The data retention cliff nobody is pricing in

Google Ads is about to quietly cap how far back you can pull granular reporting data. Thirty-seven months for the detailed stuff, eleven years for aggregates. The announcement landed without much fanfare, the trade press covered it like a housekeeping note, and the industry shrugged.

It shouldn't have.

This change arrives at the exact moment two other shifts are eating the foundations of how we measure paid and organic performance. Keyword-level optimisation is being deprecated in favour of AI-driven bidding that nobody can fully audit. And the people building agentic campaign managers are about to discover that their training data — the historical record of what worked, when, and why — has a sell-by date.

The combination isn't a tooling problem. It's a continuity-of-evidence problem. And the industry is sleepwalking into it.

What's actually changing

Starting in June 2026, Google Ads tightens its data retention. Click-through rates, impression shares, search term reports, granular conversion paths — all the campaign-level forensics most consultants rely on — will only be available for the most recent 37 months. Aggregated annual data sticks around for 11 years, which sounds generous until you remember that aggregated annual data is roughly as useful for diagnostics as a tax summary is for understanding your shopping habits.

Three years of granular data feels like a lot. Until you try to do anything serious with it.

A genuine year-over-year comparison needs at least two clean years of baseline. Detecting a multi-year seasonal trend — the kind that separates structural shifts from one-off blips — needs more. Auditing a campaign restructure from 2023 to understand whether the decisions made sense in hindsight? You've got maybe eighteen months before that evidence walks out of the building.

For consultants who inherit accounts, this is worse. You arrive at a new client, ask for context on the last three years of campaign decisions, and the platform itself is now the limiting factor. The detail you need to audit somebody else's work has been quietly aged out.

The timing is the story

Data retention limits aren't new in tech. Platforms tighten and loosen storage policies all the time, usually for cost or compliance reasons. In isolation, this would be a Tuesday.

What makes it consequential is what else is happening in the same window.

Keyword-based optimisation is, by Google's own admission, becoming obsolete. The engineer who helped build Google's keyword system has been on the conference circuit explaining why. Smart Bidding, broad match, and the various Performance Max derivatives have moved decision-making inside the auction in ways that are functionally unauditable from outside. You can see what the system did. You can rarely see why.

At the same time, the agentic layer is arriving fast. AI systems are starting to manage campaigns end-to-end — adjusting bids, rewriting ad copy, reshaping audience targeting, killing underperforming assets. The pitch is that these agents will outperform humans because they can process more data, more often, with no ego attached to a previous decision.

That pitch has a quiet assumption baked into it. The agent needs data to learn from. Specifically, it needs *historical* data — the long tail of what happened in your account during conditions that no longer exist, so it can recognise when those conditions return.

An AI agent with a 37-month memory is an agent that cannot remember the last recession, the last algorithm shift, or the last category-defining seasonal anomaly.

That isn't a tooling limitation. It's an epistemic one.

Why nobody is talking about this

The simple answer is that data retention is boring. It doesn't generate hot takes. It doesn't pair well with a thought-leadership graphic. The people who'd notice — senior PPC consultants and analytics leads — tend to publish less than the people who don't.

An AI agent with a 37-month memory is an agent that cannot remember the last recession, the last algorithm shift, or the last category-defining seasonal anomaly.

The structural answer is that the industry has trained itself to think about measurement in quarterly horizons. Most agency reporting cycles never look further back than the last 12 months. Most in-house teams don't have the institutional memory to miss data they never used. When the cap arrives, very few people will notice the day it bites, because most accounts are operating on a horizon shorter than the cap itself.

This is the same dynamic that makes long-term marketing investment so hard to defend. Nobody screams when invisible infrastructure disappears, because the people who'd have used it have already left for another role.

There's also a comfortable assumption in the air that AI will somehow solve this. That if Google is the one capping the data, Google's AI must already have access to the underlying signal in some richer form, and therefore the bidding systems will be fine. That's probably true for Google's systems. It's emphatically not true for yours, your client's, or any third-party tool that depends on the Ads API.

The data the API exposes is the data you can audit. Anything Google retains internally and doesn't expose is, by definition, not part of your evidence base.

The audit problem this creates

I do a lot of inherited audits. Somebody calls because the previous agency lost the account, or the in-house lead left, or the numbers stopped making sense. The first thing I do is pull as much historical context as I can — because every account has a story, and the story is usually written in the decisions that look strange in isolation but make sense once you see what came before.

