contributors. I tie retention to role-based access reviews, not calendar dates. A legitimate record may still need a smaller audience.

An internal review found inactive project spaces containing material no longer sensitive alone, but revealing in combination. Deletion mattered, yet access reduction mattered first. Every closure process now includes archive classification, membership removal, and a check for derived reports. This prevents dormant workspaces from becoming informal repositories with unclear accountability.

Separate Customer Patterns From Individuals

Keep data for as long as it changes a decision, and no longer. For a small online shop the honest test is whether anyone would look at a record again and act on it. If not, it is costing storage, cost and risk with nothing coming back.

The data that made me think hardest about this was the vehicle registration lookups on my site. A driver types their number plate, and the tool reads their charge port, maximum charging rate and phase from DVLA data, then sends them to the right cable. That history is useful in aggregate, because it tells me which cars people own and which cables to keep in stock. The individual plate attached to a visit is far less useful, and it is the part I would least want to explain if someone asked why I still held it.

The lesson I apply more broadly is separating the pattern from the person. The sensible approach is to keep counts of models and cable types for planning, and to set a short lifespan on anything that identifies an individual unless it is tied to an order, a warranty claim or something the law requires me to keep. Order records are different, because a warranty claim can arrive long after the sale. Everything else can be summarised and deleted on a schedule, so nobody has to remember to do it.

RMRichard MeadowsHead of Content, Streamrise

Document Every Backup Exclusion

We decide retention by asking two separate questions: what would we need in order to rebuild the business, and what do we keep only because it might be interesting one day. The answers usually have different lifespans, and mixing them is how a business ends up paying for data it never reads while missing the data it needs.

The decision that taught me this was a backup scope choice. Our nightly database dump left out several high-volume tables. That looked fine in daily reporting, because the job ran and the file arrived. But it meant the only complete copy of the database was a live replica, and a replica faithfully copies a mistake within seconds. A scope exclusion had quietly become a recovery decision that nobody had actually taken.

On 30 October 2026 we added a full physical backup with continuous archiving of the database's write-ahead log. The first full backup took 11 minutes 22 seconds and compressed 455.4 GB down to 120.5 GB. We keep one full backup plus the log archive, so storage stays predictable. One honest caveat: we have not yet run a test restore, so I would call this measured for backup and unproven for recovery.

The rule I now apply more broadly: every exclusion from a backup or a retention policy should be written down as a decision with an owner, in the words "if this is lost, we accept X". If nobody will sign that sentence, the data is not optional, whatever it costs to store.

Preserve Context for Defined Queries

I set a retention period around the decision the data has to support. For each dataset, I ask who needs it, what question they will answer and when older records stop changing that answer. Then I weigh the value of keeping them against the privacy exposure and the cost of storing, securing and retrieving them. Indefinite retention has no technical justification when nobody can name the future use. Legal or contractual obligations may impose a separate minimum, which the product policy must account for before records are deleted.

At Ronas IT, we chose to keep the full history of conversations with inbound leads in our own data warehouse. Leaving fragments in separate platform inboxes would make a later AI step reconstruct the conversation through each interface. A consistent history gives that step the sequence it needs to read. The retention lesson is about preserving context: one message may seem expendable on its own while its place in the conversation makes the rest intelligible. We keep this history for a defined downstream use; that choice doesn't assign the same lifetime to unrelated datasets.

For another system, I would begin with the query the team expects to run and test how far back it has to look to inform a decision. Keep the fields needed for that query, separate out identifiers it doesn't require, and restrict access to the retained history. The owner of the retention rule needs a review date so the team can revisit storage cost and privacy exposure as usage changes. If older records no longer affect a decision, shorten the window or reduce the data retained.

NSNishanth SirikondaCloud Solutions Architect, FirstDay Foundation

Save Rule History, Restrict PII

I start by splitting the data into two piles: the history of rules and decisions, and the personal details about people. They deserve very different retention answers.

