Trakis

How much of an SEO audit is repetition?

Most SEO audits hand you a list with a row for every page a problem sits on. Alt text missing on 80 product pages is 80 rows; a theme that leaves out image sizes is a row on every page it builds. So we asked how much of such a list is the same problem said again. We audited 330 small sites, up to 100 pages each, and counted both: the rows, and the problems behind them.

On the median site, 627 rows came down to 19 problems: 32 rows for each one. On a quarter of the sites it was more than 43 rows a problem, on a quarter fewer than 23.

Per siteLower quarterMedianUpper quarter
Rows: one for every page a problem sits on402627799
Problems: the same findings, each counted once151923
Rows for each problem22.732.143.7

Across all 330 sites, 205,941 rows were 6,281 problems. A row isn’t wrong: each one is a page where the problem really is. But the fix is rarely made page by page. A missing image size or a title that runs too long usually comes from a template, and one change to the template clears every row it made.

More pages, more rows. Not more problems

SitesHow manyMedian rowsMedian problemsRows for each problem
We read 100 pages2277162036.1
The archive had fewer than 1001033301820.6

Read 100 pages instead of 40 and the list grows with them; the problems hardly move. We stopped at 100 pages a site, so for any site bigger than that, a crawl of every page would most likely list more rows for much the same problems.

The problems small sites have most

Of the sites each check could run on, the share with that problem on at least one page:

CheckWhat it findsSites with itShareA warning or worse on
PERF-001Images without declared dimensions29790.0%297
AI-002No short answer under the heading27984.5%0
CONTENT-004Very low text-to-HTML ratio27081.8%0
LINK-002Generic anchor text26379.7%0
ONPAGE-005Meta description outside the useful range26379.7%0
ONPAGE-008Skipped heading level26379.7%0
ONPAGE-002Title too long24574.2%238
CONTENT-001Thin content23470.9%200
ONPAGE-004Meta description missing21565.2%42
ONPAGE-006H1 missing21464.8%186
INDEX-003Canonical missing20762.7%207
LINK-001Too many links on the page20461.8%0
STRUCT-001The page has no structured data20261.2%0
PERF-002Too many blocking resources in the head19860.0%0
CONTENT-002Images without alternative text18355.5%179

The last column matters as much as the share. Most sites have a generic link text or a skipped heading level somewhere, and our audit calls those notices, not warnings. Images without width and height, a title over 70 characters and a missing canonical are warnings, and they are about as common. The full table, every check, is in the data.

How we did it

  • The sites. The 10,000 domains Study 1 drew at random from ranks 100,000 to 1,000,000 of the Tranco list of 13 October 2026 (N26LW), taken in that drawn order until 330 had at least 20 HTML pages that answered 200 in Common Crawl’s October 2026 crawl. That took 981 domains.
  • The pages. We didn’t ask the sites. Every page is the copy Common Crawl archived, read from its archive. For each site: the address (https or http, with or without www) with the most pages, its home page, then pages drawn at random, up to 100. 28,483 pages in all, a median of 100 a site. Pages the site’s robots.txt keeps our crawler from were left out, as our crawler leaves them out.
  • The audit. Our own engine, version 0.34.1, the one our audits ran in November 2026: every check that can be run on an archived copy, scored and grouped as our report groups them.
  • What we counted. Rows: one for every page a problem sits on, and one for a problem of the whole site. Problems: the same findings, one for each check and severity, as our report lists them. A check that charges some pages as a warning and others as a notice counts twice.

Checking our own numbers

We drew 30 pages at random, one from each of 30 sites, and had them judged by a second reader that never saw our code or what it said: another AI model, working from the HTML alone, on 13 of the checks a page’s HTML can settle by itself (title, description, headings, language, viewport, noindex, canonical, image sizes, alt text, structured data). One copy we couldn’t keep, as our antivirus removed it on download. On the other 29 pages, 376 of 377 verdicts agreed.

