Deterministic SEO audits beat AI guesswork

"Ask the AI to audit your SEO" sounds modern. It's also a debugging nightmare: a generative model can give you a different answer on the same crawl, with no way to trace a recommendation to a piece of evidence.

This is why SEO Engine is deliberately deterministic. Here's the case for rule-based, explainable audits.

The problem with vibes-based auditing

An SEO audit is a decision-support document. It tells you what to fix and why. That only works if every finding is reproducible:

• Same input, same output. A credible auditor must produce the same findings for the same site, again and again. Change one rule version and you should know exactly which findings changed.
• Evidence over opinion. "Meta descriptions are missing" is a fact; "your content could feel more authoritative" is not a finding. Rule engines deal in the first kind.
• Traceable scores. If a page scores 78, a scoring guide should show you the exact rules that cost the 22 points.

Generative models produce delightful prose and unreliable numbers. For a tool, that is the wrong trade.

How a deterministic engine is built

The pipeline is a straight line, and each step produces something the next step consumes:

Crawler -> Rules Engine -> Evidence -> Diagnosis -> Prioritization -> Recommendation -> Repository

1. 01Crawler — walks the real site over your network, optionally rendering pages with Playwright where JavaScript matters.
2. 02Rules Engine — applies a versioned repository of rules. "Versioned" is the key word: rules change deliberately, in releases, so results can be compared across versions.
3. 03Evidence — each rule stores what it observed, verbatim.
4. 04Diagnosis — evidence becomes a diagnosis: what's wrong, on which pages.
5. 05Prioritization — diagnoses are weighted by impact so the output reads as a fix list, not a dump.
6. 06Recommendation — each item carries an actionable recommendation.
7. 07Repository — results land in a local SQLite store so the next run is a comparison, not another snapshot.

Deterministic site classification is part of the same philosophy: your site type decides which rules apply, and the decision is made by a ruleset, not by a model "feeling" what your business is.

The 49-rule baseline

The current rules repository covers the practical surface of SEO: technical crawl and indexation signals, on-page and content elements, internal linking, structured data, performance basics, international and local SEO, and commerce-related checks.

Each rule is small enough to read and audit. You know what ran, what it checked, and what it found. If you disagree with a rule's importance, the weights are visible too — nothing is hidden inside a model.

Where history pays off

One audit is a photograph. A series is a trend. Because every run is stored in the SQLite repository, the tool can answer questions photographs can't:

• Did the latest release break more internal links than it fixed?
• Are meta descriptions improving across the site, or only on the pages we fixed?
• Which priority-one issues are still open from three runs ago?

That's the difference between an audit and an audit system.

When deterministic beats generative

Generative analysis has its place — asking a model to read a page and suggest copy angles is genuinely useful. But the layer that needs to be correct, reproducible and explainable should not be the same layer that's allowed to be creative.

For SEO work specifically, the value is in the paper trail: evidence, diagnosis, priority, recommendation, history. None of that requires an LLM API key — which also means no per-crawl costs, no data leaving your machine, and no opaque verdicts.