How to reduce AI hallucinations in marketing

Hallucinations are the biggest blocker to AI-first reporting. Here's how marketers reduce them by grounding Claude in first-party data, with lessons from Layer, a marketing agency that cut reporting time by 30x.

Your client’s quarterly catch-up is going well. The deck you’re presenting is detailed and well-designed — all thanks to AI. The client seems to appreciate the rigor.

Suddenly, they spot a figure: a surprisingly low cost per click (CPC) on one of their campaigns. This kicks off an intense discussion, and before you know it, they’re using this single datapoint to rethink next quarter’s strategy.

It’s one thing to present AI-generated numbers to your clients. But when they start allocating huge portions of their budget based on data you copy and pasted without checking, you might start to feel slightly uneasy.

We often speak to marketers (both in-house and at agencies) about adopting AI-first workflows. Most share our enthusiasm — but they’re often worried about the same thing: hallucinations. According to the 2026 Marketing Data Report, just 1% of marketers have complete trust in AI.

This article takes a look at how marketers can build trust in AI workflows. It examines why hallucinations happen in the first place, how feeding Claude your own data reduces hallucinations, and how to get consistent, repeatable results using prompts and skills.

We’ll also share insights from our June 2026 webinar with Layer, an agency that used the Supermetrics for Claude integration to cut their reporting time by 30x.

Key takeaways

Why do AI tools hallucinate?

AI tools hallucinate (i.e. make things up) because they don’t know what’s true and what’s false. They’re prediction engines, not fact-checkers.

LLMs are built to guess what the next word in a sentence should be, based on their analysis of trillions of words that humans have written. Often, they get this prediction right — but sometimes they don’t. That’s what we term “hallucinations”.

In other words, hallucinations don’t happen because an LLM has gone off the rails (though it’s more likely to happen in certain situations, like when its context window — its working memory — is nearly full). It’s simply doing what it always does. The process is the same, yet the answer is wrong.

That’s why Andrej Karpathy, one of the co-founders of OpenAI, said: “I always struggle a bit with [when] I’m asked about the ‘hallucination problem’ in LLMs. Because, in some sense, hallucination is all LLMs do… Hallucination is not a bug, it is LLM’s greatest feature.”

What’s more, AI tools hate leaving blanks. They’ve been engineered to provide complete answers — even if this means guessing.

That means the answer to reducing AI hallucinations is relatively straightforward: reduce the guesswork. Stop leaving gaps for the AI to fill.

Does feeding Claude your own data make it more accurate?

Yes, feeding Claude your own trusted data makes it far more accurate. If you give it the correct data, it’s more likely that it will give you the correct answer. This is especially true if the data is presented in a format that’s easy for Claude to analyze. It can easily arrive at the answer rather than having to use significant tokens to untangle the numbers, which sometimes leads to it guessing the answer.

Take Supermetrics for Claude.

Supermetrics for Claude connects data from your ad platforms directly to Claude, via Supermetrics. It feeds Claude real-time data straight from platforms including Google Ads, Meta Ads, Microsoft Advertising, and TikTok Ads.

When you ask a question about your campaigns — like how much you spent or what the results were — it pulls straight from your own first-party data. It analyzes your numbers, but it doesn’t have to source them.

Morten Kleven, AI and Automation Lead at Layer, recently spoke about how his agency has been using Supermetrics for Claude to speed up their workflows. He offered a few tips based on Layer’s experience to help marketers get started with the integration:

• Start small, check, then scale: Morten started by connecting just seven days of Google Ads data and three metrics. He asked Claude questions and checked the results it provided. Once he saw that everything was correct, he then scaled up the number of connected data sources.
• Ask specific questions: Asking a narrow, well-defined question is far more likely to generate an accurate answer than a non-specific, sprawling one. It ensures the AI focuses only on the relevant facts.
• Be wary of data limitations: Some platforms cap how far back their data goes. For example, Morten ran into API limits from Meta and Snapchat when pulling older year-on-year comparisons. To fix this, he asks Claude what the connector can pull and where the data ends. If there’s a gap, he exports the missing data from the platform and drops it straight into the chat.

Plugging your trusted data into Claude is a great way to make it more accurate. However, it’s worth still being cautious when starting out. Apply Morten’s tips above to confirm everything is correct before you trust the output.

How do you get consistent results from Claude?

Getting consistent results from Claude requires specific prompts, building skills for repeatable tasks, using projects to hold context, letting it learn your preferences, and matching the model to the task. It might not always produce the same exact results, but the ultimate goal is for it to follow specific lanes and formats in all future answers.

You’ve probably heard someone say: “I tried doing this task in Claude. It was brilliant once, and absolutely useless the next time.” Maybe you’ve even said this yourself. This is something we constantly hear from marketers, both within our own teams at Supermetrics and with our clients.

This inconsistency is frustrating — but it usually isn’t random. Most of the time, it’s due to a difference in what you fed Claude the first time around versus the second. Perhaps the prompt was vague, you forgot to add much-needed context, or you phrased a question a new way.

Whatever the reason, here are a few best practices that Morten shared to improve Claude’s consistency.

