Precision and Recall

Precision and recall are two evaluation metrics that describe what kind of mistakes a model makes. Precision is the share of the items a model returns that are correct. Recall is the share of all the correct items that the model manages to find. They are reported together because improving one often hurts the other, and the right balance depends on the job.

How it works

Both come from comparing predictions with a hand-labelled answer key, which gives four counts: true positives (correctly found), false positives (returned but wrong), false negatives (missed) and true negatives.

• Precision = true positives divided by (true positives plus false positives). “When it says yes, how often is it right?”
• Recall = true positives divided by (true positives plus false negatives). “Of everything it should have found, how much did it find?”

Most models have a threshold you can move. A stricter threshold returns fewer, safer results (higher precision, lower recall); a looser one returns more, with more errors (higher recall, lower precision). The F1 score is the harmonic mean of the two and gives one number when you need both. Unlike accuracy, precision and recall stay informative when the class you care about is rare.

How it’s used in SEO and marketing

• Entity extraction: named entity recognition has been scored with precision, recall and F1 since benchmarks such as CoNLL-2003. A tool that returns more entities is not better if many are wrong.
• Fuzzy matching for redirect maps: MLforSEO’s tutorial notes that lower similarity thresholds (around 60 to 70) capture more matches, while higher ones (85 to 90) favour accuracy and cut irrelevant matches. That is the precision and recall trade-off in practice.
• Text classification and intent labelling: check per-class precision and recall on a labelled sample before trusting a classifier on your full dataset.
• Monitoring and alerting: anomaly alerts with high recall catch every real drop but may flood you with false alarms.

Example

A model flags 100 pages as thin content. A human check finds 80 really are thin (precision 80%). The site actually has 160 thin pages, so the model found 80 of 160 (recall 50%). If you only want a safe shortlist to fix first, that is fine. If you are planning a full content prune, you need better recall.

Limitations and things to watch

• You need a labelled test set; without one you cannot calculate either metric. The MLforSEO NLP API comparison is explicit that its small test did not compute formal precision, recall or F1 for this reason.
• How you define “correct” changes the scores. CoNLL scoring only counts an entity if its span matches exactly.
• For multi-class problems, report precision and recall per class, then average, so a strong majority class does not hide a weak minority one.

Learn it on MLforSEO

• Comparing NLP APIs for Entity and Sentiment Analysis (vs Generative AI)
• 404 and Redirect Mapping with Fuzzy Matching in Google Colab (Python)
• What Machine Learning Classification Is, and Where It Fits in SEO
• Introduction to Machine Learning for SEO