Traffic has changed. Things have gotten weird. We have more data than ever, and the temptation to throw it at ChatGPT or Claude for analysis is understandable—but it’s often the wrong approach. LLMs make things up, they always tell you your question is good even when it isn’t, and your data isn’t safe. If you’re uploading GA4 or client data to custom GPTs, your client probably didn’t approve that, and your boss definitely didn’t either.

The question we should be asking: are you using the right tool for the right job?

When to Use LLMs vs. Machine Learning

LLMs shine with messy, chaotic, unstructured data—language-heavy content where you need interpretation, summarization, or creative generation. But if your data is structured, tabular, or numbers-related, machine learning models are the superior choice. And here’s the beautiful part: you can use LLMs to help you build and run those machine learning models.

Why Machine Learning Models Win for Data Analysis

Machine learning models are built to do specific things extremely well. BERT topic modeling for keyword clustering is phenomenal, working across many languages. Many of these models have been around for 10-15+ years, tested rigorously in academic settings with transparent training data—so if you’re worried about bias, you can actually investigate it.

The advantages compound from there. ML models are deterministic—you get the same result every time, unlike ChatGPT deciding it feels differently today. They’re more efficient than LLMs, so if you’re hitting API limits or token caps, ML is more cost-effective. And critically, they’re honest with you. They won’t tell you your question was great when it wasn’t. They just show you what’s happening in the data.

Google Colab: Your Free Machine Learning Environment

The barrier to machine learning has always been the intimidating setup—getting a development environment working can take half a day and still fail. Google Colab eliminates that friction. It’s a cloud-based Jupyter notebook that sets up your environment automatically, and it’s free.

What makes Colab particularly valuable for SEO work: your data stays private. When you upload GSC or GA4 data to Google Colab, it’s not being used to train anything else. One client had developer comments in their website code calling someone an idiot—that kind of thing has shown up in ChatGPT outputs. It won’t happen with Colab.

Colab notebooks are easy to share for methodology review, they have version control built in, and you can break things without consequences—it’s a learning environment where experimentation is encouraged. If you start hitting RAM or disk limits on large datasets, you can add more CPUs, which actually means you’re using it correctly.

The Workflow: LLMs Build Your ML Tools

Here’s the approach that bridges both worlds: use Claude or your preferred LLM to generate the machine learning code, then run it in Google Colab. The workflow starts with defining what you’re trying to accomplish—what answer you’re looking for and what data you have to get there. Then ask Claude which models would work best for your use case.

Choose your model, have Claude build the entire notebook, download it, and upload to Colab. Expect some troubleshooting—it typically doesn’t work perfectly on the first try. A Shopify engineer shared a useful trick: end every answer with “are you sure?” and the LLM will rewrite the whole thing, often producing more efficient code.

Problem Types: Speaking the Right Language

Using the correct terminology in your prompts dramatically improves results. Classification is for grouping things together—useful for tagging URLs by content topic when you lack good folder structure. Regression is for predicting the future or finding patterns. Clustering finds neighborhoods of related items—similar to classification but focused on discovering natural groupings rather than assigning predefined categories.

Dimensionality reduction (or “removing noise”) strips out junk data to find the gems—helpful for cleaning up ad campaign data. And anomaly detection identifies what’s weird in your data, which uses the same models as forecasting—you’re either asking the model to point out outliers or to predict what comes next.

Data Considerations

Several factors influence which models work best. Dataset size matters—30 days versus 16 months requires different approaches. Format matters—you can even use images with appropriate models. For time series data, consider whether seasonality affects your analysis. Black Friday doesn’t fall on the same date each year, and some models can account for holidays across different geographic regions.

The number of dimensions you’re analyzing influences model selection. And data quality is crucial—some models handle null or zero values gracefully while others fall apart entirely. Getting in the habit of cleaning your data first is usually wise.

Three Favorite Models for Getting Started

Isolation Forest

A flexible, fast model that works across multiple dimensions (though it struggles past 8 dimensions). It’s robust enough to handle messy data and still produce useful insights. The tradeoff: it sometimes flags too many anomalies, but having more leads to investigate beats having none. It’s less effective at detecting gradual changes.

LOF (Local Outlier Factor)

LOF excels at finding neighbors and identifying what doesn’t belong in the neighborhood. It’s particularly strong with multiple dimensions and keyword-related analysis—differentiating between “JavaScript,” “migrations,” and “JavaScript migrations” when they’re related but distinct. Excellent for anomaly detection in A/B tests. Less effective for far-reaching pattern detection or when you have wildly different categories of data.

Prophet (Forecast Residual Approach)

Prophet, created by the Facebook team, is the favorite but also the highest maintenance. It handles time series data beautifully—exactly what SEOs typically work with—and excels at identifying gradual shifts. It can distinguish algorithm update impacts from slow content drift or entity changes, and it’s excellent for SERP volatility analysis.

Prophet supports holidays across many countries and works in numerous languages. The downside: it requires substantial data (30 days probably isn’t enough unless you’re using log files), and the model occasionally changes in ways that break existing notebooks. One notebook stopped working a week after creation because the model updated—it took two hours to fix.

Don’t Forget the Wins

Anomaly detection models show where things went wrong, but they also highlight where things went right. Too often we focus exclusively on problems. Analyzing positive anomalies reveals what to double down on—sometimes the path forward is doing more of what’s already working rather than fixing what’s broken.

Communities for Continued Learning

Lazarina Stoy’s ML for SEO community offers extensive free resources for those interested in machine learning applications. The SEO community founded by Noah Lerner is a great place to ask questions with a supportive rule: you’re not allowed to say “is this a dumb question”—you have to say “help me get smart.” Women in Tech SEO provides another strong learning environment. Brittany Muller’s Orange Labs community offers additional resources, though it’s a paid option.

Key Takeaways

Stop uploading structured data to LLMs—your data isn’t safe and the analysis isn’t reliable. Use LLMs for what they’re good at (messy, unstructured, language-heavy tasks) and machine learning for what it’s good at (structured, tabular, numerical analysis). Google Colab provides a free, private environment to run ML models without the setup headaches. The workflow that bridges both worlds: use LLMs to write your ML code, then run it in Colab. Learn the right terminology—classification, regression, clustering, dimensionality reduction, anomaly detection—to get better results from your prompts. Start with Isolation Forest, LOF, or Prophet depending on your use case. And remember to analyze your wins, not just your problems.