Personal details get the shortest life the law allows. Take I-9s, the forms that verify someone can work in the U.S.: federal rules require keeping each one three years after hire or one year after the person leaves, whichever is later. Automating that process across nine nonprofits taught me that holding forms past that date protects no one. Storage is cheap, but every extra year is more ID details exposed if there's a breach.

The rules history is the opposite. When I redesigned union pay at Sunbelt Rentals, every rate went into a table with the date it took effect. That record holds almost nothing personal, and it's what answers a grievance years later: what rate applied on that date?

Keep the history of your rules for years. Keep personal details only as long as you truly need them.

For analytics, most trends survive without names, so strip out the person before you archive. Happy to share more.

NGNoah GulaAVP at OSP, OSP

Demand Questions Before Telemetry Storage

We used to assume that more historical data meant better models, so we kept years of raw system interaction logs on the chance we'd need them. A security audit and a cloud cost review changed that. We were paying thousands of dollars a month to store unindexed telemetry that no data scientist had queried in 18 months. In a breach, that archive would have been pure liability with no business value to offset it.

That experience gave us a rule we now apply to every data architecture decision. Before we keep a dataset, someone has to name the question it will answer and who will ask it. If an aggregated feature or schema can answer that question, we don't store the raw data. Our pipelines summarize and anonymize datasets on a schedule, so raw records age out once their useful signal has been captured.

Storage costs fell, queries run faster, and a breach would now expose far less.

Anonymize Milestones After Graduation

Managing Fantasy Island Schools for Kids since 2009 with a background in psychology, I balance data retention by separating sensitive personal records from long-term developmental observations. We hold identifying documents strictly through active enrollment and regulatory compliance windows, removing them once that operational relationship ends.

The clearest lesson came from managing years of student milestone portfolios and observational tracking notes. Keeping full, personally identifiable family files indefinitely increased our privacy exposure and storage footprint without offering better educational insight.

We now scrub all identifying information at graduation, archiving only anonymized developmental milestone data to assess our gifted curriculum trends over time. Purging private identifiers immediately while keeping raw analytical metrics protects family privacy and maintains rich data for future program planning.

Follow Product Demand, Not Speculation

We sort data by what it's about before deciding how long to keep it. Market data and company filings can be fetched again, so we keep them for as long as something in the product reads them and the cost makes sense. Data about people can't be rebuilt and comes with privacy duties, so we keep as little of it as we can, for as short a time as we can.

The decision that taught me this was price history. Until Sept. 2, 2026, we fetched ten years of daily prices for every company we added. We measured our own tables: a decade cost about 2,436 price rows per company, and our data vendor bills by usage. Then we checked what actually reads those rows. Our charts ask for 365 days, and none of our valuation models looks further back than a year. Ten years was paying for nine we didn't use. We cut new history to five years, which still leaves four years of headroom, and kept one exception on purpose: the market index we benchmark against keeps ten years, because our study of analyst calls measures returns against it back to December 2011.

The lesson I now apply everywhere: don't ask how long data might be useful someday. Ask what reads it today, and keep a margin above that.

It changed our backups. The production database is copied off the server every Sunday, and we keep eight weekly copies rather than all of them, because an unlimited archive would quietly keep data that users have asked us to delete. It changed our logs, too. We log which crawlers visit the site, but each line records the bot, the time and the page it asked for: no IP address, no query string, and nothing about human visitors. If we can't say what a field is for, we don't keep it.

BRBranden RobinsonVice President of Business Development, CortiCare

Attach ROI to Clinical Retention

Two decades scaling healthcare tech companies--including building out nationwide tele-EEG infrastructure at CortiCare--means I live at the intersection of clinical data value, compliance obligations, and operational cost daily. That context shapes how I think about retention practically, not theoretically.