Three years from now, that work gets harder.

Treating data retention as a storage problem instead of an evidence problem. Your CRM has the last decade of customer data. Your Ads account, post-June 2026, will not. The asymmetry between what your business knows about its customers and what the ad platform knows about its own behaviour is about to widen sharply. That gap is where bad decisions hide.

Assuming third-party reporting tools will paper over the gap. They might, for accounts that signed up early enough and configured the integrations correctly. But most accounts didn't, and most reporting tools pull from the same Ads API that's about to be capped. If the source can't see the data, the tool can't see the data.

Trusting AI campaign-management agents to learn what they cannot remember. An agent trained on 37 months of an account's history will recognise the patterns that repeat within 37 months. It will not recognise the patterns that repeat every five years, or seven, or every economic cycle. It will treat those events as novel each time. That's not a clever system. That's an amnesiac with confidence.

What the agentic vendors aren't saying

The current wave of agentic ad-management tools is being sold on the basis that AI will optimise better than humans because it can hold more variables in mind at once. Fine. That's probably true within a given moment.

What's left unsaid is that these systems are dependent on the platform APIs for everything they know about the past. When the API stops exposing granular history, the agent's "judgement" — the part of the pitch that's supposed to justify the price tag — becomes structurally short-sighted.

The honest version of this pitch is: *our agent will optimise your account brilliantly for the last three years of conditions, and if anything older than three years matters, we won't know.*

That's not a damning admission. Plenty of marketing decisions are correctly made on recent data. The problem is the gap between what these tools claim ("comprehensive optimisation") and what they can actually see. When the gap is invisible to the buyer, the buyer ends up paying a premium for a service whose limits aren't on the spec sheet.

This is the same pattern that's played out across the AI search measurement stack — tools selling certainty about systems they can only partially observe.

What the honest response looks like

If you're running paid search of any consequence, the response is straightforward and unglamorous: get your historical data out of the platform before the cap bites.

The platform is treated as the source of truth, and the moment Google decides the source of truth is shorter than it used to be, the truth gets shorter with it.

This is not a sophisticated insight. It's data hygiene. But almost no agency I've audited has a structured archive of granular Ads data older than what the platform happens to be showing today. The platform is treated as the source of truth, and the moment Google decides the source of truth is shorter than it used to be, the truth gets shorter with it.

Concretely:

• Pull granular search term, keyword, and conversion path data on a rolling quarterly schedule into a warehouse you control.
• Snapshot account structure changes — campaign renames, audience changes, asset swaps — with timestamps, so the *why* of past decisions doesn't evaporate with the data.
• Stop assuming the API will be the durable record of your account's history. It won't be.

For consultants, this is also a service offering hiding in plain sight. Most clients haven't done this. Most clients won't do it themselves. Someone needs to.

The honest limits

I'm not predicting catastrophe here. Most accounts will keep running. Most agencies will keep reporting on quarterly cycles, the way they always have. The cap will arrive, nothing visible will break, and the industry will move on.

What I am claiming is narrower. The combination of shrinking data retention, declining keyword-level transparency, and rising agentic decision-making is creating a system where the long-term audit trail is being thinned out at exactly the moment AI systems are being given more authority. The people best placed to notice the thinning are the people least likely to be in the room when the decisions get made.

It's also possible that Google reverses course, extends the retention window, or offers a paid tier for longer history. Any of those would soften the problem. None of them are announced.

And it's worth saying: a 37-month window is still vastly more than what most platforms offered a decade ago. The comparison isn't to a perfect past. It's to the gap between what AI campaign management *promises* and what its evidence base can actually support.

The wider pattern

The protocol layer is changing underneath us. Schema is being deprecated. Rich results are quietly dying. The evaluation layer for AI search has moved into runtimes that SEO tools cannot inspect. And now the historical data layer for paid search is contracting too.

None of these are headline events on their own. Together, they describe a discipline where the substrate is becoming less inspectable, less durable, and less owned by the practitioner. The platforms are telling us what they think we need to see. The window into the rest is closing.

You don't have to panic about this. You do have to notice it. The agencies and consultants who'll still be credible in five years are the ones building their own evidence base now — because the platforms aren't going to do it for them.

That's the loop. And we're about to be locked out of half of it.