AI has big fat blind spots.

AI is very good at what it does but it has huge blinkers. It optimises whatever data it can see, and that data often doesn’t match where the business is heading. Any agency will tell you it uses AI, so that isn’t the useful question. Ask any AI marketing agency who keeps checking what the AI can’t see and keeps it pointed at your long-term plan.

Here’s what that looks like in a real account, the two reasons it goes wrong, what fixes it, and the question to ask.

The change nobody remembers making

Earlier this year our PPC team wrote about why Google Ads conversion rates are sliding. In it, they described taking on accounts where AI Max was already running and the marketing manager didn’t remember turning it on. Someone had clicked through a setup prompt three weeks earlier, and that was that.

AI Max lets Google go and find search terms beyond the keywords you chose, which means it decides where a share of your budget goes. Plenty of accounts were opted in through upgrade prompts, with nobody making a deliberate call to use it.

Another account we took on, for an Australian skincare brand, had conversion tracking firing on add-to-cart instead of completed purchase. Every number the account reported, and anything bidding toward those numbers, was aimed at the wrong thing.

In both accounts the tools did what they were set up to do, with the data and settings they had. Nobody was checking them against the business.

For any account using AI, start with a plain question: who signed off the last change, and what did they check it against?

AI tells the person using it they’re right

The first problem sits with whoever is using the tool. AI makes people feel more right than they are, and the more AI-savvy they believe they are, the stronger the effect.

Researchers at Aalto University had about 500 people work through logical reasoning tasks from the US Law School Admission Test. Half used ChatGPT and half didn’t. The people using ChatGPT did score better, but all of them overestimated how well they’d done. Those who considered themselves more AI-literate overestimated more. As Professor Robin Welsch put it, “higher AI literacy brings more overconfidence”. Most of the ChatGPT users rarely prompted it more than once per question, and often took its answer without checking.

The tools lean the same way. In April 2025, OpenAI rolled back a GPT-4o update that was “overly flattering or agreeable”. Its own explanation was that it had focused too much on short-term feedback. Anthropic’s 2023 research found the same tendency, known as AI sycophancy, across five leading AI assistants. It described sycophancy as a general behaviour of models trained on human feedback, likely driven in part by people preferring answers that match their views.

Ask a chat assistant whether to widen match types, lift a budget or switch on an automated feature. The answer can easily lean toward however you framed it, and you walk away more certain of a call nobody has checked against the business. It’s the same trap behind why telling AI it’s an expert makes it worse: confident errors are the ones that slip through review.

That applies to anyone making account decisions with a chat assistant open, agencies and business owners alike. An agency telling you it’s great with AI hasn’t yet told you who checks its work.

It optimises what it can see, and your strategy usually isn’t in the data

Too much weight on short-term feedback is also how an ad account drifts, and it’s the second, bigger problem. AI optimises the data it’s given, not your profit or your five-year plan, and it’s very good at it.

Smart Bidding weighs more signals in every auction than a person or team could compute. It looks at the device, the location, the time of day, the exact search, and whether someone is already on a remarketing list. All of that is pointed at one target: the conversions you’ve told it to count. The conversion goals you set drive the bidding, and misconfigured conversion actions stop Smart Bidding optimising effectively.

Google’s AI is like a dog bringing in a dead bird because it thinks its supposed to, you can’t get mad at it because you haven’t told it what it should be trying to do instead.

Google DeepMind’s researchers call this specification gaming: behaviour that satisfies the literal objective without achieving the intended outcome. They describe it as showing “the ingenuity and power of algorithms to find ways to do exactly what we tell them to do”. They also expect the problem to get harder as AI systems become more capable. The more capable the tool, the more its blinkers cost you.

Optimise purely to return on ad spend and the budget slides toward people already close to buying, through retargeting and brand search. One of Google’s documented bidding signals is whether someone added a product to their cart last week, because they’re more likely to buy soon. The dashboard improves while the pool of people who have heard of you shrinks. It’s the loop behind why your most profitable ads aren’t the ones that grow your business.

The data that matches your strategy might not exist yet

Long-term goals often leave no conversion data yet: becoming the go-to name in your category, moving upmarket, growing a newer service, winning a new region. The customers you want in two years aren’t clicking today.

The same blinkers show up beyond Google Ads. An AI trained on your current website learns the brand you have today, including the parts you’re trying to move away from. Optimise social creative purely to clicks and engagement and it drifts the same way, toward whatever performed this week, on-brand or not. Even an AI summary of your monthly report can only describe what was measured, which isn’t always what matters.