The one that didn’t was a page asking not to be indexed, like most pages built from the same template on its site. Our audit reports that once, as a setting of the template, not as a problem on every page, so it makes no row. The reader, going by the check alone, counted it.

Before that, an earlier check found a bug in our own crawler. It read every page as UTF-8, so on sites written in another encoding (windows-1251, GBK, ISO-8859-1), each letter outside ASCII came out as a replacement character: titles turned into rows of them, and a Chinese title counted up to twice its length. 23 of the 330 sites were read that way. We fixed the crawler; since version 0.34.1 it reads a page in the encoding the page declares, in our audits too. Those sites were read again, and every figure here is from after the fix. The medians didn’t move.

We also left out 16 pages whose copy in the archive stops partway through: part of a page, with its links and text cut, isn’t a page to audit.

What this can’t tell you

  • Small sites only. Ranks 100,000 to 1,000,000, and only those with 20 pages or more in the archive: 330 of 981. Sites that block Common Crawl, send every page to another domain, or have only a few pages aren’t here.
  • October’s copies, not the sites today. And no JavaScript was run. On 17 sites most pages build their content in the browser; without them, the median is 30.8 rows a problem instead of 32.1.
  • Fewer checks than a full audit. Some checks need what an archive doesn’t keep: a sitemap, the site’s other addresses, redirects followed live, page speed, Search Console. They aren’t counted, so a full audit lists more rows and more problems than this one. The list is below.
  • Some shares only one way. A link counts as dead only if Common Crawl kept the dead page’s answer, and duplicate titles and descriptions are found only among the pages we read: those are at least what the table says. Pages that look empty without JavaScript are counted without the browser our audits open for them: that one is at most.
  • Our rows, not every tool’s. The count of rows depends on which checks a tool runs. Another tool, with other checks, would list a different number. What this measures is the shape of a per-page list, not any one product’s.

The checks an archive can’t answer, and why:

CheckWhat it findsWhy it isn’t counted
CRAWL-001Sitemap URLs blocked by robots.txtno sitemap is read
CRAWL-002Response other than 200only pages that answered 200 are drawn
CRAWL-003Redirect chainredirects are not followed from the archive
CRAWL-004Page missing from the sitemapno sitemap is read
CRAWL-007Pages that could not be downloaded at alla page Common Crawl could not download is not in the archive
CRAWL-008The site is published on two addressesthe site’s other addresses are not asked for
CRAWL-009http:// serves the site instead of redirecting to https://the site’s other addresses are not asked for
CRAWL-010Sitemap lists pages that should not be thereno sitemap is read
GSC-001Many impressions, very few clicksno Search Console data
GSC-002Close to the first pageno Search Console data
GSC-003Page with traffic that asks not to be indexedno Search Console data
GSC-004Pages competing for the same searchno Search Console data
INDEX-008Missing pages answer 200no address that cannot exist is asked for
LINK-005Orphan pagesa sample of pages, not the whole site, so no page can be called an orphan
LINK-007Pages reached by one internal link onlya sample of pages, not the whole site, so no page can be called an orphan
PERF-003LCP above 2.5 sno PageSpeed measurement
PERF-005Layout shifts while loading (CLS)no PageSpeed measurement
PERF-006Slow response to clicks and taps (INP)no PageSpeed measurement
PERF-007Images heavier than 200 KBimages are not downloaded and weighed

The data

On Zenodo, doi.org/10.5281/zenodo.24173724: the totals, every check with how many sites it ran on and found, and one line for each site with its rows, problems and the checks it failed, with the method in full. No domain and no rank is in the data: below the top 10,000, as our bot page promises, nothing is published site by site that names a site. The data is CC BY 4.0. The audit engine is our product and its code isn’t published; what each check finds, and why, is on its own page. To have a domain taken out of the data, write to bot@trakis.io.

Sources

Last changed 13 November 2026.