How to prioritize an SEO backlog

An SEO backlog dies when every row is “high.” Priority is a sort on facts you can recount, then a judgment you can defend.

The facts are dull: how many URLs, which template, what status code, whether the URL is in the sitemap, whether anything links to it. The judgment is whether that pattern touches a page you need in search. Mix them into one red cell and the list cannot be audited.

Count the facts before you score the judgment

Write the finding as a count before anyone labels it severe.

A useful row has at least:

• URL pattern, not a single pretty example
• Number of URLs, and whether that is a sample or a full crawl
• HTTP status, or the index reason from Search Console
• Template or directory
• Whether the URL is linked, redirected, or only present in a sitemap

How to read Googlebot in your server logs is the source for the status and the recrawl rate. A crawler you run yourself is the source for internal links. Search Console is the source for indexing exclusions. None of the three replaces the others.

Google’s crawl budget guide, updated 22 July 2026, is explicit about who should even be in this conversation. The rough bands are 1 million unique pages that change about once a week, or 10,000 unique pages that change daily. Add sites with a large share of “Discovered, currently not indexed.” Below that, an outdated sitemap and the Page indexing report usually beat a 40-row audit.

Fix the errors that change crawl behavior first

Some issues change what the crawler does next. Those outrank a missing meta description on a page that already returns 200.

From the same crawl-budget page, and from Google’s status-code notes:

• 5xx and 429 lower the host’s crawl capacity. The limit is shared by Google’s crawlers on that hostname, so one sick app can starve the rest.
• Soft 404s stay in the crawl. They waste budget until the page either becomes real or returns a proper 404 or 410.
• noindex does not save a fetch. Google still requests the URL.
• Long redirect chains have a negative effect on crawling. Point internal links at the final URL.
• Duplicate URLs soak demand. Consolidating them is the inventory lever Google says you control most.

A title tag that is a bit vague does not do any of that. Park it behind the host errors. If you need a second sort key, use the template that earns the visits, not the template with the most rows.

A typed score belongs after the crawl data

Once the row has counts, a model can label it. It should not invent the counts.

Jev’s documentation says the model evaluates questions against state you send and returns typed values, probabilities, and, for Choice and Score, a confidence number. It does not write the ticket. TypeSafe’s own pattern is to split a big judgment into small questions and combine them in code. “How bad is SEO?” is a bad question. “Is this status code a server error?” is a small one, and you often do not need a model for it.

Their 15 September 2026 post publishes end-to-end times of 70ms to 500ms and an input price of $0.042 per million tokens. Speed is not accuracy. A confident score on an incomplete export is a confident score about the wrong site.

If you use a score, keep the formula visible. Example, not a standard: affected URLs, times a business weight you set, times a severity you defined, divided by an effort bucket a person assigned. When the weight changes, change the coefficient. Do not ask the model to “reprioritize the roadmap” in prose. And do not read the output as a forecast of rankings. No third-party model has Google’s ranking systems. Google says as much in its generative AI guide: no outside tool has internal ranking metrics.

Ship the smallest change you can remeasure

Pick one pattern. Collapse a redirect chain on one template, or stop linking a parameter that should not be crawled. Recrawl, or pull the next log window, and see if that pattern shrank.

A backlog that only grows is a taxonomy, not a queue. Close rows when the count moves, including when the count was never going to move because the site is too small for the issue to matter.