You can’t always get data that lines up with the long-term plan, so someone has to hold the plan.

Give it the whole business, or steer it

There are two ways to work around the blinkers.

Give it more of the business. You can theoretically push this as far as you want and give AI keys to the castle all the way to long term strategy decisions that your leadership has locked in. To start with though, feed it real outcomes. Google Ads lets you import qualified and converted leads from your CRM, so the system can learn which leads became customers. Then give your tools, and whoever sets them up, the context the AI can’t infer:

• your margin on each product or service
• which products or services you want more of
• how much work you can take on
• your positioning and brand guidelines.

Steer where the data can’t reach. Some of what matters will never show up in the data. A person who knows the strategy decides what the AI isn’t allowed to optimise away: the brand budget, the premium positioning, the new service. That person checks changes against margin and the plan, and overrules the numbers when the numbers and the plan disagree.

If you’re not sure what your own account is optimising toward, we’re happy to take a look.

What we do differently: a named person keeps checking what the tool can’t see

Every piece of work at Distl has a named person in front of it, doing it and signing it off. There are no AI agents tinkering away on client accounts behind the scenes. Our specialists all use AI, and with it they can check more of an account, more thoroughly, than they could on their own. Having a named person also means you can ask why any change was made and get the answer from the person who made it.

In Google Ads management, that starts before anything is built with an existing account getting an audit first, covering conversion tracking, campaign structure and historical performance. The automation will optimise toward whatever it’s told to value, so deciding what counts as a conversion sits with a named PPC specialist. Goals are tied to business outcomes, such as a revenue figure or a number of leads, and reset monthly or quarterly depending on the account.

From there, a specialist who works only in Google Ads reviews the account constantly. Health checks cover spend, impression share, cost per lead or sale and return on ad spend. Search terms get the same attention, audited regularly with irrelevant queries excluded. Each month the specialist records what changed and why, and reports it back to you. The monthly report answers four questions: are you on track to the goal, what can be adjusted, what was actioned, and what did it do?

The account can still use Google’s automation. A person who knows your goal reads what it did and decides whether it fits the plan.

Should you just run it yourself?

For some businesses, yes. The person who knows the strategy best is often the owner.

If you’re a local service business on one or two Search campaigns and you know your margin and enquiry-to-job rate, run it yourself. Use AI as your analyst and give the search terms report an hour a week. Let AI find things and make the calls yourself.

If you’re running Performance Max or several campaign types, hand the checking to a specialist. The same goes if you sell product lines with different margins, split budget between brand and performance, or won’t look weekly. Most of the job at that point is checking what the automation did, that’s what you’d be paying an agency for, so judge them on it.

Volume matters too. If your campaigns sit under about 30 conversions a month, lean on what you know about the business. Google recommends judging Smart Bidding results over periods with at least 30 conversions, such as a month or longer (50 for Target ROAS). Below that, there’s little for the algorithm to learn from.

Those lines are where we’d draw them today, and they’ll move as the tools get beefier.

The questions to ask any AI marketing agency

• What is the AI optimising toward in my account? A specific answer names the conversion action and how it connects to your margin and where you want the business in two years.
• What doesn’t it know about my business? Think margins, priorities, capacity and brand goals. A specific answer names what’s missing and who fills it in.
• Who signed off the last change, and what did they check it against? A specific answer gives you a name, the change and what it was checked against.

The shape of the answer is the test. If the answers stay vague and lean on “the algorithm”, that’s the part of your marketing to look at first.

Put the same question to us, too. Whichever AI marketing agency you end up with, the answer should name the person checking your account and the last thing they changed. If you’d like to hear how we’d answer it for yours, let’s talk.

Frequently asked questions

What is an AI marketing agency?

Any agency that uses AI tools in its ads, content or reporting work. That now covers plenty of agencies. The tools themselves are widely available. The difference worth paying for is who checks the AI against your business, and whether you can put a name to that person.

Can AI run Google Ads on its own?

It already makes plenty of decisions by itself. Smart Bidding sets bids in every auction, and AI Max goes and finds search terms beyond the keywords you chose. Both optimise toward whatever conversions you track. Neither knows your margin or your plans unless someone gives it that context, so a person still needs to check them against both.

Jack Headford

Account Manager & Operations Lead