Most SEO reports end with a list of recommendations. "Improve your meta descriptions." "Add internal links." "Increase content depth." Some of them are right, some are generic, and a few are simply wrong for the site in question. The reader usually has no way to tell which is which, because the recommendation arrives without the reason behind it.

Evidence-based SEO is a simple idea: every recommendation should carry the observation that produced it — what was measured, where, by which source and when. This article explains why that matters, what it looks like in practice and how it changes the way AI should be used in SEO work.

The problem with recommendations without proof

A recommendation without evidence asks to be trusted. That creates three practical problems.

  • You cannot prioritise. If ten recommendations each say "important", there is no basis for deciding which to do first. The decision falls back on intuition.
  • You cannot verify. If a report says a page is slow, but not how slow, on which test or on which day, there is no way to confirm the problem still exists or that a fix worked.
  • You cannot spot errors. Tools and people make mistakes. A finding with its evidence attached can be checked in a minute; a finding without it has to be investigated from scratch or taken on faith.

The rise of AI-generated advice has made this worse. A language model can produce a confident, well-written recommendation for any website, including one it has never seen. Without evidence attached, a fluent but invented recommendation looks exactly like a correct one.

What evidence means in SEO

Evidence is not a long explanation. It is a small set of facts that anyone can check:

  • The observation — for example, "the title tag is 92 characters long" or "the page returns a 404 status".
  • The location — the exact page or query it applies to.
  • The source — a crawl of the site, a search results check, a speed test.
  • The date — when it was observed. Websites change; a finding from three months ago may no longer be true.

With these four items, a recommendation stops being an opinion. It becomes a statement that can be confirmed, disputed or re-measured.

From findings to opportunities

The second idea behind evidence-based SEO is that a finding is not a line in a report but an object with a life cycle. A missing description is found, accepted as worth fixing, worked on, fixed — and sometimes comes back after a redesign or a CMS update.

Treating each finding as a record with a state gives you things a report cannot:

  • No duplicates. When the site is crawled again, the same problem updates the existing record instead of appearing as a new one.
  • History. You can see when a problem first appeared, when it was fixed and whether it returned.
  • Honest closure. Finishing a task is not the same as fixing the problem. The problem is fixed when a new measurement no longer shows it.

In SEO Melody: this is how the Opportunities screen (Action Center) works. Every item has a source, evidence with the date it was observed and a state: open, accepted, in progress, postponed, ignored, fixed or returned. A repeat collection updates existing items by their fingerprint instead of adding copies, and closing a task does not close the opportunity on its own.

Priority you can read

Priority is where most tools ask for the most trust. A score of "87" or a red label says that something is important but not why. Evidence-based SEO asks for the priority to be calculated by a visible rule — how many pages are affected, how important those pages are, how serious the problem is — and for that calculation to be shown next to the result.

A visible calculation has a useful side effect: when you disagree with a priority, you can see exactly which input you disagree with. Perhaps the rule treats a page as important that is not important to your business. That is a conversation about facts, not about whose instinct is better.

In SEO Melody: each opportunity shows its priority together with the line of calculation behind it. The priority is computed by a transparent rule, not by a model, and the screen works even when AI features are switched off.

Where AI fits: explain, do not measure

AI is genuinely useful in SEO work, but its role should be limited to what it does reliably. Language models are good at explaining a known problem in plain language and drafting a fix. They are not measuring instruments. A model cannot tell you a page's load time, its position in search or its number of links unless it is given those facts — and if it is not, it may produce plausible numbers anyway.

An evidence-based workflow therefore separates the two jobs:

  1. Tools measure. Crawls, position checks and speed tests produce the facts.
  2. The service compares. Calculations — this value is above the threshold, this position fell by five places — are done by code, not by the model.
  3. The model explains. It receives the conclusion and the evidence and writes why it matters and what to do.
  4. The answer is checked. If the model's answer refers to something that was not in the evidence it received, the answer is not used.

In SEO Melody: AI explanations follow exactly these rules. The model receives conclusions already calculated by the service, its answer must match a fixed structure, user data is passed as a marked block whose instructions the model does not follow, and an answer that mentions a page outside the evidence is rejected entirely. Each explanation is stored with a fingerprint of its subject, so reopening it is free until the subject changes.

How to apply this without special tools

The approach does not depend on any particular software. A spreadsheet can carry most of it:

  • For every recommendation, add columns for the page, the observed value, the source and the date.
  • Write the priority rule down once and apply it to every row.
  • Keep a status column and never delete rows — mark them fixed, and reopen them if the problem returns.
  • When you use AI to explain or draft fixes, paste in the evidence and check that the answer refers only to it.

Summary

The value of an SEO recommendation depends on whether it can be checked. Attaching the observation, location, source and date to every finding makes priorities readable, progress measurable and mistakes visible. AI fits into this model well — as long as it explains what was measured rather than inventing what was not.