Ground Truth

Ground truth is the correct answer for each example in a dataset: the label you trust enough to treat as reality. It is what a supervised model learns from and, just as importantly, what you compare any model’s output against to see whether it works. In SEO terms, ground truth is the hand-checked list of intents, entities or categories that tells you whether a tool, script or LLM got it right.

How it works

Ground truth usually comes from people. Someone reads each keyword, page or comment and assigns the correct label, a process called data labelling. That labelled set then plays two roles:

• Training data: in supervised learning, the model learns the pattern between inputs and their true labels.
• Evaluation data: a separate held-out portion is used to score predictions with a confusion matrix, precision and recall or accuracy.

MLforSEO’s framework for choosing an ML task starts from the same question: do you already know the correct answer for at least a subset of your data? If yes, you can use supervised methods and check outcomes. If no, you are in unsupervised territory, and in clustering you fall back on internal measures such as the silhouette score because there are no ground truth labels.

How it’s used in SEO and marketing

• Testing a search intent classifier on a few hundred keywords you labelled yourself before running it on thousands.
• Benchmarking entity extraction and sentiment tools against a known answer key.
• Evaluating LLM outputs: a small golden set of prompts with agreed correct answers lets you track hallucinations and regressions, and calibrate an LLM-as-a-judge setup against human judgement.
• Auditing vendor claims: a tool’s own confidence score is not ground truth, so spot-check it against labels you trust.

Example

In MLforSEO’s comparison of NLP APIs and generative AI for entity extraction, the test text contained 13 core named entities (the people, organisations and brands). That list acted as the ground truth: Google, ChatGPT, Qwen and DeepSeek found all 13, Amazon Comprehend found 12 and IBM found 7. For sentiment, the post relies on an independent study in which about 6,000 social media comments were labelled by hand; against those human labels the three APIs scored between 67% and 72% accuracy. The post’s advice follows directly: keep a human review step, and do not treat any single API’s sentiment label as ground truth.

Limitations and things to watch

• Labels can be wrong or subjective. Intent and sentiment are often ambiguous, so have two people label a sample and check how often they agree.
• Ground truth ages. Search intent for a query can shift, so refresh labelled sets for long-running models.
• A biased or unrepresentative labelled set produces biased models and misleading scores (see bias).
• Never evaluate on the same examples you trained on; keep a held-out set.

Learn it on MLforSEO

• Comparing NLP APIs for Entity and Sentiment Analysis (vs Generative AI)
• The Characteristics of a Machine Learning Problem: A Task, Data and Solution Framework for SEOs
• What Machine Learning Classification Is, and Where It Fits in SEO
• Introduction to Machine Learning for SEO