The lesson that changed my thinking came from watching continuous EEG monitoring data accumulate across ICUs and hospital networks. The analytical signal in that longitudinal data is genuinely valuable--outcomes research, utilization trends, protocol refinement. But storing raw patient-level neurodiagnostic data indefinitely without a clear downstream use case is expensive and creates real HIPAA exposure. We learned to ask "what decision will this data inform, and by when?" before defaulting to "keep everything."

That question now drives our retention logic: aggregate and de-identify data that has served its primary clinical purpose, retain identified records only as long as active care relationships or contractual obligations demand, and document the reasoning. The discipline of forcing that conversation early--before the data lake gets unmanageable--is the habit I'd push anyone to build.

The broader lesson I apply everywhere now: data you can't articulate a future use case for isn't an asset, it's a liability. Treat retention decisions like you'd treat any resource allocation--with a clear ROI threshold attached.

Write Policies Around Legal Exposure

My background is in both Chemical Engineering and business transactional law, so I think about data the way engineers think about materials: you keep what has load-bearing value, and you dispose of what creates liability sitting on a shelf. The practical answer is to match retention periods to your actual legal exposure windows -- in Texas, most contract claims carry a four-year statute of limitations, so that's a natural floor for anything touching a business relationship.

The lesson that stuck with me came from a collections matter. A client had purged invoicing records before the limitations clock even started running, essentially destroying their own sworn account claim before they knew they needed it. We salvaged what we could, but it was a painful reminder: data you think is "done" often isn't done legally.

The flip side I've seen equally often is clients hoarding everything indefinitely because it might be useful someday. That creates discovery exposure in litigation -- the more you've kept, the more opposing counsel can dig through. A written retention policy with scheduled purges is actually a litigation defense tool, not just a housekeeping exercise.

The broader principle I now apply everywhere: treat your data retention policy like a contract. Put it in writing, tie it to specific business and legal purposes, review it annually, and make sure someone is actually responsible for executing it.

Store Daily Snapshots, Expire Feeds

At SellerQI we analyse Amazon seller accounts, and the retention question is partly forced on us by Amazon itself. Advertising API report data only goes back a limited window, roughly 60 to 95 days depending on the ad type, and Seller Central doesn't keep every report forever either. If you don't store the data yourself, it's gone, and so is your ability to answer "why are sales down compared with last Prime Day?"

So we split data by what it's for, not by how big it is.

Performance data with no personal information in it (daily sales and sessions by ASIN, ad spend and clicks by keyword, Buy Box percentage, inventory levels) we keep long term. Stored in aggregated form it's cheap, and its analytical value grows with time, because seasonality and year over year comparisons are where most root cause analysis happens.

Anything that touches a buyer, like order level records with names or addresses, gets the opposite treatment. Amazon's Data Protection Policy for developers already limits how long that information can be held, and we don't need it for analysis. We aggregate what we need and drop the personal fields early.

The lesson I now apply everywhere: the thing you regret is rarely deleting raw data. It's not having saved the daily snapshot. Rolled up daily figures at ASIN and keyword level answer nearly every question that comes up months later, at a fraction of the storage cost of raw logs. So for any new data source, our first question is "what is the smallest daily summary that would let us explain a change a year from now?" We keep that indefinitely, give the raw feed a short, defined expiry, and keep personal data only as long as the task in front of us needs it.

VRVishnuvardhan Reddy KaithapuramSoftware Development Engineer, Amazon

Use Verified Outputs, Release Inputs

I decide how long to keep business data by starting with what the data will actually be used for, not by keeping everything just in case. If a dataset supports reporting, billing, or a decision that may need to be revisited, it needs a longer window and clear ownership. If it is raw input that has already been cleaned, checked, and turned into a trusted dataset, keeping the raw copy for a long time adds cost and privacy risk without adding much value.

In my work on data pipelines at Amazon Advertising, I have seen this trade off play out in practice. We ingest large volumes of data through services I help build and maintain, and the question is always which parts need to stay close at hand and which parts can be summarized, archived, or deleted on a schedule. My rule is simple. Keep the trusted, well described output that teams actually query. Keep the raw input only long enough to reprocess it if a check fails or a source changes. After that, retention should be a decision, not a default.