Be specific in your prompts

The quality of your prompts dictates the quality of your outputs. Be specific and tell Claude exactly what you want, and how it should provide the answer.

If you ask “How are my campaigns doing?”, you’ll get a vague answer. It will cherry-pick data that it considers relevant and present it in whatever format it chooses.

On the other hand, if you say “Create a table comparing cost per acquisition across my three Google Ads campaigns for the last 30 days, and highlight in red any results above €50”, then it’s obvious what you’re looking for.

Constraints guide Claude’s output and tell it what you’re looking for. The more precise the ask, the more reliable the reply.

Build skills for repeatable tasks

Do you find yourself regularly typing out (or copy-pasting) the same prompts, over and over again, in different Claude chats?

If so, there’s a better alternative: create a skill instead.

Morten likens creating a skill to training a new employee. “In Claude, you can create a skill that is a specialist in doing different kinds of analysis within a channel, or see the big picture, or create presentations in a certain way, so you get the same output all the time”.

Once you’ve built the skill, the output will be provided and formatted the same way every time. This saves you from having to re-explain yourself with each prompt.

Use projects to hold context

Skills teach Claude how to complete a task, but projects give it the context to do the task well.

For example, Layer creates a separate project for each client, with the client’s goals, targets, background, and more. When a member of the team asks Claude a question, it has the full picture. Its answers are grounded in the client’s specific context. It knows all about the client and what they want to achieve.

Let it learn your preferences

If you correct Claude then it will learn your preferences over time. It’s easy to think “It’s quicker if I correct this myself rather than giving Claude feedback”. Sure, it might be quicker — but over the long run, Claude will never know what you’re looking for.

On the other hand, if you correct the response and capture the fix, Claude will remember this for next time.

As Morten explains: “You’re kind of building a memory bank where it knows what is correct, what I like, what I don’t like. The more you use it, the better it gets.” For instance, you can correct Claude and then tell it to rewrite the skill or prompt to account for the correction the next time you use it.

Do you need to check the AI’s work?

Yes, you should always check the AI’s output to confirm the data is accurate.

Would you double check the numbers if you crunched the data manually? Of course you would. So you should do the same when using AI. It’s just best practice when you’re analyzing data.

It’s safe to say that by now, Layer trusts Supermetrics for Claude. “I still haven’t experienced any errors myself, and I don’t think my team has any either”, adds Morten. Yet despite this, they always make sure to manually check the data instead of blindly trusting it.

Why? Because the cost of being wrong in front of a client is high, and the cost of a ten-second check is almost nothing. It’s worth the time and effort — both of which are minimal.

How do you build trust in AI across a whole team?

Building trust in AI requires being open, sharing best practices, and being honest about what hasn’t worked.

In most agencies, there are a handful of people who are AI whizzes and others who aren’t (yet) as skillful. If this sounds like you, don’t worry — this is perfectly normal.

The trick is to create an environment where those most skilled at using Claude regularly share what’s working and what isn’t. This way, your entire team can replicate their approach, standardizing the output while also making everyone more efficient.

Let’s dig into a few approaches that helped Layer build trust in AI across their whole team.

Standardize and share the good stuff

When someone builds a skill that works, it gets rolled out to everyone, not kept on one person’s account. Layer documents all the skills that are live across the team. Everyone works from the same toolkit and produces output in the same recognizable shape.

Keep your skills tidy

This matters more than it sounds. Skills are often triggered by certain words, so if two skills are too similar, the wrong one can fire. A bit of naming discipline and a clear record of what each skill does keeps that from happening.

Layer also mixes shared, team-wide skills with personal ones, so individuals can build their own without cluttering everyone else’s setup.

Talk about it, regularly

Layer runs a dedicated Slack channel and a meeting every two weeks just for this — a standing space to ask “why didn’t I get the output I expected?” and “how do I do this better?”

Sharing what didn’t work is almost as useful as sharing what did. It helps others avoid the same mistakes and brings the team together in problem-solving mode.

Let the results do the talking

Marketers are smart and adaptable. If they see that something works, they’ll do it — even if it means changing their workflow.

According to Morten, nobody at Layer needed to be persuaded to use Supermetrics for Claude. Its value was obvious the moment people saw it in action.

That’s the real lesson for anyone rolling this out: you don’t win a team over by forcing them to do something. You win them over by showing them why they should do it. If you can teach them how to compress a task from hours into minutes, you won’t need to do any convincing.

Trust in AI isn’t something you either have or you don’t. It’s something you build, step by step.

It starts with grounding Claude in your own first-party data using a connection like Supermetrics for Claude, so it’s working from real numbers instead of guesswork. You then refine your questions, sharpen your prompts, and turn repeatable tasks into skills and projects that produce consistent output every time. Keep checking the output and share what works, so everyone on your team can benefit.

None of this requires taking a leap of faith. Layer didn’t trust the output on day one. And why should they? They built that trust by starting small, checking the numbers, and building from there — and now they’ve turned a 10-hour manual task into a 20-minute workflow.

You can take the same first step today. Watch the full Layer webinar to find out more about Layer’s story or take the plunge and get started.

FAQs