That decision should also be written down in plain terms. Who owns the data, why it is kept, how long it is kept, and what happens at the end of that period. When those points are not written down, data tends to stay forever, and nobody wants to be the person who deletes it. When they are written down, teams can agree on a shorter window because they know what they will lose and what they will keep.

The lesson I apply more broadly came from keeping data longer than we needed to. A longer window felt safer at first, because it seemed to leave every option open. In practice it made the data harder to trust. Older records were mixed with newer definitions, and it was harder to explain where a number came from or whether it should still be used. Cutting the window back, and keeping better descriptions of what remained, made the data easier to defend and easier to use.

Cost matters here too, but it is not only storage cost. It is the cost of checking, securing, and explaining data that nobody is using. Privacy matters for the same reason. If data is kept without a clear purpose, it is harder to protect and harder to justify. A shorter, purposeful retention window, reviewed on a regular schedule, gives teams a dataset they can actually stand behind. That is the standard I use now. Keep what has a use, describe it well, and let the rest go on time.

DGDaniel GohChief Marketing Officer, RankAISearch

Capture Baselines Before Site Changes

1. I no longer worry about storing too much information. There are only a few instances in which I need to store small amounts of information to be able to prove that something worked. Google Search Console only stores data for 16 months on Search Console. As such, on a monthly basis, I create a simple text file that stores click and impression data for all queries for all pages on a website. For RankAISearch's free AI Visibility Checker, we store the information that each check found, i.e., did we find the business name and did we find any citations to the business on any sites. I don't store any more information about a person than is necessary. Will this information help to answer a before versus after type question? If so, then create a simple summary of the information that you've found and be done. If not, then it's not necessary to store the information and therefore you shouldn't.

2. I remember changing title tags and internal links on a website and going back to the analytics after a while to check if there were any positive results. However, Search Console data had already moved on from the 'before' period and I had not saved a copy of the data. I had a feeling that things had improved but there was no concrete evidence to back this up. I now write down the date, the change(s) made and a copy of the relevant metrics before making any changes. This has become second nature to me and I now apply this to all audits that I carry out, including the recent audit of the RankAISearch website where we reduced a 2MB page to 330KB. I'm glad I kept a note of the 'before' numbers. Keeping less but keeping the right information and labeling it and dating it is so much better.

Move Archives Beyond Retired Systems

I start from the question the data must answer, not from the data. Every record class gets three dates: how long the law requires it (tax, audit, employment and contract rules), how long the business realistically asks questions of it, and the point where it holds personal information nobody has a reason to see anymore. The retention period is the longest legitimate need, not the longest period storage allows. Beyond that, the default is delete or anonymize.

Cost rarely comes from storage. It comes from keeping data inside the system that created it. In our work archiving ERP systems, we regularly see companies paying for full licences of a retired system for years, only because finance may need to look up an old invoice during an audit. The data was worth keeping. The system was not.

The lesson I now apply broadly: separate the retention decision from the system decision. Decide what must be kept and for how long, then move it to a low-cost, read-only, queryable store with its own access controls and a deletion date per record class. That changes the trade-off in three ways. Analytical value survives, because history stays queryable in open formats instead of sitting in a system nobody can log in to. Privacy improves, because a read-only archive with narrow access and audit logging exposes far less than a live system carrying years of accumulated user accounts. And cost falls to the level of storage instead of licences.

A second lesson came with it: "keep everything, just in case" is not a neutral choice. Every year of retained personal data is a year of breach exposure and of data-subject requests that must be answered. So we ask for an owner and a reason for each data set. If nobody can name the question it will answer, or the rule that requires it, it gets an expiry date. Aggregates and anonymized history usually carry the analytical value; the identifiable detail rarely does.